llvm/test/CodeGen/X86/vector-interleaved-load-i64-stride-8.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx  | FileCheck %s --check-prefixes=AVX
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX2
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX512
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX512DQ
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512BW
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX512DQ-BW
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-BW-FCP

; These patterns are produced by LoopVectorizer for interleaved loads.

define void @load_i64_stride8_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf2:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; SSE-NEXT:    movaps 112(%rdi), %xmm0
; SSE-NEXT:    movaps 96(%rdi), %xmm1
; SSE-NEXT:    movaps 80(%rdi), %xmm2
; SSE-NEXT:    movaps 64(%rdi), %xmm3
; SSE-NEXT:    movaps (%rdi), %xmm4
; SSE-NEXT:    movaps 16(%rdi), %xmm5
; SSE-NEXT:    movaps 32(%rdi), %xmm6
; SSE-NEXT:    movaps 48(%rdi), %xmm7
; SSE-NEXT:    movaps %xmm4, %xmm8
; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm3[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]
; SSE-NEXT:    movaps %xmm5, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
; SSE-NEXT:    movaps %xmm6, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm1[1]
; SSE-NEXT:    movaps %xmm7, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT:    movaps %xmm8, (%rsi)
; SSE-NEXT:    movaps %xmm4, (%rdx)
; SSE-NEXT:    movaps %xmm3, (%rcx)
; SSE-NEXT:    movaps %xmm5, (%r8)
; SSE-NEXT:    movaps %xmm2, (%r9)
; SSE-NEXT:    movaps %xmm6, (%r11)
; SSE-NEXT:    movaps %xmm1, (%r10)
; SSE-NEXT:    movaps %xmm7, (%rax)
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i64_stride8_vf2:
; AVX:       # %bb.0:
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX-NEXT:    vmovaps (%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX-NEXT:    vmovaps 16(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm1[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm1[1]
; AVX-NEXT:    vmovaps 96(%rdi), %xmm3
; AVX-NEXT:    vmovaps 32(%rdi), %xmm5
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm3[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm3[1]
; AVX-NEXT:    vmovaps 112(%rdi), %xmm5
; AVX-NEXT:    vmovaps 48(%rdi), %xmm7
; AVX-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm5[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm7[1],xmm5[1]
; AVX-NEXT:    vmovaps %xmm2, (%rsi)
; AVX-NEXT:    vmovaps %xmm0, (%rdx)
; AVX-NEXT:    vmovaps %xmm4, (%rcx)
; AVX-NEXT:    vmovaps %xmm1, (%r8)
; AVX-NEXT:    vmovaps %xmm6, (%r9)
; AVX-NEXT:    vmovaps %xmm3, (%r11)
; AVX-NEXT:    vmovaps %xmm8, (%r10)
; AVX-NEXT:    vmovaps %xmm5, (%rax)
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i64_stride8_vf2:
; AVX2:       # %bb.0:
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-NEXT:    vmovaps 96(%rdi), %ymm0
; AVX2-NEXT:    vmovaps 32(%rdi), %ymm1
; AVX2-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-NEXT:    vmovaps (%rdi), %ymm3
; AVX2-NEXT:    vmovaps (%rdi), %xmm4
; AVX2-NEXT:    vmovaps 32(%rdi), %xmm5
; AVX2-NEXT:    vmovaps 64(%rdi), %xmm6
; AVX2-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm4[0],xmm6[0]
; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-NEXT:    vmovaps 96(%rdi), %xmm3
; AVX2-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm3[0]
; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm3[1]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovaps %xmm7, (%rsi)
; AVX2-NEXT:    vmovaps %xmm4, (%rdx)
; AVX2-NEXT:    vextractf128 $1, %ymm6, (%rcx)
; AVX2-NEXT:    vextractf128 $1, %ymm2, (%r8)
; AVX2-NEXT:    vmovaps %xmm8, (%r9)
; AVX2-NEXT:    vmovaps %xmm3, (%r11)
; AVX2-NEXT:    vextractf128 $1, %ymm5, (%r10)
; AVX2-NEXT:    vextractf128 $1, %ymm0, (%rax)
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i64_stride8_vf2:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm3
; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm4
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm5
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm6
; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm4[0],xmm6[0]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm3
; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm3[0]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm3[1]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovaps %xmm7, (%rsi)
; AVX2-FP-NEXT:    vmovaps %xmm4, (%rdx)
; AVX2-FP-NEXT:    vextractf128 $1, %ymm6, (%rcx)
; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, (%r8)
; AVX2-FP-NEXT:    vmovaps %xmm8, (%r9)
; AVX2-FP-NEXT:    vmovaps %xmm3, (%r11)
; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, (%r10)
; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, (%rax)
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i64_stride8_vf2:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm5
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm6
; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm4[0],xmm6[0]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm3
; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm5[0],xmm3[0]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm5[1],xmm3[1]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovaps %xmm7, (%rsi)
; AVX2-FCP-NEXT:    vmovaps %xmm4, (%rdx)
; AVX2-FCP-NEXT:    vextractf128 $1, %ymm6, (%rcx)
; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, (%r8)
; AVX2-FCP-NEXT:    vmovaps %xmm8, (%r9)
; AVX2-FCP-NEXT:    vmovaps %xmm3, (%r11)
; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, (%r10)
; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, (%rax)
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i64_stride8_vf2:
; AVX512:       # %bb.0:
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512-NEXT:    vmovaps (%rdi), %xmm1
; AVX512-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512-NEXT:    vmovaps %xmm1, (%r8)
; AVX512-NEXT:    vmovaps %xmm7, (%r9)
; AVX512-NEXT:    vmovaps %xmm2, (%r11)
; AVX512-NEXT:    vmovaps %xmm8, (%r10)
; AVX512-NEXT:    vmovaps %xmm3, (%rax)
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i64_stride8_vf2:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-FCP-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512-FCP-NEXT:    vmovaps (%rdi), %xmm1
; AVX512-FCP-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512-FCP-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512-FCP-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512-FCP-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512-FCP-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512-FCP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512-FCP-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512-FCP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512-FCP-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512-FCP-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512-FCP-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512-FCP-NEXT:    vmovaps %xmm1, (%r8)
; AVX512-FCP-NEXT:    vmovaps %xmm7, (%r9)
; AVX512-FCP-NEXT:    vmovaps %xmm2, (%r11)
; AVX512-FCP-NEXT:    vmovaps %xmm8, (%r10)
; AVX512-FCP-NEXT:    vmovaps %xmm3, (%rax)
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i64_stride8_vf2:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512DQ-NEXT:    vmovaps (%rdi), %xmm1
; AVX512DQ-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512DQ-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512DQ-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512DQ-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512DQ-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512DQ-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512DQ-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512DQ-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512DQ-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512DQ-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512DQ-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512DQ-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512DQ-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512DQ-NEXT:    vmovaps %xmm1, (%r8)
; AVX512DQ-NEXT:    vmovaps %xmm7, (%r9)
; AVX512DQ-NEXT:    vmovaps %xmm2, (%r11)
; AVX512DQ-NEXT:    vmovaps %xmm8, (%r10)
; AVX512DQ-NEXT:    vmovaps %xmm3, (%rax)
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i64_stride8_vf2:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-FCP-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vmovaps (%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512DQ-FCP-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512DQ-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512DQ-FCP-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512DQ-FCP-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512DQ-FCP-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512DQ-FCP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512DQ-FCP-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovaps %xmm1, (%r8)
; AVX512DQ-FCP-NEXT:    vmovaps %xmm7, (%r9)
; AVX512DQ-FCP-NEXT:    vmovaps %xmm2, (%r11)
; AVX512DQ-FCP-NEXT:    vmovaps %xmm8, (%r10)
; AVX512DQ-FCP-NEXT:    vmovaps %xmm3, (%rax)
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride8_vf2:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512BW-NEXT:    vmovaps (%rdi), %xmm1
; AVX512BW-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512BW-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512BW-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512BW-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512BW-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512BW-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512BW-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512BW-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512BW-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512BW-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512BW-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512BW-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512BW-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512BW-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512BW-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512BW-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512BW-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512BW-NEXT:    vmovaps %xmm1, (%r8)
; AVX512BW-NEXT:    vmovaps %xmm7, (%r9)
; AVX512BW-NEXT:    vmovaps %xmm2, (%r11)
; AVX512BW-NEXT:    vmovaps %xmm8, (%r10)
; AVX512BW-NEXT:    vmovaps %xmm3, (%rax)
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i64_stride8_vf2:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512BW-FCP-NEXT:    vmovaps (%rdi), %xmm1
; AVX512BW-FCP-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512BW-FCP-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512BW-FCP-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512BW-FCP-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512BW-FCP-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512BW-FCP-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512BW-FCP-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512BW-FCP-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512BW-FCP-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512BW-FCP-NEXT:    vmovaps %xmm1, (%r8)
; AVX512BW-FCP-NEXT:    vmovaps %xmm7, (%r9)
; AVX512BW-FCP-NEXT:    vmovaps %xmm2, (%r11)
; AVX512BW-FCP-NEXT:    vmovaps %xmm8, (%r10)
; AVX512BW-FCP-NEXT:    vmovaps %xmm3, (%rax)
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i64_stride8_vf2:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vmovaps (%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512DQ-BW-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512DQ-BW-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512DQ-BW-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512DQ-BW-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512DQ-BW-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512DQ-BW-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512DQ-BW-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512DQ-BW-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512DQ-BW-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512DQ-BW-NEXT:    vmovaps %xmm1, (%r8)
; AVX512DQ-BW-NEXT:    vmovaps %xmm7, (%r9)
; AVX512DQ-BW-NEXT:    vmovaps %xmm2, (%r11)
; AVX512DQ-BW-NEXT:    vmovaps %xmm8, (%r10)
; AVX512DQ-BW-NEXT:    vmovaps %xmm3, (%rax)
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i64_stride8_vf2:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX512DQ-BW-FCP-NEXT:    vmovaps (%rdi), %xmm1
; AVX512DQ-BW-FCP-NEXT:    vmovaps 16(%rdi), %xmm2
; AVX512DQ-BW-FCP-NEXT:    vmovaps 32(%rdi), %xmm3
; AVX512DQ-BW-FCP-NEXT:    vmovaps 48(%rdi), %xmm4
; AVX512DQ-BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm1[0],xmm0[0]
; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX512DQ-BW-FCP-NEXT:    vmovaps 80(%rdi), %xmm1
; AVX512DQ-BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
; AVX512DQ-BW-FCP-NEXT:    vmovaps 96(%rdi), %xmm2
; AVX512DQ-BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm3[0],xmm2[0]
; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX512DQ-BW-FCP-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX512DQ-BW-FCP-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm4[0],xmm3[0]
; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm5, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm0, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm6, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm1, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm7, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm2, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm8, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %xmm3, (%rax)
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <16 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 0, i32 8>
  %strided.vec1 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 1, i32 9>
  %strided.vec2 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 2, i32 10>
  %strided.vec3 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 3, i32 11>
  %strided.vec4 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 4, i32 12>
  %strided.vec5 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 5, i32 13>
  %strided.vec6 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 6, i32 14>
  %strided.vec7 = shufflevector <16 x i64> %wide.vec, <16 x i64> poison, <2 x i32> <i32 7, i32 15>
  store <2 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <2 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <2 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <2 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <2 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <2 x i64> %strided.vec5, ptr %out.vec5, align 64
  store <2 x i64> %strided.vec6, ptr %out.vec6, align 64
  store <2 x i64> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i64_stride8_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf4:
; SSE:       # %bb.0:
; SSE-NEXT:    movaps 112(%rdi), %xmm6
; SSE-NEXT:    movaps 240(%rdi), %xmm0
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 96(%rdi), %xmm9
; SSE-NEXT:    movaps 224(%rdi), %xmm11
; SSE-NEXT:    movaps 160(%rdi), %xmm0
; SSE-NEXT:    movaps 80(%rdi), %xmm14
; SSE-NEXT:    movaps 208(%rdi), %xmm15
; SSE-NEXT:    movaps 144(%rdi), %xmm2
; SSE-NEXT:    movaps 64(%rdi), %xmm12
; SSE-NEXT:    movaps (%rdi), %xmm7
; SSE-NEXT:    movaps 16(%rdi), %xmm5
; SSE-NEXT:    movaps 32(%rdi), %xmm3
; SSE-NEXT:    movaps 48(%rdi), %xmm4
; SSE-NEXT:    movaps 192(%rdi), %xmm13
; SSE-NEXT:    movaps 128(%rdi), %xmm8
; SSE-NEXT:    movaps %xmm8, %xmm10
; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm13[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm13[1]
; SSE-NEXT:    movaps %xmm7, %xmm13
; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm12[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm12[1]
; SSE-NEXT:    movaps %xmm2, %xmm12
; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm15[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm15[1]
; SSE-NEXT:    movaps %xmm5, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm14[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm14[1]
; SSE-NEXT:    movaps %xmm0, %xmm14
; SSE-NEXT:    movlhps {{.*#+}} xmm14 = xmm14[0],xmm11[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm11[1]
; SSE-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm3, %xmm11
; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm9[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm9[1]
; SSE-NEXT:    movaps %xmm4, %xmm9
; SSE-NEXT:    movlhps {{.*#+}} xmm9 = xmm9[0],xmm6[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]
; SSE-NEXT:    movaps 176(%rdi), %xmm6
; SSE-NEXT:    movaps %xmm6, %xmm1
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]
; SSE-NEXT:    movaps %xmm13, (%rsi)
; SSE-NEXT:    movaps %xmm10, 16(%rsi)
; SSE-NEXT:    movaps %xmm7, (%rdx)
; SSE-NEXT:    movaps %xmm8, 16(%rdx)
; SSE-NEXT:    movaps %xmm15, (%rcx)
; SSE-NEXT:    movaps %xmm12, 16(%rcx)
; SSE-NEXT:    movaps %xmm5, (%r8)
; SSE-NEXT:    movaps %xmm2, 16(%r8)
; SSE-NEXT:    movaps %xmm11, (%r9)
; SSE-NEXT:    movaps %xmm14, 16(%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm3, (%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm9, (%rax)
; SSE-NEXT:    movaps %xmm1, 16(%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm6, 16(%rax)
; SSE-NEXT:    movaps %xmm4, (%rax)
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i64_stride8_vf4:
; AVX:       # %bb.0:
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX-NEXT:    vmovaps 224(%rdi), %ymm0
; AVX-NEXT:    vmovaps 160(%rdi), %ymm1
; AVX-NEXT:    vmovaps 192(%rdi), %ymm4
; AVX-NEXT:    vmovaps 128(%rdi), %ymm5
; AVX-NEXT:    vmovaps 192(%rdi), %xmm3
; AVX-NEXT:    vmovaps 128(%rdi), %xmm6
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm6[0],xmm3[0]
; AVX-NEXT:    vmovaps 64(%rdi), %xmm7
; AVX-NEXT:    vmovaps (%rdi), %xmm8
; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm7[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm6[1],xmm3[1]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm6 = xmm8[1],xmm7[1]
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX-NEXT:    vmovaps 80(%rdi), %xmm8
; AVX-NEXT:    vmovaps 16(%rdi), %xmm10
; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm10[0],xmm8[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm10[1],xmm8[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
; AVX-NEXT:    vmovaps 224(%rdi), %xmm5
; AVX-NEXT:    vmovaps 160(%rdi), %xmm8
; AVX-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm8[0],xmm5[0]
; AVX-NEXT:    vmovaps 96(%rdi), %xmm11
; AVX-NEXT:    vmovaps 32(%rdi), %xmm12
; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm8[1],xmm5[1]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm12[1],xmm11[1]
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX-NEXT:    vmovaps 112(%rdi), %xmm12
; AVX-NEXT:    vmovaps 48(%rdi), %xmm14
; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm14[0],xmm12[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm14[1],xmm12[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovaps %xmm9, (%rsi)
; AVX-NEXT:    vmovaps %xmm2, 16(%rsi)
; AVX-NEXT:    vmovaps %xmm6, (%rdx)
; AVX-NEXT:    vmovaps %xmm3, 16(%rdx)
; AVX-NEXT:    vmovaps %ymm7, (%rcx)
; AVX-NEXT:    vmovaps %ymm4, (%r8)
; AVX-NEXT:    vmovaps %xmm13, (%r9)
; AVX-NEXT:    vmovaps %xmm10, 16(%r9)
; AVX-NEXT:    vmovaps %xmm8, (%r11)
; AVX-NEXT:    vmovaps %xmm5, 16(%r11)
; AVX-NEXT:    vmovaps %ymm11, (%r10)
; AVX-NEXT:    vmovaps %ymm0, (%rax)
; AVX-NEXT:    vzeroupper
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i64_stride8_vf4:
; AVX2:       # %bb.0:
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-NEXT:    vmovaps 96(%rdi), %ymm0
; AVX2-NEXT:    vmovaps 32(%rdi), %ymm1
; AVX2-NEXT:    vmovaps 224(%rdi), %ymm2
; AVX2-NEXT:    vmovaps 160(%rdi), %ymm3
; AVX2-NEXT:    vmovaps 64(%rdi), %ymm4
; AVX2-NEXT:    vmovaps (%rdi), %ymm5
; AVX2-NEXT:    vmovaps 192(%rdi), %ymm6
; AVX2-NEXT:    vmovaps 128(%rdi), %ymm7
; AVX2-NEXT:    vmovaps (%rdi), %xmm8
; AVX2-NEXT:    vmovaps 32(%rdi), %xmm9
; AVX2-NEXT:    vmovaps 64(%rdi), %xmm10
; AVX2-NEXT:    vinsertf128 $1, 192(%rdi), %ymm10, %ymm10
; AVX2-NEXT:    vinsertf128 $1, 128(%rdi), %ymm8, %ymm8
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm6[2,3]
; AVX2-NEXT:    vmovaps 96(%rdi), %xmm5
; AVX2-NEXT:    vinsertf128 $1, 224(%rdi), %ymm5, %ymm5
; AVX2-NEXT:    vinsertf128 $1, 160(%rdi), %ymm9, %ymm6
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm6 = ymm9[2,3],ymm6[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
; AVX2-NEXT:    vmovaps %ymm11, (%rsi)
; AVX2-NEXT:    vmovaps %ymm8, (%rdx)
; AVX2-NEXT:    vmovaps %ymm10, (%rcx)
; AVX2-NEXT:    vmovaps %ymm4, (%r8)
; AVX2-NEXT:    vmovaps %ymm7, (%r9)
; AVX2-NEXT:    vmovaps %ymm5, (%r11)
; AVX2-NEXT:    vmovaps %ymm6, (%r10)
; AVX2-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i64_stride8_vf4:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovaps 224(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm5
; AVX2-FP-NEXT:    vmovaps 192(%rdi), %ymm6
; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm7
; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm8
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm9
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm10
; AVX2-FP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm10, %ymm10
; AVX2-FP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm8, %ymm8
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm6[2,3]
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm5
; AVX2-FP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm5, %ymm5
; AVX2-FP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm9, %ymm6
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm6 = ymm9[2,3],ymm6[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
; AVX2-FP-NEXT:    vmovaps %ymm11, (%rsi)
; AVX2-FP-NEXT:    vmovaps %ymm8, (%rdx)
; AVX2-FP-NEXT:    vmovaps %ymm10, (%rcx)
; AVX2-FP-NEXT:    vmovaps %ymm4, (%r8)
; AVX2-FP-NEXT:    vmovaps %ymm7, (%r9)
; AVX2-FP-NEXT:    vmovaps %ymm5, (%r11)
; AVX2-FP-NEXT:    vmovaps %ymm6, (%r10)
; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i64_stride8_vf4:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm5
; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm7
; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm8
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm9
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm10
; AVX2-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm10, %ymm10
; AVX2-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm8, %ymm8
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm6[2,3]
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm5
; AVX2-FCP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm5, %ymm5
; AVX2-FCP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm9, %ymm6
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm6 = ymm9[2,3],ymm6[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
; AVX2-FCP-NEXT:    vmovaps %ymm11, (%rsi)
; AVX2-FCP-NEXT:    vmovaps %ymm8, (%rdx)
; AVX2-FCP-NEXT:    vmovaps %ymm10, (%rcx)
; AVX2-FCP-NEXT:    vmovaps %ymm4, (%r8)
; AVX2-FCP-NEXT:    vmovaps %ymm7, (%r9)
; AVX2-FCP-NEXT:    vmovaps %ymm5, (%r11)
; AVX2-FCP-NEXT:    vmovaps %ymm6, (%r10)
; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i64_stride8_vf4:
; AVX512:       # %bb.0:
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512-NEXT:    vmovaps (%rdi), %xmm0
; AVX512-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512-NEXT:    vmovaps (%rdi), %ymm10
; AVX512-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512-NEXT:    vmovaps %ymm1, (%r8)
; AVX512-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i64_stride8_vf4:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512-FCP-NEXT:    vmovaps (%rdi), %xmm0
; AVX512-FCP-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512-FCP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512-FCP-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512-FCP-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512-FCP-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512-FCP-NEXT:    vmovaps (%rdi), %ymm10
; AVX512-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512-FCP-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512-FCP-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512-FCP-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512-FCP-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512-FCP-NEXT:    vmovaps %ymm1, (%r8)
; AVX512-FCP-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512-FCP-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512-FCP-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512-FCP-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i64_stride8_vf4:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-NEXT:    vmovaps (%rdi), %xmm0
; AVX512DQ-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512DQ-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512DQ-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512DQ-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512DQ-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512DQ-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512DQ-NEXT:    vmovaps (%rdi), %ymm10
; AVX512DQ-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512DQ-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512DQ-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512DQ-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512DQ-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512DQ-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512DQ-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512DQ-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512DQ-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512DQ-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512DQ-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512DQ-NEXT:    vmovaps %ymm1, (%r8)
; AVX512DQ-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512DQ-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512DQ-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512DQ-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i64_stride8_vf4:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vmovaps (%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512DQ-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-FCP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-FCP-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512DQ-FCP-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512DQ-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512DQ-FCP-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512DQ-FCP-NEXT:    vmovaps (%rdi), %ymm10
; AVX512DQ-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512DQ-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512DQ-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512DQ-FCP-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512DQ-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovaps %ymm1, (%r8)
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride8_vf4:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512BW-NEXT:    vmovaps (%rdi), %xmm0
; AVX512BW-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512BW-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512BW-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512BW-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512BW-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512BW-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512BW-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512BW-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512BW-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512BW-NEXT:    vmovaps (%rdi), %ymm10
; AVX512BW-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512BW-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512BW-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512BW-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512BW-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512BW-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512BW-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512BW-NEXT:    vmovaps %ymm1, (%r8)
; AVX512BW-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512BW-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512BW-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512BW-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i64_stride8_vf4:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vmovaps (%rdi), %xmm0
; AVX512BW-FCP-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512BW-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512BW-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512BW-FCP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512BW-FCP-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512BW-FCP-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512BW-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512BW-FCP-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512BW-FCP-NEXT:    vmovaps (%rdi), %ymm10
; AVX512BW-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512BW-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512BW-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512BW-FCP-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512BW-FCP-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512BW-FCP-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512BW-FCP-NEXT:    vmovaps %ymm1, (%r8)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i64_stride8_vf4:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vmovaps (%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512DQ-BW-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-BW-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512DQ-BW-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512DQ-BW-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512DQ-BW-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512DQ-BW-NEXT:    vmovaps (%rdi), %ymm10
; AVX512DQ-BW-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512DQ-BW-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512DQ-BW-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512DQ-BW-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512DQ-BW-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512DQ-BW-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512DQ-BW-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512DQ-BW-NEXT:    vmovaps %ymm1, (%r8)
; AVX512DQ-BW-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512DQ-BW-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512DQ-BW-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512DQ-BW-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i64_stride8_vf4:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovaps (%rdi), %xmm0
; AVX512DQ-BW-FCP-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX512DQ-BW-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm1, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-BW-FCP-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX512DQ-BW-FCP-NEXT:    vmovaps 128(%rdi), %ymm2
; AVX512DQ-BW-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX512DQ-BW-FCP-NEXT:    vmovaps 64(%rdi), %ymm9
; AVX512DQ-BW-FCP-NEXT:    vmovaps (%rdi), %ymm10
; AVX512DQ-BW-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[2],ymm9[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX512DQ-BW-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm9[1],ymm10[3],ymm9[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm2 = [0,0,4,12]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm9 = [4,12]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm9 = [0,0,5,13]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm10 = [5,13]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm10 = [0,0,6,14]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm11 = [6,14]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} ymm11 = [0,0,7,15]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm5, %zmm6, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbq {{.*#+}} xmm5 = [7,15]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm4, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovaps %ymm7, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %ymm8, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %ymm0, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %ymm1, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm2, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm9, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm10, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm3, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <32 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
  %strided.vec1 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 1, i32 9, i32 17, i32 25>
  %strided.vec2 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 2, i32 10, i32 18, i32 26>
  %strided.vec3 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 3, i32 11, i32 19, i32 27>
  %strided.vec4 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 4, i32 12, i32 20, i32 28>
  %strided.vec5 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 5, i32 13, i32 21, i32 29>
  %strided.vec6 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 6, i32 14, i32 22, i32 30>
  %strided.vec7 = shufflevector <32 x i64> %wide.vec, <32 x i64> poison, <4 x i32> <i32 7, i32 15, i32 23, i32 31>
  store <4 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <4 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <4 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <4 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <4 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <4 x i64> %strided.vec5, ptr %out.vec5, align 64
  store <4 x i64> %strided.vec6, ptr %out.vec6, align 64
  store <4 x i64> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i64_stride8_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf8:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $152, %rsp
; SSE-NEXT:    movaps 336(%rdi), %xmm0
; SSE-NEXT:    movaps 464(%rdi), %xmm1
; SSE-NEXT:    movaps 400(%rdi), %xmm8
; SSE-NEXT:    movaps 80(%rdi), %xmm2
; SSE-NEXT:    movaps 208(%rdi), %xmm3
; SSE-NEXT:    movaps 144(%rdi), %xmm9
; SSE-NEXT:    movaps 320(%rdi), %xmm4
; SSE-NEXT:    movaps 256(%rdi), %xmm11
; SSE-NEXT:    movaps 448(%rdi), %xmm5
; SSE-NEXT:    movaps 384(%rdi), %xmm12
; SSE-NEXT:    movaps 64(%rdi), %xmm6
; SSE-NEXT:    movaps (%rdi), %xmm13
; SSE-NEXT:    movaps 16(%rdi), %xmm10
; SSE-NEXT:    movaps 192(%rdi), %xmm7
; SSE-NEXT:    movaps 128(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]
; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm7[1]
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm13, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]
; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm12, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm5[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm5
; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm4[0]
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm4[1]
; SSE-NEXT:    movaps %xmm11, (%rsp) # 16-byte Spill
; SSE-NEXT:    movaps %xmm9, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm3[1]
; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm10, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm2[1]
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 272(%rdi), %xmm15
; SSE-NEXT:    movaps %xmm15, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]
; SSE-NEXT:    movaps 96(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
; SSE-NEXT:    movaps 224(%rdi), %xmm0
; SSE-NEXT:    movaps 160(%rdi), %xmm9
; SSE-NEXT:    movaps %xmm9, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
; SSE-NEXT:    movaps 352(%rdi), %xmm0
; SSE-NEXT:    movaps 288(%rdi), %xmm10
; SSE-NEXT:    movaps %xmm10, %xmm12
; SSE-NEXT:    movlhps {{.*#+}} xmm12 = xmm12[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT:    movaps 480(%rdi), %xmm0
; SSE-NEXT:    movaps 416(%rdi), %xmm8
; SSE-NEXT:    movaps %xmm8, %xmm11
; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
; SSE-NEXT:    movaps 112(%rdi), %xmm0
; SSE-NEXT:    movaps 48(%rdi), %xmm7
; SSE-NEXT:    movaps %xmm7, %xmm13
; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT:    movaps 240(%rdi), %xmm0
; SSE-NEXT:    movaps 176(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm5, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 304(%rdi), %xmm2
; SSE-NEXT:    movaps %xmm2, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; SSE-NEXT:    movaps 496(%rdi), %xmm0
; SSE-NEXT:    movaps 432(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rsi)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rcx)
; SSE-NEXT:    movaps %xmm15, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r8)
; SSE-NEXT:    movaps %xmm12, 32(%r9)
; SSE-NEXT:    movaps %xmm11, 48(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm8, 48(%rax)
; SSE-NEXT:    movaps %xmm10, 32(%rax)
; SSE-NEXT:    movaps %xmm9, 16(%rax)
; SSE-NEXT:    movaps %xmm14, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm3, 48(%rax)
; SSE-NEXT:    movaps %xmm4, 32(%rax)
; SSE-NEXT:    movaps %xmm6, 16(%rax)
; SSE-NEXT:    movaps %xmm13, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm1, 48(%rax)
; SSE-NEXT:    movaps %xmm2, 32(%rax)
; SSE-NEXT:    movaps %xmm5, 16(%rax)
; SSE-NEXT:    movaps %xmm7, (%rax)
; SSE-NEXT:    addq $152, %rsp
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i64_stride8_vf8:
; AVX:       # %bb.0:
; AVX-NEXT:    subq $184, %rsp
; AVX-NEXT:    vmovaps 192(%rdi), %ymm0
; AVX-NEXT:    vmovaps 128(%rdi), %ymm1
; AVX-NEXT:    vmovaps 448(%rdi), %ymm2
; AVX-NEXT:    vmovaps 384(%rdi), %ymm3
; AVX-NEXT:    vmovaps 320(%rdi), %xmm4
; AVX-NEXT:    vmovaps 256(%rdi), %xmm5
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 448(%rdi), %xmm7
; AVX-NEXT:    vmovaps 384(%rdi), %xmm8
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm8[0],xmm7[0]
; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 192(%rdi), %xmm9
; AVX-NEXT:    vmovaps 128(%rdi), %xmm10
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm10[0],xmm9[0]
; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 64(%rdi), %xmm11
; AVX-NEXT:    vmovaps (%rdi), %xmm12
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm12[0],xmm11[0]
; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
; AVX-NEXT:    vmovaps %xmm4, (%rsp) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm8[1],xmm7[1]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm10[1],xmm9[1]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm12[1],xmm11[1]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX-NEXT:    vmovaps 336(%rdi), %xmm10
; AVX-NEXT:    vmovaps 272(%rdi), %xmm11
; AVX-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm11[0],xmm10[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX-NEXT:    vmovaps 80(%rdi), %xmm12
; AVX-NEXT:    vmovaps 16(%rdi), %xmm13
; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm13[0],xmm12[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm11[1],xmm10[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
; AVX-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm13[1],xmm12[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 480(%rdi), %xmm0
; AVX-NEXT:    vmovaps 416(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 352(%rdi), %xmm2
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 288(%rdi), %xmm0
; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm0[0],xmm2[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 224(%rdi), %xmm0
; AVX-NEXT:    vmovaps 160(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm1[0],xmm0[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps 96(%rdi), %xmm1
; AVX-NEXT:    vmovaps 32(%rdi), %xmm2
; AVX-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm2[0],xmm1[0]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm11 = xmm2[1],xmm1[1]
; AVX-NEXT:    vmovaps 480(%rdi), %ymm6
; AVX-NEXT:    vmovaps 416(%rdi), %ymm5
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
; AVX-NEXT:    vmovaps 368(%rdi), %xmm4
; AVX-NEXT:    vmovaps 304(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm3[0],xmm4[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovaps 224(%rdi), %ymm9
; AVX-NEXT:    vmovaps 160(%rdi), %ymm2
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm2[0],ymm9[0],ymm2[2],ymm9[2]
; AVX-NEXT:    vmovaps 112(%rdi), %xmm1
; AVX-NEXT:    vmovaps 48(%rdi), %xmm0
; AVX-NEXT:    vmovlhps {{.*#+}} xmm14 = xmm0[0],xmm1[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm4[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm9[1],ymm2[3],ymm9[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rdx)
; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rdx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX-NEXT:    vmovaps %xmm8, (%r9)
; AVX-NEXT:    vmovaps %xmm12, 16(%r9)
; AVX-NEXT:    vmovaps %xmm15, 32(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%r9)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %xmm11, (%rax)
; AVX-NEXT:    vmovaps %xmm13, 16(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm7, (%rax)
; AVX-NEXT:    vmovaps %ymm10, 32(%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm0, (%rax)
; AVX-NEXT:    vmovaps %ymm3, 32(%rax)
; AVX-NEXT:    addq $184, %rsp
; AVX-NEXT:    vzeroupper
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i64_stride8_vf8:
; AVX2:       # %bb.0:
; AVX2-NEXT:    subq $136, %rsp
; AVX2-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 64(%rdi), %ymm8
; AVX2-NEXT:    vmovaps (%rdi), %ymm10
; AVX2-NEXT:    vmovaps 192(%rdi), %ymm11
; AVX2-NEXT:    vmovaps 128(%rdi), %ymm13
; AVX2-NEXT:    vmovaps 320(%rdi), %ymm7
; AVX2-NEXT:    vmovaps 256(%rdi), %ymm14
; AVX2-NEXT:    vmovaps 448(%rdi), %ymm12
; AVX2-NEXT:    vmovaps 384(%rdi), %ymm15
; AVX2-NEXT:    vmovaps 320(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 448(%rdi), %ymm1, %ymm2
; AVX2-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm4
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps (%rdi), %xmm3
; AVX2-NEXT:    vmovaps 64(%rdi), %xmm5
; AVX2-NEXT:    vinsertf128 $1, 192(%rdi), %ymm5, %ymm5
; AVX2-NEXT:    vinsertf128 $1, 128(%rdi), %ymm3, %ymm6
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX2-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm15[0],ymm12[0],ymm15[2],ymm12[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm14[0],ymm7[0],ymm14[2],ymm7[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm6[2,3],ymm5[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm13[0],ymm11[0],ymm13[2],ymm11[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm10[0],ymm8[0],ymm10[2],ymm8[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm6[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 288(%rdi), %ymm9
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm15[1],ymm12[1],ymm15[3],ymm12[3]
; AVX2-NEXT:    vmovaps 480(%rdi), %ymm12
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm14[1],ymm7[1],ymm14[3],ymm7[3]
; AVX2-NEXT:    vmovaps 32(%rdi), %xmm14
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm15[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm13[1],ymm11[1],ymm13[3],ymm11[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm10[1],ymm8[1],ymm10[3],ymm8[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm8 = ymm8[2,3],ymm11[2,3]
; AVX2-NEXT:    vmovaps 352(%rdi), %xmm10
; AVX2-NEXT:    vinsertf128 $1, 480(%rdi), %ymm10, %ymm11
; AVX2-NEXT:    vmovaps 288(%rdi), %xmm10
; AVX2-NEXT:    vinsertf128 $1, 416(%rdi), %ymm10, %ymm13
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm13[0],ymm11[0],ymm13[2],ymm11[2]
; AVX2-NEXT:    vmovaps 96(%rdi), %xmm15
; AVX2-NEXT:    vinsertf128 $1, 224(%rdi), %ymm15, %ymm15
; AVX2-NEXT:    vinsertf128 $1, 160(%rdi), %ymm14, %ymm14
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm13[1],ymm11[1],ymm13[3],ymm11[3]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm15[0],ymm14[2],ymm15[2]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm14 = ymm14[1],ymm15[1],ymm14[3],ymm15[3]
; AVX2-NEXT:    vmovaps 416(%rdi), %ymm15
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm12[0],ymm15[2],ymm12[2]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm9[0],ymm3[0],ymm9[2],ymm3[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm7 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovaps 96(%rdi), %ymm4
; AVX2-NEXT:    vmovaps 32(%rdi), %ymm2
; AVX2-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm4[0],ymm2[2],ymm4[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm15[1],ymm12[1],ymm15[3],ymm12[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm3[1],ymm9[3],ymm3[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm6 = ymm9[2,3],ymm6[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm4[1],ymm2[3],ymm4[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-NEXT:    vmovaps %ymm8, (%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-NEXT:    vmovaps %ymm13, (%r9)
; AVX2-NEXT:    vmovaps %ymm10, 32(%r9)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovaps %ymm14, (%rax)
; AVX2-NEXT:    vmovaps %ymm11, 32(%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovaps %ymm5, (%rax)
; AVX2-NEXT:    vmovaps %ymm7, 32(%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-NEXT:    vmovaps %ymm6, 32(%rax)
; AVX2-NEXT:    addq $136, %rsp
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i64_stride8_vf8:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    subq $136, %rsp
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm8
; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm10
; AVX2-FP-NEXT:    vmovaps 192(%rdi), %ymm11
; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm13
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm7
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %ymm14
; AVX2-FP-NEXT:    vmovaps 448(%rdi), %ymm12
; AVX2-FP-NEXT:    vmovaps 384(%rdi), %ymm15
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm1, %ymm2
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm4
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm3
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm5
; AVX2-FP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm5, %ymm5
; AVX2-FP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm3, %ymm6
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX2-FP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm15[0],ymm12[0],ymm15[2],ymm12[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm14[0],ymm7[0],ymm14[2],ymm7[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm6[2,3],ymm5[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm13[0],ymm11[0],ymm13[2],ymm11[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm10[0],ymm8[0],ymm10[2],ymm8[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm6[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm9
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm15[1],ymm12[1],ymm15[3],ymm12[3]
; AVX2-FP-NEXT:    vmovaps 480(%rdi), %ymm12
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm14[1],ymm7[1],ymm14[3],ymm7[3]
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm14
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm15[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm13[1],ymm11[1],ymm13[3],ymm11[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm10[1],ymm8[1],ymm10[3],ymm8[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm8 = ymm8[2,3],ymm11[2,3]
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %xmm10
; AVX2-FP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm10, %ymm11
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %xmm10
; AVX2-FP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm10, %ymm13
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm13[0],ymm11[0],ymm13[2],ymm11[2]
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm15
; AVX2-FP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm15, %ymm15
; AVX2-FP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm14, %ymm14
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm13[1],ymm11[1],ymm13[3],ymm11[3]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm15[0],ymm14[2],ymm15[2]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm14 = ymm14[1],ymm15[1],ymm14[3],ymm15[3]
; AVX2-FP-NEXT:    vmovaps 416(%rdi), %ymm15
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm12[0],ymm15[2],ymm12[2]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm9[0],ymm3[0],ymm9[2],ymm3[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm7 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm4[0],ymm2[2],ymm4[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm15[1],ymm12[1],ymm15[3],ymm12[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm3[1],ymm9[3],ymm3[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm6 = ymm9[2,3],ymm6[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm4[1],ymm2[3],ymm4[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-FP-NEXT:    vmovaps %ymm8, (%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-FP-NEXT:    vmovaps %ymm13, (%r9)
; AVX2-FP-NEXT:    vmovaps %ymm10, 32(%r9)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovaps %ymm14, (%rax)
; AVX2-FP-NEXT:    vmovaps %ymm11, 32(%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovaps %ymm5, (%rax)
; AVX2-FP-NEXT:    vmovaps %ymm7, 32(%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FP-NEXT:    vmovaps %ymm6, 32(%rax)
; AVX2-FP-NEXT:    addq $136, %rsp
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i64_stride8_vf8:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    subq $136, %rsp
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm8
; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm10
; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %ymm11
; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm13
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm7
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %ymm14
; AVX2-FCP-NEXT:    vmovaps 448(%rdi), %ymm12
; AVX2-FCP-NEXT:    vmovaps 384(%rdi), %ymm15
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm1, %ymm2
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm4
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm3
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm5
; AVX2-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm5, %ymm5
; AVX2-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm3, %ymm6
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX2-FCP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm15[0],ymm12[0],ymm15[2],ymm12[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm14[0],ymm7[0],ymm14[2],ymm7[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm6[2,3],ymm5[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm13[0],ymm11[0],ymm13[2],ymm11[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm10[0],ymm8[0],ymm10[2],ymm8[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm6[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm9
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm15[1],ymm12[1],ymm15[3],ymm12[3]
; AVX2-FCP-NEXT:    vmovaps 480(%rdi), %ymm12
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm14[1],ymm7[1],ymm14[3],ymm7[3]
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm14
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm15[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm13[1],ymm11[1],ymm13[3],ymm11[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm10[1],ymm8[1],ymm10[3],ymm8[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm8 = ymm8[2,3],ymm11[2,3]
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %xmm10
; AVX2-FCP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm10, %ymm11
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %xmm10
; AVX2-FCP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm10, %ymm13
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm13[0],ymm11[0],ymm13[2],ymm11[2]
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm15
; AVX2-FCP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm15, %ymm15
; AVX2-FCP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm14, %ymm14
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm11 = ymm13[1],ymm11[1],ymm13[3],ymm11[3]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm15[0],ymm14[2],ymm15[2]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm14 = ymm14[1],ymm15[1],ymm14[3],ymm15[3]
; AVX2-FCP-NEXT:    vmovaps 416(%rdi), %ymm15
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm12[0],ymm15[2],ymm12[2]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm9[0],ymm3[0],ymm9[2],ymm3[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm7 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm2[0],ymm4[0],ymm2[2],ymm4[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm15[1],ymm12[1],ymm15[3],ymm12[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm9[1],ymm3[1],ymm9[3],ymm3[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm6 = ymm9[2,3],ymm6[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm4[1],ymm2[3],ymm4[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-FCP-NEXT:    vmovaps %ymm8, (%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-FCP-NEXT:    vmovaps %ymm13, (%r9)
; AVX2-FCP-NEXT:    vmovaps %ymm10, 32(%r9)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovaps %ymm14, (%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm11, 32(%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovaps %ymm5, (%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm7, 32(%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm6, 32(%rax)
; AVX2-FCP-NEXT:    addq $136, %rsp
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i64_stride8_vf8:
; AVX512:       # %bb.0:
; AVX512-NEXT:    pushq %rbx
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512-NEXT:    movb $-64, %bl
; AVX512-NEXT:    kmovw %ebx, %k1
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512-NEXT:    popq %rbx
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i64_stride8_vf8:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    pushq %rbx
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512-FCP-NEXT:    movb $-64, %bl
; AVX512-FCP-NEXT:    kmovw %ebx, %k1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512-FCP-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512-FCP-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512-FCP-NEXT:    popq %rbx
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i64_stride8_vf8:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    pushq %rbx
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512DQ-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512DQ-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-NEXT:    movb $-64, %bl
; AVX512DQ-NEXT:    kmovw %ebx, %k1
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512DQ-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512DQ-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512DQ-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512DQ-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512DQ-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512DQ-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512DQ-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512DQ-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512DQ-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512DQ-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512DQ-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-NEXT:    popq %rbx
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i64_stride8_vf8:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    pushq %rbx
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512DQ-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-FCP-NEXT:    movb $-64, %bl
; AVX512DQ-FCP-NEXT:    kmovw %ebx, %k1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512DQ-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512DQ-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512DQ-FCP-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512DQ-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512DQ-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512DQ-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512DQ-FCP-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512DQ-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-FCP-NEXT:    popq %rbx
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride8_vf8:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    pushq %rbx
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512BW-NEXT:    movb $-64, %bl
; AVX512BW-NEXT:    kmovd %ebx, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512BW-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512BW-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512BW-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512BW-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512BW-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512BW-NEXT:    popq %rbx
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i64_stride8_vf8:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    pushq %rbx
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512BW-FCP-NEXT:    movb $-64, %bl
; AVX512BW-FCP-NEXT:    kmovd %ebx, %k1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512BW-FCP-NEXT:    popq %rbx
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i64_stride8_vf8:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    pushq %rbx
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-BW-NEXT:    movb $-64, %bl
; AVX512DQ-BW-NEXT:    kmovd %ebx, %k1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512DQ-BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512DQ-BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512DQ-BW-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-BW-NEXT:    popq %rbx
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i64_stride8_vf8:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    pushq %rbx
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm5
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-BW-FCP-NEXT:    movb $-64, %bl
; AVX512DQ-BW-FCP-NEXT:    kmovd %ebx, %k1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm8 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm10
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm10, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm9, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm11, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm9 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm11, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm10, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %ymm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm8 = ymm8[2,3],ymm13[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm10, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm8, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm11 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm12 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm8, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm11 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm11, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm11, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 = zmm0[0],zmm3[0],zmm0[2],zmm3[2],zmm0[4],zmm3[4],zmm0[6],zmm3[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm13 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm12, %zmm13, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm12, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm4, %zmm5, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm0[1],zmm3[1],zmm0[3],zmm3[3],zmm0[5],zmm3[5],zmm0[7],zmm3[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm14, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm13 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm13, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm13, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm3, %zmm0, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm13 {%k1} = zmm5[0],zmm4[0],zmm5[2],zmm4[2],zmm5[4],zmm4[4],zmm5[6],zmm4[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm14, %zmm13, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm14 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm14, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm14, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm14, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm5[1],zmm4[1],zmm5[3],zmm4[3],zmm5[5],zmm4[5],zmm5[7],zmm4[7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    popq %rbx
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <64 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56>
  %strided.vec1 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57>
  %strided.vec2 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58>
  %strided.vec3 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59>
  %strided.vec4 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60>
  %strided.vec5 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61>
  %strided.vec6 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62>
  %strided.vec7 = shufflevector <64 x i64> %wide.vec, <64 x i64> poison, <8 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63>
  store <8 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <8 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <8 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <8 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <8 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <8 x i64> %strided.vec5, ptr %out.vec5, align 64
  store <8 x i64> %strided.vec6, ptr %out.vec6, align 64
  store <8 x i64> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i64_stride8_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf16:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $664, %rsp # imm = 0x298
; SSE-NEXT:    movaps 832(%rdi), %xmm0
; SSE-NEXT:    movaps 320(%rdi), %xmm1
; SSE-NEXT:    movaps 256(%rdi), %xmm8
; SSE-NEXT:    movaps 960(%rdi), %xmm2
; SSE-NEXT:    movaps 896(%rdi), %xmm9
; SSE-NEXT:    movaps 448(%rdi), %xmm3
; SSE-NEXT:    movaps 384(%rdi), %xmm10
; SSE-NEXT:    movaps 576(%rdi), %xmm4
; SSE-NEXT:    movaps 512(%rdi), %xmm11
; SSE-NEXT:    movaps 64(%rdi), %xmm5
; SSE-NEXT:    movaps (%rdi), %xmm12
; SSE-NEXT:    movaps 704(%rdi), %xmm6
; SSE-NEXT:    movaps 640(%rdi), %xmm13
; SSE-NEXT:    movaps 192(%rdi), %xmm7
; SSE-NEXT:    movaps 128(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]
; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm7[1]
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm13, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]
; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm12, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm5[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm5
; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm4[0]
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm4[1]
; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm10, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm3[1]
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm9, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm2[1]
; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 768(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 16(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 208(%rdi), %xmm0
; SSE-NEXT:    movaps 144(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 336(%rdi), %xmm0
; SSE-NEXT:    movaps 272(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 464(%rdi), %xmm0
; SSE-NEXT:    movaps 400(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 592(%rdi), %xmm0
; SSE-NEXT:    movaps 528(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 720(%rdi), %xmm0
; SSE-NEXT:    movaps 656(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 848(%rdi), %xmm0
; SSE-NEXT:    movaps 784(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 976(%rdi), %xmm0
; SSE-NEXT:    movaps 912(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 96(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 224(%rdi), %xmm0
; SSE-NEXT:    movaps 160(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 352(%rdi), %xmm0
; SSE-NEXT:    movaps 288(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 480(%rdi), %xmm0
; SSE-NEXT:    movaps 416(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, (%rsp) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 608(%rdi), %xmm0
; SSE-NEXT:    movaps 544(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 736(%rdi), %xmm0
; SSE-NEXT:    movaps 672(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 864(%rdi), %xmm0
; SSE-NEXT:    movaps 800(%rdi), %xmm12
; SSE-NEXT:    movaps %xmm12, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]
; SSE-NEXT:    movaps 992(%rdi), %xmm0
; SSE-NEXT:    movaps 928(%rdi), %xmm9
; SSE-NEXT:    movaps %xmm9, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
; SSE-NEXT:    movaps 112(%rdi), %xmm0
; SSE-NEXT:    movaps 48(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 240(%rdi), %xmm0
; SSE-NEXT:    movaps 176(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 304(%rdi), %xmm13
; SSE-NEXT:    movaps %xmm13, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]
; SSE-NEXT:    movaps 496(%rdi), %xmm0
; SSE-NEXT:    movaps 432(%rdi), %xmm10
; SSE-NEXT:    movaps %xmm10, %xmm11
; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
; SSE-NEXT:    movaps 624(%rdi), %xmm0
; SSE-NEXT:    movaps 560(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm5, %xmm8
; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT:    movaps 752(%rdi), %xmm0
; SSE-NEXT:    movaps 688(%rdi), %xmm6
; SSE-NEXT:    movaps %xmm6, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]
; SSE-NEXT:    movaps 880(%rdi), %xmm0
; SSE-NEXT:    movaps 816(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps 1008(%rdi), %xmm0
; SSE-NEXT:    movaps 944(%rdi), %xmm2
; SSE-NEXT:    movaps %xmm2, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r9)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm9, 112(%rax)
; SSE-NEXT:    movaps %xmm12, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm3, 112(%rax)
; SSE-NEXT:    movaps %xmm4, 96(%rax)
; SSE-NEXT:    movaps %xmm7, 80(%rax)
; SSE-NEXT:    movaps %xmm8, 64(%rax)
; SSE-NEXT:    movaps %xmm11, 48(%rax)
; SSE-NEXT:    movaps %xmm15, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm2, 112(%rax)
; SSE-NEXT:    movaps %xmm1, 96(%rax)
; SSE-NEXT:    movaps %xmm6, 80(%rax)
; SSE-NEXT:    movaps %xmm5, 64(%rax)
; SSE-NEXT:    movaps %xmm10, 48(%rax)
; SSE-NEXT:    movaps %xmm13, 32(%rax)
; SSE-NEXT:    movaps %xmm14, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    addq $664, %rsp # imm = 0x298
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i64_stride8_vf16:
; AVX:       # %bb.0:
; AVX-NEXT:    subq $808, %rsp # imm = 0x328
; AVX-NEXT:    vmovaps 448(%rdi), %xmm0
; AVX-NEXT:    vmovaps 384(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 320(%rdi), %xmm2
; AVX-NEXT:    vmovaps 256(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 832(%rdi), %xmm4
; AVX-NEXT:    vmovaps 768(%rdi), %xmm5
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 960(%rdi), %xmm6
; AVX-NEXT:    vmovaps 896(%rdi), %xmm7
; AVX-NEXT:    vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm6[0]
; AVX-NEXT:    vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 704(%rdi), %xmm8
; AVX-NEXT:    vmovaps 640(%rdi), %xmm9
; AVX-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm9[0],xmm8[0]
; AVX-NEXT:    vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 576(%rdi), %xmm10
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 512(%rdi), %xmm0
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm2[1]
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm10[0]
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm6[1]
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 64(%rdi), %xmm1
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm5[1],xmm4[1]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 192(%rdi), %xmm2
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm9[1],xmm8[1]
; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 128(%rdi), %xmm3
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm10[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm3[0],xmm2[0]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps (%rdi), %xmm0
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 448(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 384(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX-NEXT:    vmovaps 336(%rdi), %xmm3
; AVX-NEXT:    vmovaps 272(%rdi), %xmm4
; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm3[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm2[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 960(%rdi), %ymm2
; AVX-NEXT:    vmovaps 896(%rdi), %ymm5
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
; AVX-NEXT:    vmovaps 848(%rdi), %xmm7
; AVX-NEXT:    vmovaps 784(%rdi), %xmm8
; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm7[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm6[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 192(%rdi), %ymm6
; AVX-NEXT:    vmovaps 128(%rdi), %ymm9
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm6[0],ymm9[2],ymm6[2]
; AVX-NEXT:    vmovaps 80(%rdi), %xmm11
; AVX-NEXT:    vmovaps 16(%rdi), %xmm12
; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm12[0],xmm11[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm10[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 704(%rdi), %ymm10
; AVX-NEXT:    vmovaps 640(%rdi), %ymm13
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm13[0],ymm10[0],ymm13[2],ymm10[2]
; AVX-NEXT:    vmovaps 592(%rdi), %xmm15
; AVX-NEXT:    vmovaps 528(%rdi), %xmm0
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm15[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX-NEXT:    # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm8[1],xmm7[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm6[1],ymm9[3],ymm6[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm12[1],xmm11[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm10[1],ymm13[3],ymm10[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 480(%rdi), %xmm0
; AVX-NEXT:    vmovaps 416(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 992(%rdi), %xmm0
; AVX-NEXT:    vmovaps 928(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 224(%rdi), %xmm0
; AVX-NEXT:    vmovaps 160(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
; AVX-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 736(%rdi), %xmm0
; AVX-NEXT:    vmovaps 672(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 480(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 416(%rdi), %ymm14
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm14[0],ymm0[0],ymm14[2],ymm0[2]
; AVX-NEXT:    vmovaps 368(%rdi), %xmm12
; AVX-NEXT:    vmovaps 304(%rdi), %xmm6
; AVX-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm6[0],xmm12[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 992(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 928(%rdi), %ymm15
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX-NEXT:    vmovaps 880(%rdi), %xmm11
; AVX-NEXT:    vmovaps 816(%rdi), %xmm10
; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm10[0],xmm11[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 224(%rdi), %ymm9
; AVX-NEXT:    vmovaps 160(%rdi), %ymm4
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm4[0],ymm9[0],ymm4[2],ymm9[2]
; AVX-NEXT:    vmovaps 112(%rdi), %xmm3
; AVX-NEXT:    vmovaps 48(%rdi), %xmm0
; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm0[0],xmm3[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm13[0,1,2,3],ymm8[4,5,6,7]
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 736(%rdi), %ymm7
; AVX-NEXT:    vmovaps 672(%rdi), %ymm5
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm5[0],ymm7[0],ymm5[2],ymm7[2]
; AVX-NEXT:    vmovaps 624(%rdi), %xmm2
; AVX-NEXT:    vmovaps 560(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm4[1],ymm9[1],ymm4[3],ymm9[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm14[1],mem[1],ymm14[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm6[1],xmm12[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm3[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm7[1],ymm5[3],ymm7[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm10[1],xmm11[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rdx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%r9)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rax)
; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm8, 64(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX-NEXT:    vmovaps %ymm4, 64(%rax)
; AVX-NEXT:    vmovaps %ymm6, 32(%rax)
; AVX-NEXT:    vmovaps %ymm9, (%rax)
; AVX-NEXT:    addq $808, %rsp # imm = 0x328
; AVX-NEXT:    vzeroupper
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i64_stride8_vf16:
; AVX2:       # %bb.0:
; AVX2-NEXT:    subq $808, %rsp # imm = 0x328
; AVX2-NEXT:    vmovaps 832(%rdi), %ymm1
; AVX2-NEXT:    vmovaps 768(%rdi), %ymm2
; AVX2-NEXT:    vmovaps 320(%rdi), %ymm3
; AVX2-NEXT:    vmovaps 256(%rdi), %ymm4
; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 448(%rdi), %ymm5
; AVX2-NEXT:    vmovaps 384(%rdi), %ymm13
; AVX2-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 320(%rdi), %xmm6
; AVX2-NEXT:    vinsertf128 $1, 448(%rdi), %ymm6, %ymm6
; AVX2-NEXT:    vmovaps 256(%rdi), %xmm7
; AVX2-NEXT:    vinsertf128 $1, 384(%rdi), %ymm7, %ymm7
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 832(%rdi), %xmm8
; AVX2-NEXT:    vinsertf128 $1, 960(%rdi), %ymm8, %ymm8
; AVX2-NEXT:    vmovaps 768(%rdi), %xmm9
; AVX2-NEXT:    vinsertf128 $1, 896(%rdi), %ymm9, %ymm9
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps (%rdi), %xmm10
; AVX2-NEXT:    vmovaps 64(%rdi), %xmm11
; AVX2-NEXT:    vinsertf128 $1, 192(%rdi), %ymm11, %ymm11
; AVX2-NEXT:    vinsertf128 $1, 128(%rdi), %ymm10, %ymm10
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 576(%rdi), %xmm12
; AVX2-NEXT:    vinsertf128 $1, 704(%rdi), %ymm12, %ymm12
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 512(%rdi), %xmm6
; AVX2-NEXT:    vinsertf128 $1, 640(%rdi), %ymm6, %ymm6
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm12[0],ymm6[2],ymm12[2]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm12[1],ymm6[3],ymm12[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm13[0],ymm5[0],ymm13[2],ymm5[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm4[0],ymm3[0],ymm4[2],ymm3[2]
; AVX2-NEXT:    vmovaps %ymm3, %ymm4
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm6[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 960(%rdi), %ymm6
; AVX2-NEXT:    vmovaps 896(%rdi), %ymm7
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-NEXT:    vmovaps %ymm2, %ymm3
; AVX2-NEXT:    vmovaps %ymm1, %ymm2
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 64(%rdi), %ymm8
; AVX2-NEXT:    vmovaps (%rdi), %ymm9
; AVX2-NEXT:    vmovaps 192(%rdi), %ymm10
; AVX2-NEXT:    vmovaps 128(%rdi), %ymm11
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm12[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 576(%rdi), %ymm12
; AVX2-NEXT:    vmovaps 512(%rdi), %ymm13
; AVX2-NEXT:    vmovaps 704(%rdi), %ymm14
; AVX2-NEXT:    vmovaps 640(%rdi), %ymm15
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm13[0],ymm12[0],ymm13[2],ymm12[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm5[1],ymm0[3],ymm5[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm12[1],ymm13[3],ymm12[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 288(%rdi), %ymm15
; AVX2-NEXT:    vmovaps 480(%rdi), %ymm12
; AVX2-NEXT:    vmovaps 416(%rdi), %ymm10
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm10[0],ymm12[0],ymm10[2],ymm12[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm4[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 864(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 800(%rdi), %ymm14
; AVX2-NEXT:    vmovaps 992(%rdi), %ymm11
; AVX2-NEXT:    vmovaps 928(%rdi), %ymm7
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm11[0],ymm7[2],ymm11[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm14[0],ymm0[0],ymm14[2],ymm0[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 96(%rdi), %ymm9
; AVX2-NEXT:    vmovaps 32(%rdi), %ymm6
; AVX2-NEXT:    vmovaps 224(%rdi), %ymm3
; AVX2-NEXT:    vmovaps 160(%rdi), %ymm2
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm8[2,3]
; AVX2-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-NEXT:    vmovaps 608(%rdi), %ymm5
; AVX2-NEXT:    vmovaps 544(%rdi), %ymm4
; AVX2-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm8 = ymm13[2,3],ymm8[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm9 = ymm3[2,3],ymm2[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm12[1],ymm10[3],ymm12[3]
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm3 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm3 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm3[2,3],ymm2[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm7[1],ymm11[1],ymm7[3],ymm11[3]
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm3 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm3 = ymm14[1],mem[1],ymm14[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 64(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, (%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 96(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 32(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 64(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, (%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 96(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 32(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 64(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, (%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 96(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 32(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 64(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, (%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 96(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 32(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 64(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, (%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 96(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 32(%r9)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 64(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, (%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 32(%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovaps %ymm8, 64(%rax)
; AVX2-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, (%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm3, 32(%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-NEXT:    vmovaps %ymm2, 32(%rax)
; AVX2-NEXT:    vmovaps %ymm9, (%rax)
; AVX2-NEXT:    addq $808, %rsp # imm = 0x328
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i64_stride8_vf16:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    subq $808, %rsp # imm = 0x328
; AVX2-FP-NEXT:    vmovaps 832(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovaps 768(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 448(%rdi), %ymm5
; AVX2-FP-NEXT:    vmovaps 384(%rdi), %ymm13
; AVX2-FP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %xmm6
; AVX2-FP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm6, %ymm6
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm7
; AVX2-FP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm7, %ymm7
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 832(%rdi), %xmm8
; AVX2-FP-NEXT:    vinsertf128 $1, 960(%rdi), %ymm8, %ymm8
; AVX2-FP-NEXT:    vmovaps 768(%rdi), %xmm9
; AVX2-FP-NEXT:    vinsertf128 $1, 896(%rdi), %ymm9, %ymm9
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm10
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm11
; AVX2-FP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm11, %ymm11
; AVX2-FP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm10, %ymm10
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 576(%rdi), %xmm12
; AVX2-FP-NEXT:    vinsertf128 $1, 704(%rdi), %ymm12, %ymm12
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 512(%rdi), %xmm6
; AVX2-FP-NEXT:    vinsertf128 $1, 640(%rdi), %ymm6, %ymm6
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm12[0],ymm6[2],ymm12[2]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm12[1],ymm6[3],ymm12[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm13[0],ymm5[0],ymm13[2],ymm5[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm4[0],ymm3[0],ymm4[2],ymm3[2]
; AVX2-FP-NEXT:    vmovaps %ymm3, %ymm4
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm6[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 960(%rdi), %ymm6
; AVX2-FP-NEXT:    vmovaps 896(%rdi), %ymm7
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-FP-NEXT:    vmovaps %ymm2, %ymm3
; AVX2-FP-NEXT:    vmovaps %ymm1, %ymm2
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm8
; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm9
; AVX2-FP-NEXT:    vmovaps 192(%rdi), %ymm10
; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm11
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm12[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 576(%rdi), %ymm12
; AVX2-FP-NEXT:    vmovaps 512(%rdi), %ymm13
; AVX2-FP-NEXT:    vmovaps 704(%rdi), %ymm14
; AVX2-FP-NEXT:    vmovaps 640(%rdi), %ymm15
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm13[0],ymm12[0],ymm13[2],ymm12[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm5[1],ymm0[3],ymm5[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm12[1],ymm13[3],ymm12[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm15
; AVX2-FP-NEXT:    vmovaps 480(%rdi), %ymm12
; AVX2-FP-NEXT:    vmovaps 416(%rdi), %ymm10
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm10[0],ymm12[0],ymm10[2],ymm12[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm4[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 864(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 800(%rdi), %ymm14
; AVX2-FP-NEXT:    vmovaps 992(%rdi), %ymm11
; AVX2-FP-NEXT:    vmovaps 928(%rdi), %ymm7
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm11[0],ymm7[2],ymm11[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm14[0],ymm0[0],ymm14[2],ymm0[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm9
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm6
; AVX2-FP-NEXT:    vmovaps 224(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm2
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm8[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 608(%rdi), %ymm5
; AVX2-FP-NEXT:    vmovaps 544(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm8 = ymm13[2,3],ymm8[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm9 = ymm3[2,3],ymm2[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm12[1],ymm10[3],ymm12[3]
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm3 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm3 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm3[2,3],ymm2[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm7[1],ymm11[1],ymm7[3],ymm11[3]
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm3 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm3 = ymm14[1],mem[1],ymm14[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, (%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 96(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, (%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 96(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, (%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 96(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, (%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 96(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, (%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 96(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%r9)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, (%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovaps %ymm8, 64(%rax)
; AVX2-FP-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, (%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm3, 32(%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm2, 32(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm9, (%rax)
; AVX2-FP-NEXT:    addq $808, %rsp # imm = 0x328
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i64_stride8_vf16:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    subq $808, %rsp # imm = 0x328
; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 448(%rdi), %ymm5
; AVX2-FCP-NEXT:    vmovaps 384(%rdi), %ymm13
; AVX2-FCP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %xmm6
; AVX2-FCP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm6, %ymm6
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm7
; AVX2-FCP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm7, %ymm7
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %xmm8
; AVX2-FCP-NEXT:    vinsertf128 $1, 960(%rdi), %ymm8, %ymm8
; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %xmm9
; AVX2-FCP-NEXT:    vinsertf128 $1, 896(%rdi), %ymm9, %ymm9
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm10
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm11
; AVX2-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm11, %ymm11
; AVX2-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm10, %ymm10
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %xmm12
; AVX2-FCP-NEXT:    vinsertf128 $1, 704(%rdi), %ymm12, %ymm12
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %xmm6
; AVX2-FCP-NEXT:    vinsertf128 $1, 640(%rdi), %ymm6, %ymm6
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm12[0],ymm6[2],ymm12[2]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm12[1],ymm6[3],ymm12[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm13[0],ymm5[0],ymm13[2],ymm5[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm4[0],ymm3[0],ymm4[2],ymm3[2]
; AVX2-FCP-NEXT:    vmovaps %ymm3, %ymm4
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm6[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 960(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovaps 896(%rdi), %ymm7
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-FCP-NEXT:    vmovaps %ymm2, %ymm3
; AVX2-FCP-NEXT:    vmovaps %ymm1, %ymm2
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm8
; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm9
; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %ymm10
; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm11
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm12[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %ymm12
; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %ymm13
; AVX2-FCP-NEXT:    vmovaps 704(%rdi), %ymm14
; AVX2-FCP-NEXT:    vmovaps 640(%rdi), %ymm15
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm13[0],ymm12[0],ymm13[2],ymm12[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm5[1],ymm0[3],ymm5[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm12[1],ymm13[3],ymm12[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm15
; AVX2-FCP-NEXT:    vmovaps 480(%rdi), %ymm12
; AVX2-FCP-NEXT:    vmovaps 416(%rdi), %ymm10
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm10[0],ymm12[0],ymm10[2],ymm12[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm4[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %ymm14
; AVX2-FCP-NEXT:    vmovaps 992(%rdi), %ymm11
; AVX2-FCP-NEXT:    vmovaps 928(%rdi), %ymm7
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm11[0],ymm7[2],ymm11[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm14[0],ymm0[0],ymm14[2],ymm0[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm8[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm9
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm2
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm8[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %ymm5
; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm8 = ymm13[2,3],ymm8[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm9 = ymm3[2,3],ymm2[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm10[1],ymm12[1],ymm10[3],ymm12[3]
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm3 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm3 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm3[2,3],ymm2[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm7[1],ymm11[1],ymm7[3],ymm11[3]
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm3 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm3 = ymm14[1],mem[1],ymm14[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, (%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 96(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, (%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 96(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, (%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 96(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, (%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 96(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, (%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 96(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%r9)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, (%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovaps %ymm8, 64(%rax)
; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, (%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm3, 32(%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm9, (%rax)
; AVX2-FCP-NEXT:    addq $808, %rsp # imm = 0x328
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i64_stride8_vf16:
; AVX512:       # %bb.0:
; AVX512-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512-NEXT:    vmovaps (%rdi), %zmm0
; AVX512-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512-NEXT:    movb $-64, %al
; AVX512-NEXT:    kmovw %eax, %k1
; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i64_stride8_vf16:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512-FCP-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512-FCP-NEXT:    vmovaps (%rdi), %zmm0
; AVX512-FCP-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512-FCP-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512-FCP-NEXT:    movb $-64, %al
; AVX512-FCP-NEXT:    kmovw %eax, %k1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512-FCP-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512-FCP-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512-FCP-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512-FCP-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512-FCP-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i64_stride8_vf16:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512DQ-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512DQ-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512DQ-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512DQ-NEXT:    vmovaps (%rdi), %zmm0
; AVX512DQ-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512DQ-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512DQ-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512DQ-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512DQ-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512DQ-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512DQ-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512DQ-NEXT:    movb $-64, %al
; AVX512DQ-NEXT:    kmovw %eax, %k1
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512DQ-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512DQ-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512DQ-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512DQ-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512DQ-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512DQ-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512DQ-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512DQ-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512DQ-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512DQ-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512DQ-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512DQ-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512DQ-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512DQ-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512DQ-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512DQ-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512DQ-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512DQ-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512DQ-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512DQ-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512DQ-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i64_stride8_vf16:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512DQ-FCP-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512DQ-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512DQ-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512DQ-FCP-NEXT:    vmovaps (%rdi), %zmm0
; AVX512DQ-FCP-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512DQ-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512DQ-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512DQ-FCP-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512DQ-FCP-NEXT:    movb $-64, %al
; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512DQ-FCP-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512DQ-FCP-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512DQ-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512DQ-FCP-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512DQ-FCP-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512DQ-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512DQ-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512DQ-FCP-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride8_vf16:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512BW-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512BW-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512BW-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512BW-NEXT:    vmovaps (%rdi), %zmm0
; AVX512BW-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512BW-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512BW-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512BW-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512BW-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512BW-NEXT:    movb $-64, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512BW-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512BW-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512BW-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512BW-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512BW-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512BW-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512BW-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512BW-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512BW-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512BW-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512BW-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i64_stride8_vf16:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512BW-FCP-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512BW-FCP-NEXT:    vmovaps (%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512BW-FCP-NEXT:    movb $-64, %al
; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512BW-FCP-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512BW-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512BW-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512BW-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512BW-FCP-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i64_stride8_vf16:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512DQ-BW-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512DQ-BW-NEXT:    vmovaps (%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512DQ-BW-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512DQ-BW-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512DQ-BW-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512DQ-BW-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512DQ-BW-NEXT:    movb $-64, %al
; AVX512DQ-BW-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512DQ-BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512DQ-BW-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512DQ-BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512DQ-BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512DQ-BW-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512DQ-BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512DQ-BW-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512DQ-BW-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i64_stride8_vf16:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    subq $264, %rsp # imm = 0x108
; AVX512DQ-BW-FCP-NEXT:    vmovaps 576(%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovaps 512(%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovaps (%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovups %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm31
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm30
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm28
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm15
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm29 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    # zmm29 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm20
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm29, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm18
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm29, %zmm18
; AVX512DQ-BW-FCP-NEXT:    movb $-64, %al
; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm18 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm19 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm19, %zmm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm19, %zmm21
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm21 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %ymm22
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %ymm23
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm23[0],ymm22[0],ymm23[2],ymm22[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm24
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %ymm25
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm12 = ymm25[0],ymm24[0],ymm25[2],ymm24[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm12[2,3],ymm11[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm21, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm19, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm19 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 704(%rdi), %ymm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 640(%rdi), %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm12[0],ymm11[0],ymm12[2],ymm11[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm26
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm27
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm27[0],ymm26[0],ymm27[2],ymm26[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm13[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm20
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm13 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm23[1],ymm22[1],ymm23[3],ymm22[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm23
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm21, %zmm23
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm25[1],ymm24[1],ymm25[3],ymm24[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm20
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm12[1],ymm11[1],ymm12[3],ymm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm27[1],ymm26[1],ymm27[3],ymm26[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm22
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm5[0],zmm16[0],zmm5[2],zmm16[2],zmm5[4],zmm16[4],zmm5[6],zmm16[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm24
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm14[0],zmm10[0],zmm14[2],zmm10[2],zmm14[4],zmm10[4],zmm14[6],zmm10[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm27 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm27, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm27, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 = zmm5[1],zmm16[1],zmm5[3],zmm16[3],zmm5[5],zmm16[5],zmm5[7],zmm16[7]
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm26 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm26, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm1 {%k1} = zmm15[0],zmm3[0],zmm15[2],zmm3[2],zmm15[4],zmm3[4],zmm15[6],zmm3[6]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm15[1],zmm3[1],zmm15[3],zmm3[3],zmm15[5],zmm3[5],zmm15[7],zmm3[7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm21, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm27, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm0, %zmm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm27, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm27, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm30, %zmm28, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 = zmm14[1],zmm10[1],zmm14[3],zmm10[3],zmm14[5],zmm10[5],zmm14[7],zmm10[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm9 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm9, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm26, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm26, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm29, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm29
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm26, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm26, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm21, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm21
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm10, %zmm14, %zmm26
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm2, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm10 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm10, %zmm18, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm29 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 576(%rdi), %xmm1
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm1, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm26 {%k1} = zmm28[0],zmm30[0],zmm28[2],zmm30[2],zmm28[4],zmm30[4],zmm28[6],zmm30[6]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 {%k1} = zmm28[1],zmm30[1],zmm28[3],zmm30[3],zmm28[5],zmm30[5],zmm28[7],zmm30[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm29, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm23 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm23, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm21 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm21, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm26, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm5, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm14, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, 64(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 64(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, 64(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, 64(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, 64(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, (%r9)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, (%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)
; AVX512DQ-BW-FCP-NEXT:    addq $264, %rsp # imm = 0x108
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <128 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120>
  %strided.vec1 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121>
  %strided.vec2 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122>
  %strided.vec3 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123>
  %strided.vec4 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124>
  %strided.vec5 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125>
  %strided.vec6 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126>
  %strided.vec7 = shufflevector <128 x i64> %wide.vec, <128 x i64> poison, <16 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127>
  store <16 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <16 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <16 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <16 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <16 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <16 x i64> %strided.vec5, ptr %out.vec5, align 64
  store <16 x i64> %strided.vec6, ptr %out.vec6, align 64
  store <16 x i64> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i64_stride8_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf32:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $1688, %rsp # imm = 0x698
; SSE-NEXT:    movaps 832(%rdi), %xmm0
; SSE-NEXT:    movaps 320(%rdi), %xmm2
; SSE-NEXT:    movaps 256(%rdi), %xmm8
; SSE-NEXT:    movaps 960(%rdi), %xmm1
; SSE-NEXT:    movaps 896(%rdi), %xmm10
; SSE-NEXT:    movaps 448(%rdi), %xmm4
; SSE-NEXT:    movaps 384(%rdi), %xmm9
; SSE-NEXT:    movaps 576(%rdi), %xmm3
; SSE-NEXT:    movaps 512(%rdi), %xmm12
; SSE-NEXT:    movaps 64(%rdi), %xmm6
; SSE-NEXT:    movaps (%rdi), %xmm11
; SSE-NEXT:    movaps 704(%rdi), %xmm5
; SSE-NEXT:    movaps 640(%rdi), %xmm14
; SSE-NEXT:    movaps 192(%rdi), %xmm7
; SSE-NEXT:    movaps 128(%rdi), %xmm13
; SSE-NEXT:    movaps %xmm13, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]
; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm7[1]
; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm6[1]
; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm9, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm4[0]
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm4[1]
; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm2[0]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm2[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm14, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm5[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm5[1]
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm12, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm3[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm3[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm10, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm1[1]
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 768(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1216(%rdi), %xmm0
; SSE-NEXT:    movaps 1152(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1088(%rdi), %xmm0
; SSE-NEXT:    movaps 1024(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1472(%rdi), %xmm0
; SSE-NEXT:    movaps 1408(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1344(%rdi), %xmm0
; SSE-NEXT:    movaps 1280(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1728(%rdi), %xmm0
; SSE-NEXT:    movaps 1664(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1600(%rdi), %xmm0
; SSE-NEXT:    movaps 1536(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1984(%rdi), %xmm0
; SSE-NEXT:    movaps 1920(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1856(%rdi), %xmm0
; SSE-NEXT:    movaps 1792(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 16(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 208(%rdi), %xmm0
; SSE-NEXT:    movaps 144(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 336(%rdi), %xmm0
; SSE-NEXT:    movaps 272(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 464(%rdi), %xmm0
; SSE-NEXT:    movaps 400(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 592(%rdi), %xmm0
; SSE-NEXT:    movaps 528(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 720(%rdi), %xmm0
; SSE-NEXT:    movaps 656(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 848(%rdi), %xmm0
; SSE-NEXT:    movaps 784(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 976(%rdi), %xmm0
; SSE-NEXT:    movaps 912(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1104(%rdi), %xmm0
; SSE-NEXT:    movaps 1040(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1232(%rdi), %xmm0
; SSE-NEXT:    movaps 1168(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1360(%rdi), %xmm0
; SSE-NEXT:    movaps 1296(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1488(%rdi), %xmm0
; SSE-NEXT:    movaps 1424(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1616(%rdi), %xmm0
; SSE-NEXT:    movaps 1552(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1744(%rdi), %xmm0
; SSE-NEXT:    movaps 1680(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1872(%rdi), %xmm0
; SSE-NEXT:    movaps 1808(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2000(%rdi), %xmm0
; SSE-NEXT:    movaps 1936(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 96(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 224(%rdi), %xmm0
; SSE-NEXT:    movaps 160(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 352(%rdi), %xmm0
; SSE-NEXT:    movaps 288(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 480(%rdi), %xmm0
; SSE-NEXT:    movaps 416(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 608(%rdi), %xmm0
; SSE-NEXT:    movaps 544(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 736(%rdi), %xmm0
; SSE-NEXT:    movaps 672(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 864(%rdi), %xmm0
; SSE-NEXT:    movaps 800(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 992(%rdi), %xmm0
; SSE-NEXT:    movaps 928(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1120(%rdi), %xmm0
; SSE-NEXT:    movaps 1056(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1248(%rdi), %xmm0
; SSE-NEXT:    movaps 1184(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1376(%rdi), %xmm0
; SSE-NEXT:    movaps 1312(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1504(%rdi), %xmm0
; SSE-NEXT:    movaps 1440(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1632(%rdi), %xmm0
; SSE-NEXT:    movaps 1568(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1760(%rdi), %xmm0
; SSE-NEXT:    movaps 1696(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1888(%rdi), %xmm0
; SSE-NEXT:    movaps 1824(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2016(%rdi), %xmm0
; SSE-NEXT:    movaps 1952(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 112(%rdi), %xmm0
; SSE-NEXT:    movaps 48(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 240(%rdi), %xmm0
; SSE-NEXT:    movaps 176(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 304(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 496(%rdi), %xmm0
; SSE-NEXT:    movaps 432(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 624(%rdi), %xmm0
; SSE-NEXT:    movaps 560(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 752(%rdi), %xmm0
; SSE-NEXT:    movaps 688(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill
; SSE-NEXT:    movaps 880(%rdi), %xmm0
; SSE-NEXT:    movaps 816(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1008(%rdi), %xmm0
; SSE-NEXT:    movaps 944(%rdi), %xmm12
; SSE-NEXT:    movaps %xmm12, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]
; SSE-NEXT:    movaps 1136(%rdi), %xmm0
; SSE-NEXT:    movaps 1072(%rdi), %xmm11
; SSE-NEXT:    movaps %xmm11, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1]
; SSE-NEXT:    movaps 1264(%rdi), %xmm0
; SSE-NEXT:    movaps 1200(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
; SSE-NEXT:    movaps 1392(%rdi), %xmm0
; SSE-NEXT:    movaps 1328(%rdi), %xmm13
; SSE-NEXT:    movaps %xmm13, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]
; SSE-NEXT:    movaps 1520(%rdi), %xmm0
; SSE-NEXT:    movaps 1456(%rdi), %xmm7
; SSE-NEXT:    movaps %xmm7, %xmm10
; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
; SSE-NEXT:    movaps 1648(%rdi), %xmm0
; SSE-NEXT:    movaps 1584(%rdi), %xmm5
; SSE-NEXT:    movaps %xmm5, %xmm9
; SSE-NEXT:    movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
; SSE-NEXT:    movaps 1776(%rdi), %xmm0
; SSE-NEXT:    movaps 1712(%rdi), %xmm4
; SSE-NEXT:    movaps %xmm4, %xmm8
; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]
; SSE-NEXT:    movaps 1904(%rdi), %xmm0
; SSE-NEXT:    movaps 1840(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps 2032(%rdi), %xmm0
; SSE-NEXT:    movaps 1968(%rdi), %xmm2
; SSE-NEXT:    movaps %xmm2, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm3, 240(%rax)
; SSE-NEXT:    movaps %xmm6, 224(%rax)
; SSE-NEXT:    movaps %xmm8, 208(%rax)
; SSE-NEXT:    movaps %xmm9, 192(%rax)
; SSE-NEXT:    movaps %xmm10, 176(%rax)
; SSE-NEXT:    movaps %xmm15, 160(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm2, 240(%rax)
; SSE-NEXT:    movaps %xmm1, 224(%rax)
; SSE-NEXT:    movaps %xmm4, 208(%rax)
; SSE-NEXT:    movaps %xmm5, 192(%rax)
; SSE-NEXT:    movaps %xmm7, 176(%rax)
; SSE-NEXT:    movaps %xmm13, 160(%rax)
; SSE-NEXT:    movaps %xmm14, 144(%rax)
; SSE-NEXT:    movaps %xmm11, 128(%rax)
; SSE-NEXT:    movaps %xmm12, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rax)
; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    addq $1688, %rsp # imm = 0x698
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i64_stride8_vf32:
; AVX:       # %bb.0:
; AVX-NEXT:    subq $2216, %rsp # imm = 0x8A8
; AVX-NEXT:    vmovaps 192(%rdi), %xmm2
; AVX-NEXT:    vmovaps 128(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm3[0],xmm2[0]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 704(%rdi), %xmm0
; AVX-NEXT:    vmovaps 640(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1216(%rdi), %xmm4
; AVX-NEXT:    vmovaps 1152(%rdi), %xmm6
; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm6[0],xmm4[0]
; AVX-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 64(%rdi), %xmm8
; AVX-NEXT:    vmovaps (%rdi), %xmm9
; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm9[0],xmm8[0]
; AVX-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 576(%rdi), %xmm5
; AVX-NEXT:    vmovaps 512(%rdi), %xmm7
; AVX-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm7[0],xmm5[0]
; AVX-NEXT:    vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1088(%rdi), %xmm10
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1024(%rdi), %xmm2
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm9[1],xmm8[1]
; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm10[0]
; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm6[1],xmm4[1]
; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1600(%rdi), %xmm3
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm10[1]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1728(%rdi), %xmm2
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1664(%rdi), %xmm0
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm5[1]
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm2[0]
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1536(%rdi), %xmm0
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm3[0]
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 448(%rdi), %xmm0
; AVX-NEXT:    vmovaps 384(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1472(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1408(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1344(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1280(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 960(%rdi), %xmm0
; AVX-NEXT:    vmovaps 896(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 832(%rdi), %xmm0
; AVX-NEXT:    vmovaps 768(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1984(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1920(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1856(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1792(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 128(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 80(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 16(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 704(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 640(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 592(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 528(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1216(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1152(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1104(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1040(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1728(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1664(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1616(%rdi), %xmm14
; AVX-NEXT:    vmovaps 1552(%rdi), %xmm12
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm12[0],xmm14[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1872(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1808(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1472(%rdi), %ymm11
; AVX-NEXT:    vmovaps 1408(%rdi), %ymm10
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX-NEXT:    vmovaps 1360(%rdi), %xmm9
; AVX-NEXT:    vmovaps 1296(%rdi), %xmm8
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm8[0],xmm9[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 960(%rdi), %ymm7
; AVX-NEXT:    vmovaps 896(%rdi), %ymm6
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
; AVX-NEXT:    vmovaps 848(%rdi), %xmm5
; AVX-NEXT:    vmovaps 784(%rdi), %xmm4
; AVX-NEXT:    vmovlhps {{.*#+}} xmm13 = xmm4[0],xmm5[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 448(%rdi), %ymm3
; AVX-NEXT:    vmovaps 384(%rdi), %ymm13
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm3[0],ymm13[2],ymm3[2]
; AVX-NEXT:    vmovaps 336(%rdi), %xmm2
; AVX-NEXT:    vmovaps 272(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm12 = xmm12[1],xmm14[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm13[1],ymm3[1],ymm13[3],ymm3[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm5[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm8[1],xmm9[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 480(%rdi), %xmm0
; AVX-NEXT:    vmovaps 416(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 992(%rdi), %xmm0
; AVX-NEXT:    vmovaps 928(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1248(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1184(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1504(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1440(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 736(%rdi), %xmm0
; AVX-NEXT:    vmovaps 672(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 224(%rdi), %xmm0
; AVX-NEXT:    vmovaps 160(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1760(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1696(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2016(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1952(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 112(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 48(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 368(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 304(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 624(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 560(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, (%rsp) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 880(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 816(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1248(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1184(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1136(%rdi), %xmm14
; AVX-NEXT:    vmovaps 1072(%rdi), %xmm13
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1504(%rdi), %ymm12
; AVX-NEXT:    vmovaps 1440(%rdi), %ymm11
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX-NEXT:    vmovaps 1392(%rdi), %xmm10
; AVX-NEXT:    vmovaps 1328(%rdi), %xmm9
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1760(%rdi), %ymm8
; AVX-NEXT:    vmovaps 1696(%rdi), %ymm7
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX-NEXT:    vmovaps 1648(%rdi), %xmm6
; AVX-NEXT:    vmovaps 1584(%rdi), %xmm5
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2016(%rdi), %ymm4
; AVX-NEXT:    vmovaps 1952(%rdi), %ymm3
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX-NEXT:    vmovaps 1904(%rdi), %xmm2
; AVX-NEXT:    vmovaps 1840(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rdx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r8)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%r9)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX-NEXT:    vmovaps %ymm5, 192(%rax)
; AVX-NEXT:    vmovaps %ymm9, 160(%rax)
; AVX-NEXT:    vmovaps %ymm13, 128(%rax)
; AVX-NEXT:    vmovaps %ymm15, 96(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, (%rax)
; AVX-NEXT:    addq $2216, %rsp # imm = 0x8A8
; AVX-NEXT:    vzeroupper
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i64_stride8_vf32:
; AVX2:       # %bb.0:
; AVX2-NEXT:    subq $2248, %rsp # imm = 0x8C8
; AVX2-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 448(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 832(%rdi), %xmm2
; AVX2-NEXT:    vinsertf128 $1, 960(%rdi), %ymm2, %ymm2
; AVX2-NEXT:    vmovaps 768(%rdi), %xmm3
; AVX2-NEXT:    vinsertf128 $1, 896(%rdi), %ymm3, %ymm3
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1344(%rdi), %xmm4
; AVX2-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm4, %ymm4
; AVX2-NEXT:    vmovaps 1280(%rdi), %xmm5
; AVX2-NEXT:    vinsertf128 $1, 1408(%rdi), %ymm5, %ymm5
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1856(%rdi), %xmm6
; AVX2-NEXT:    vinsertf128 $1, 1984(%rdi), %ymm6, %ymm6
; AVX2-NEXT:    vmovaps 1792(%rdi), %xmm7
; AVX2-NEXT:    vinsertf128 $1, 1920(%rdi), %ymm7, %ymm7
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps (%rdi), %xmm9
; AVX2-NEXT:    vmovaps 64(%rdi), %xmm8
; AVX2-NEXT:    vinsertf128 $1, 192(%rdi), %ymm8, %ymm8
; AVX2-NEXT:    vinsertf128 $1, 128(%rdi), %ymm9, %ymm9
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 576(%rdi), %xmm10
; AVX2-NEXT:    vinsertf128 $1, 704(%rdi), %ymm10, %ymm10
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 512(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 640(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1088(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1216(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1024(%rdi), %xmm2
; AVX2-NEXT:    vinsertf128 $1, 1152(%rdi), %ymm2, %ymm2
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1600(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1728(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1536(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 320(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 256(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 448(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 384(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 832(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 768(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 960(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1344(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1280(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1408(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1856(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1792(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps (%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill
; AVX2-NEXT:    vmovaps 192(%rdi), %ymm14
; AVX2-NEXT:    vmovaps 128(%rdi), %ymm12
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm12[0],ymm14[0],ymm12[2],ymm14[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 576(%rdi), %ymm11
; AVX2-NEXT:    vmovaps 512(%rdi), %ymm10
; AVX2-NEXT:    vmovaps 704(%rdi), %ymm8
; AVX2-NEXT:    vmovaps 640(%rdi), %ymm7
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1088(%rdi), %ymm6
; AVX2-NEXT:    vmovaps 1024(%rdi), %ymm9
; AVX2-NEXT:    vmovaps 1216(%rdi), %ymm5
; AVX2-NEXT:    vmovaps 1152(%rdi), %ymm3
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm6[0],ymm9[2],ymm6[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1600(%rdi), %ymm4
; AVX2-NEXT:    vmovaps 1536(%rdi), %ymm13
; AVX2-NEXT:    vmovaps 1728(%rdi), %ymm2
; AVX2-NEXT:    vmovaps 1664(%rdi), %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm13[0],ymm4[0],ymm13[2],ymm4[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm12[1],ymm14[1],ymm12[3],ymm14[3]
; AVX2-NEXT:    vmovups (%rsp), %ymm12 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm12 = ymm12[1],mem[1],ymm12[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm12[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm6[1],ymm9[3],ymm6[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm4[1],ymm13[3],ymm4[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1184(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1760(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1952(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 32(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 352(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 288(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 608(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 544(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 864(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 800(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1120(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1056(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1248(%rdi), %ymm12
; AVX2-NEXT:    vmovaps 1184(%rdi), %ymm11
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1376(%rdi), %ymm10
; AVX2-NEXT:    vmovaps 1312(%rdi), %ymm9
; AVX2-NEXT:    vmovaps 1504(%rdi), %ymm8
; AVX2-NEXT:    vmovaps 1440(%rdi), %ymm7
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1632(%rdi), %ymm6
; AVX2-NEXT:    vmovaps 1568(%rdi), %ymm15
; AVX2-NEXT:    vmovaps 1760(%rdi), %ymm5
; AVX2-NEXT:    vmovaps 1696(%rdi), %ymm3
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm15[0],ymm6[0],ymm15[2],ymm6[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1888(%rdi), %ymm4
; AVX2-NEXT:    vmovaps 1824(%rdi), %ymm14
; AVX2-NEXT:    vmovaps 2016(%rdi), %ymm2
; AVX2-NEXT:    vmovaps 1952(%rdi), %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm11 = ymm11[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm15[1],ymm6[1],ymm15[3],ymm6[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm3[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-NEXT:    vmovaps %ymm3, 192(%rax)
; AVX2-NEXT:    vmovaps %ymm7, 160(%rax)
; AVX2-NEXT:    vmovaps %ymm11, 128(%rax)
; AVX2-NEXT:    vmovaps %ymm13, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-NEXT:    addq $2248, %rsp # imm = 0x8C8
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i64_stride8_vf32:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    subq $2248, %rsp # imm = 0x8C8
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 832(%rdi), %xmm2
; AVX2-FP-NEXT:    vinsertf128 $1, 960(%rdi), %ymm2, %ymm2
; AVX2-FP-NEXT:    vmovaps 768(%rdi), %xmm3
; AVX2-FP-NEXT:    vinsertf128 $1, 896(%rdi), %ymm3, %ymm3
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1344(%rdi), %xmm4
; AVX2-FP-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm4, %ymm4
; AVX2-FP-NEXT:    vmovaps 1280(%rdi), %xmm5
; AVX2-FP-NEXT:    vinsertf128 $1, 1408(%rdi), %ymm5, %ymm5
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1856(%rdi), %xmm6
; AVX2-FP-NEXT:    vinsertf128 $1, 1984(%rdi), %ymm6, %ymm6
; AVX2-FP-NEXT:    vmovaps 1792(%rdi), %xmm7
; AVX2-FP-NEXT:    vinsertf128 $1, 1920(%rdi), %ymm7, %ymm7
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm9
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm8
; AVX2-FP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm8, %ymm8
; AVX2-FP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm9, %ymm9
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 576(%rdi), %xmm10
; AVX2-FP-NEXT:    vinsertf128 $1, 704(%rdi), %ymm10, %ymm10
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 512(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 640(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1088(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1216(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1024(%rdi), %xmm2
; AVX2-FP-NEXT:    vinsertf128 $1, 1152(%rdi), %ymm2, %ymm2
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1600(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1728(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1536(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 448(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 384(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 832(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 768(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 960(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1344(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1280(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1408(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1856(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1792(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 192(%rdi), %ymm14
; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm12
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm12[0],ymm14[0],ymm12[2],ymm14[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 576(%rdi), %ymm11
; AVX2-FP-NEXT:    vmovaps 512(%rdi), %ymm10
; AVX2-FP-NEXT:    vmovaps 704(%rdi), %ymm8
; AVX2-FP-NEXT:    vmovaps 640(%rdi), %ymm7
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1088(%rdi), %ymm6
; AVX2-FP-NEXT:    vmovaps 1024(%rdi), %ymm9
; AVX2-FP-NEXT:    vmovaps 1216(%rdi), %ymm5
; AVX2-FP-NEXT:    vmovaps 1152(%rdi), %ymm3
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm6[0],ymm9[2],ymm6[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1600(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovaps 1536(%rdi), %ymm13
; AVX2-FP-NEXT:    vmovaps 1728(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps 1664(%rdi), %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm13[0],ymm4[0],ymm13[2],ymm4[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm12[1],ymm14[1],ymm12[3],ymm14[3]
; AVX2-FP-NEXT:    vmovups (%rsp), %ymm12 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm12 = ymm12[1],mem[1],ymm12[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm12[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm6[1],ymm9[3],ymm6[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm4[1],ymm13[3],ymm4[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1184(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1760(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1952(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 608(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 544(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 864(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 800(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1120(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1056(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1248(%rdi), %ymm12
; AVX2-FP-NEXT:    vmovaps 1184(%rdi), %ymm11
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1376(%rdi), %ymm10
; AVX2-FP-NEXT:    vmovaps 1312(%rdi), %ymm9
; AVX2-FP-NEXT:    vmovaps 1504(%rdi), %ymm8
; AVX2-FP-NEXT:    vmovaps 1440(%rdi), %ymm7
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1632(%rdi), %ymm6
; AVX2-FP-NEXT:    vmovaps 1568(%rdi), %ymm15
; AVX2-FP-NEXT:    vmovaps 1760(%rdi), %ymm5
; AVX2-FP-NEXT:    vmovaps 1696(%rdi), %ymm3
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm15[0],ymm6[0],ymm15[2],ymm6[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1888(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovaps 1824(%rdi), %ymm14
; AVX2-FP-NEXT:    vmovaps 2016(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps 1952(%rdi), %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm11 = ymm11[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm15[1],ymm6[1],ymm15[3],ymm6[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm3[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-FP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm3, 192(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm7, 160(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm11, 128(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm13, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FP-NEXT:    addq $2248, %rsp # imm = 0x8C8
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i64_stride8_vf32:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    subq $2248, %rsp # imm = 0x8C8
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %xmm2
; AVX2-FCP-NEXT:    vinsertf128 $1, 960(%rdi), %ymm2, %ymm2
; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %xmm3
; AVX2-FCP-NEXT:    vinsertf128 $1, 896(%rdi), %ymm3, %ymm3
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1344(%rdi), %xmm4
; AVX2-FCP-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm4, %ymm4
; AVX2-FCP-NEXT:    vmovaps 1280(%rdi), %xmm5
; AVX2-FCP-NEXT:    vinsertf128 $1, 1408(%rdi), %ymm5, %ymm5
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1856(%rdi), %xmm6
; AVX2-FCP-NEXT:    vinsertf128 $1, 1984(%rdi), %ymm6, %ymm6
; AVX2-FCP-NEXT:    vmovaps 1792(%rdi), %xmm7
; AVX2-FCP-NEXT:    vinsertf128 $1, 1920(%rdi), %ymm7, %ymm7
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FCP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm9
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm8
; AVX2-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm8, %ymm8
; AVX2-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm9, %ymm9
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FCP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %xmm10
; AVX2-FCP-NEXT:    vinsertf128 $1, 704(%rdi), %ymm10, %ymm10
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 640(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1088(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1216(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1024(%rdi), %xmm2
; AVX2-FCP-NEXT:    vinsertf128 $1, 1152(%rdi), %ymm2, %ymm2
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1600(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1728(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1536(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 448(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 384(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 960(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1344(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1280(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1408(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1856(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1792(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %ymm14
; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm12
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm12[0],ymm14[0],ymm12[2],ymm14[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %ymm11
; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %ymm10
; AVX2-FCP-NEXT:    vmovaps 704(%rdi), %ymm8
; AVX2-FCP-NEXT:    vmovaps 640(%rdi), %ymm7
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm9[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1088(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovaps 1024(%rdi), %ymm9
; AVX2-FCP-NEXT:    vmovaps 1216(%rdi), %ymm5
; AVX2-FCP-NEXT:    vmovaps 1152(%rdi), %ymm3
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm9[0],ymm6[0],ymm9[2],ymm6[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1600(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovaps 1536(%rdi), %ymm13
; AVX2-FCP-NEXT:    vmovaps 1728(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps 1664(%rdi), %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm13[0],ymm4[0],ymm13[2],ymm4[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm12[1],ymm14[1],ymm12[3],ymm14[3]
; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm12 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm12 = ymm12[1],mem[1],ymm12[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm12[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm6[1],ymm9[3],ymm6[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm13[1],ymm4[1],ymm13[3],ymm4[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1184(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1760(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1952(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1120(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1056(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1248(%rdi), %ymm12
; AVX2-FCP-NEXT:    vmovaps 1184(%rdi), %ymm11
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1376(%rdi), %ymm10
; AVX2-FCP-NEXT:    vmovaps 1312(%rdi), %ymm9
; AVX2-FCP-NEXT:    vmovaps 1504(%rdi), %ymm8
; AVX2-FCP-NEXT:    vmovaps 1440(%rdi), %ymm7
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1632(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovaps 1568(%rdi), %ymm15
; AVX2-FCP-NEXT:    vmovaps 1760(%rdi), %ymm5
; AVX2-FCP-NEXT:    vmovaps 1696(%rdi), %ymm3
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm15[0],ymm6[0],ymm15[2],ymm6[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1888(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovaps 1824(%rdi), %ymm14
; AVX2-FCP-NEXT:    vmovaps 2016(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps 1952(%rdi), %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm11 = ymm11[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm5[1],ymm3[3],ymm5[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm15[1],ymm6[1],ymm15[3],ymm6[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm3[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm3, 192(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm7, 160(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm11, 128(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm13, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FCP-NEXT:    addq $2248, %rsp # imm = 0x8C8
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i64_stride8_vf32:
; AVX512:       # %bb.0:
; AVX512-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512-NEXT:    movb $-64, %al
; AVX512-NEXT:    kmovw %eax, %k1
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i64_stride8_vf32:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512-FCP-NEXT:    movb $-64, %al
; AVX512-FCP-NEXT:    kmovw %eax, %k1
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512-FCP-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512-FCP-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512-FCP-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512-FCP-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512-FCP-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512-FCP-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512-FCP-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512-FCP-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512-FCP-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512-FCP-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512-FCP-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512-FCP-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512-FCP-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512-FCP-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i64_stride8_vf32:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512DQ-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512DQ-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512DQ-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512DQ-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512DQ-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512DQ-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512DQ-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512DQ-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512DQ-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512DQ-NEXT:    movb $-64, %al
; AVX512DQ-NEXT:    kmovw %eax, %k1
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512DQ-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512DQ-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512DQ-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512DQ-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512DQ-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512DQ-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512DQ-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512DQ-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512DQ-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512DQ-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512DQ-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512DQ-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512DQ-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512DQ-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512DQ-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512DQ-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512DQ-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512DQ-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512DQ-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512DQ-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512DQ-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512DQ-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512DQ-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512DQ-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512DQ-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512DQ-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512DQ-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512DQ-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512DQ-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512DQ-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512DQ-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512DQ-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512DQ-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512DQ-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512DQ-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512DQ-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512DQ-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512DQ-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512DQ-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512DQ-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512DQ-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512DQ-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i64_stride8_vf32:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512DQ-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512DQ-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512DQ-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512DQ-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512DQ-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512DQ-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512DQ-FCP-NEXT:    movb $-64, %al
; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512DQ-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512DQ-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512DQ-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512DQ-FCP-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512DQ-FCP-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512DQ-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512DQ-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512DQ-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512DQ-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512DQ-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512DQ-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512DQ-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512DQ-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512DQ-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512DQ-FCP-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512DQ-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512DQ-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512DQ-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512DQ-FCP-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512DQ-FCP-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride8_vf32:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512BW-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512BW-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512BW-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512BW-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512BW-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512BW-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512BW-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512BW-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512BW-NEXT:    movb $-64, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512BW-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512BW-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512BW-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512BW-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512BW-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512BW-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512BW-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512BW-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512BW-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512BW-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512BW-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512BW-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512BW-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512BW-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512BW-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512BW-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512BW-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512BW-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512BW-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512BW-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512BW-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512BW-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512BW-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512BW-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512BW-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512BW-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512BW-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512BW-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512BW-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512BW-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512BW-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512BW-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512BW-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512BW-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512BW-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512BW-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i64_stride8_vf32:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512BW-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512BW-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512BW-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512BW-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512BW-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512BW-FCP-NEXT:    movb $-64, %al
; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512BW-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512BW-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512BW-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512BW-FCP-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512BW-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512BW-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512BW-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512BW-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512BW-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512BW-FCP-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512BW-FCP-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512BW-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512BW-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512BW-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512BW-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512BW-FCP-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512BW-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512BW-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512BW-FCP-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512BW-FCP-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i64_stride8_vf32:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512DQ-BW-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512DQ-BW-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512DQ-BW-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512DQ-BW-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512DQ-BW-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512DQ-BW-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512DQ-BW-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512DQ-BW-NEXT:    movb $-64, %al
; AVX512DQ-BW-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512DQ-BW-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512DQ-BW-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512DQ-BW-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512DQ-BW-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512DQ-BW-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512DQ-BW-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512DQ-BW-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512DQ-BW-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512DQ-BW-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512DQ-BW-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512DQ-BW-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512DQ-BW-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512DQ-BW-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512DQ-BW-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512DQ-BW-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512DQ-BW-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512DQ-BW-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512DQ-BW-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512DQ-BW-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512DQ-BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512DQ-BW-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512DQ-BW-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i64_stride8_vf32:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    subq $3208, %rsp # imm = 0xC88
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm19
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm18
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm27
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm31
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm20
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm23
; AVX512DQ-BW-FCP-NEXT:    movb $-64, %al
; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %ymm22
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm10
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm10[0],ymm22[0],ymm10[2],ymm22[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm13 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm4 = ymm13[2,3],ymm4[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 704(%rdi), %ymm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 640(%rdi), %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm14 = ymm13[0],ymm7[0],ymm13[2],ymm7[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm21
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm21[0],ymm17[0],ymm21[2],ymm17[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm14[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm4, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm14
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm1 = ymm14[0],ymm4[0],ymm14[2],ymm4[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %ymm28
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm25[0],ymm28[2],ymm25[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm8[2,3],ymm1[2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm29
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm11, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm26
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm11 = ymm8[0],ymm2[0],ymm8[2],ymm2[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %ymm30
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm1[0],ymm30[0],ymm1[2],ymm30[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm11 = ymm15[2,3],ymm11[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm0, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm24
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm18
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm18 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm10[1],ymm22[1],ymm10[3],ymm22[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm6[1],ymm9[1],ymm6[3],ymm9[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm18, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm13[1],ymm7[1],ymm13[3],ymm7[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm21[1],ymm17[1],ymm21[3],ymm17[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm5[2,3],ymm3[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm23, (%rsp) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm5 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm14[1],ymm4[1],ymm14[3],ymm4[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm7
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm28[1],ymm25[1],ymm28[3],ymm25[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm13
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm5, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm12, %zmm26, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm2 = ymm8[1],ymm2[1],ymm8[3],ymm2[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm1[1],ymm30[1],ymm1[3],ymm30[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm2[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm17[0],zmm19[0],zmm17[2],zmm19[2],zmm17[4],zmm19[4],zmm17[6],zmm19[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm26
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm29
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm23
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm22
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm18
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm16, %zmm15, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm9[0],zmm12[0],zmm9[2],zmm12[2],zmm9[4],zmm12[4],zmm9[6],zmm12[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm7, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm1[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm21 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm21, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm16 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm16, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm7 = zmm17[1],zmm19[1],zmm17[3],zmm19[3],zmm17[5],zmm19[5],zmm17[7],zmm19[7]
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm17 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    # zmm17 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm27[0],zmm24[0],zmm27[2],zmm24[2],zmm27[4],zmm24[4],zmm27[6],zmm24[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm27[1],zmm24[1],zmm27[3],zmm24[3],zmm27[5],zmm24[5],zmm27[7],zmm24[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm21, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm16, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm7, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm2, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm26, %zmm2, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm21, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm16, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm30
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm23, %zmm30
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm11 = zmm11[1],zmm10[1],zmm11[3],zmm10[3],zmm11[5],zmm10[5],zmm11[7],zmm10[7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm17, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm30 {%k1} = zmm29[0],zmm31[0],zmm29[2],zmm31[2],zmm29[4],zmm31[4],zmm29[6],zmm31[6]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm29[1],zmm31[1],zmm29[3],zmm31[3],zmm29[5],zmm31[5],zmm29[7],zmm31[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm21, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm16, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm2, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm11 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm11, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm28, %zmm2, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm11 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm23, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm20[1],zmm22[3],zmm20[3],zmm22[5],zmm20[5],zmm22[7],zmm20[7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm17, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm31 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm27 {%k1} = zmm31[0],zmm1[0],zmm31[2],zmm1[2],zmm31[4],zmm1[4],zmm31[6],zmm1[6]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm0 {%k1} = zmm31[1],zmm1[1],zmm31[3],zmm1[3],zmm31[5],zmm1[5],zmm31[7],zmm1[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm24
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm21, %zmm31
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm16, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm24
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm14 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm20, %zmm5, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm14 = zmm22[1],zmm12[1],zmm22[3],zmm12[3],zmm22[5],zmm12[5],zmm22[7],zmm12[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm14 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm14, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm21, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm21
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm21
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm16, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm12, %zmm22, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm23, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm26, %zmm23, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm23, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm29
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm23, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm28, %zmm23, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm23, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm23, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm23, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm16, %zmm22, %zmm23
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm17, %zmm22
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm21[0],zmm20[0],zmm21[2],zmm20[2],zmm21[4],zmm20[4],zmm21[6],zmm20[6]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm22 {%k1} = zmm21[1],zmm20[1],zmm21[3],zmm20[3],zmm21[5],zmm20[5],zmm21[7],zmm20[7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm29, %zmm17, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm17, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm28, %zmm17, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm17, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm17, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm17, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm17, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm26, %zmm17, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm11[0,1,2,3],ymm14[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm0, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm8[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm30, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm27, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm25 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm1, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm2, %zmm25, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 576(%rdi), %xmm5
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm5, %ymm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 512(%rdi), %xmm6
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 640(%rdi), %ymm6, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm18 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm18, %zmm7, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm13
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm13, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %xmm18
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 1152(%rdi), %ymm18, %ymm18
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm19 = ymm18[0],ymm13[0],ymm18[2],ymm13[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm19, %zmm7, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %xmm21
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 1728(%rdi), %ymm21, %ymm21
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1536(%rdi), %xmm25
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 1664(%rdi), %ymm25, %ymm25
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[2],ymm21[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm26, %zmm7, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm18[1],ymm13[1],ymm18[3],ymm13[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm7, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm7, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm7, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm1 = ymm25[1],ymm21[1],ymm25[3],ymm21[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm6, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm6 = ymm6[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm23, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vblendps $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinsertf64x4 $0, %ymm7, %zmm8, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm8 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm8 = mem[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm9, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm9 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm9, %zmm22, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm11 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm11 = ymm3[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm11, %zmm3, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, 192(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, 128(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, 64(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 192(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, 64(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, 128(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, 192(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, 128(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, 192(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm7, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    addq $3208, %rsp # imm = 0xC88
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <256 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120, i32 128, i32 136, i32 144, i32 152, i32 160, i32 168, i32 176, i32 184, i32 192, i32 200, i32 208, i32 216, i32 224, i32 232, i32 240, i32 248>
  %strided.vec1 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121, i32 129, i32 137, i32 145, i32 153, i32 161, i32 169, i32 177, i32 185, i32 193, i32 201, i32 209, i32 217, i32 225, i32 233, i32 241, i32 249>
  %strided.vec2 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122, i32 130, i32 138, i32 146, i32 154, i32 162, i32 170, i32 178, i32 186, i32 194, i32 202, i32 210, i32 218, i32 226, i32 234, i32 242, i32 250>
  %strided.vec3 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123, i32 131, i32 139, i32 147, i32 155, i32 163, i32 171, i32 179, i32 187, i32 195, i32 203, i32 211, i32 219, i32 227, i32 235, i32 243, i32 251>
  %strided.vec4 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124, i32 132, i32 140, i32 148, i32 156, i32 164, i32 172, i32 180, i32 188, i32 196, i32 204, i32 212, i32 220, i32 228, i32 236, i32 244, i32 252>
  %strided.vec5 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125, i32 133, i32 141, i32 149, i32 157, i32 165, i32 173, i32 181, i32 189, i32 197, i32 205, i32 213, i32 221, i32 229, i32 237, i32 245, i32 253>
  %strided.vec6 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126, i32 134, i32 142, i32 150, i32 158, i32 166, i32 174, i32 182, i32 190, i32 198, i32 206, i32 214, i32 222, i32 230, i32 238, i32 246, i32 254>
  %strided.vec7 = shufflevector <256 x i64> %wide.vec, <256 x i64> poison, <32 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127, i32 135, i32 143, i32 151, i32 159, i32 167, i32 175, i32 183, i32 191, i32 199, i32 207, i32 215, i32 223, i32 231, i32 239, i32 247, i32 255>
  store <32 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <32 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <32 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <32 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <32 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <32 x i64> %strided.vec5, ptr %out.vec5, align 64
  store <32 x i64> %strided.vec6, ptr %out.vec6, align 64
  store <32 x i64> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i64_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i64_stride8_vf64:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $3720, %rsp # imm = 0xE88
; SSE-NEXT:    movaps 960(%rdi), %xmm0
; SSE-NEXT:    movaps 832(%rdi), %xmm1
; SSE-NEXT:    movaps 768(%rdi), %xmm8
; SSE-NEXT:    movaps 704(%rdi), %xmm2
; SSE-NEXT:    movaps 640(%rdi), %xmm9
; SSE-NEXT:    movaps 576(%rdi), %xmm3
; SSE-NEXT:    movaps 512(%rdi), %xmm10
; SSE-NEXT:    movaps 448(%rdi), %xmm4
; SSE-NEXT:    movaps 384(%rdi), %xmm11
; SSE-NEXT:    movaps 320(%rdi), %xmm5
; SSE-NEXT:    movaps 256(%rdi), %xmm12
; SSE-NEXT:    movaps 192(%rdi), %xmm6
; SSE-NEXT:    movaps 128(%rdi), %xmm13
; SSE-NEXT:    movaps 64(%rdi), %xmm7
; SSE-NEXT:    movaps (%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm15
; SSE-NEXT:    movlhps {{.*#+}} xmm15 = xmm15[0],xmm7[0]
; SSE-NEXT:    movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm7[1]
; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm13, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]
; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]
; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm12, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]
; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm5[1]
; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm11, %xmm5
; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm4[0]
; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm4[1]
; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm10, %xmm4
; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm3[1]
; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm9, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm2[1]
; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps %xmm8, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]
; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 896(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1088(%rdi), %xmm0
; SSE-NEXT:    movaps 1024(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1216(%rdi), %xmm0
; SSE-NEXT:    movaps 1152(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1344(%rdi), %xmm0
; SSE-NEXT:    movaps 1280(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1472(%rdi), %xmm0
; SSE-NEXT:    movaps 1408(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1600(%rdi), %xmm0
; SSE-NEXT:    movaps 1536(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1728(%rdi), %xmm0
; SSE-NEXT:    movaps 1664(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1856(%rdi), %xmm0
; SSE-NEXT:    movaps 1792(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1984(%rdi), %xmm0
; SSE-NEXT:    movaps 1920(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2112(%rdi), %xmm0
; SSE-NEXT:    movaps 2048(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2240(%rdi), %xmm0
; SSE-NEXT:    movaps 2176(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2368(%rdi), %xmm0
; SSE-NEXT:    movaps 2304(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2496(%rdi), %xmm0
; SSE-NEXT:    movaps 2432(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2624(%rdi), %xmm0
; SSE-NEXT:    movaps 2560(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2752(%rdi), %xmm0
; SSE-NEXT:    movaps 2688(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2880(%rdi), %xmm0
; SSE-NEXT:    movaps 2816(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3008(%rdi), %xmm0
; SSE-NEXT:    movaps 2944(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3136(%rdi), %xmm0
; SSE-NEXT:    movaps 3072(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3264(%rdi), %xmm0
; SSE-NEXT:    movaps 3200(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3392(%rdi), %xmm0
; SSE-NEXT:    movaps 3328(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3520(%rdi), %xmm0
; SSE-NEXT:    movaps 3456(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3648(%rdi), %xmm0
; SSE-NEXT:    movaps 3584(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3776(%rdi), %xmm0
; SSE-NEXT:    movaps 3712(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3904(%rdi), %xmm0
; SSE-NEXT:    movaps 3840(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 4032(%rdi), %xmm0
; SSE-NEXT:    movaps 3968(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 80(%rdi), %xmm0
; SSE-NEXT:    movaps 16(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 208(%rdi), %xmm0
; SSE-NEXT:    movaps 144(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 336(%rdi), %xmm0
; SSE-NEXT:    movaps 272(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 464(%rdi), %xmm0
; SSE-NEXT:    movaps 400(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 592(%rdi), %xmm0
; SSE-NEXT:    movaps 528(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 720(%rdi), %xmm0
; SSE-NEXT:    movaps 656(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 848(%rdi), %xmm0
; SSE-NEXT:    movaps 784(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 976(%rdi), %xmm0
; SSE-NEXT:    movaps 912(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1104(%rdi), %xmm0
; SSE-NEXT:    movaps 1040(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1232(%rdi), %xmm0
; SSE-NEXT:    movaps 1168(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1360(%rdi), %xmm0
; SSE-NEXT:    movaps 1296(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1488(%rdi), %xmm0
; SSE-NEXT:    movaps 1424(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1616(%rdi), %xmm0
; SSE-NEXT:    movaps 1552(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1744(%rdi), %xmm0
; SSE-NEXT:    movaps 1680(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1872(%rdi), %xmm0
; SSE-NEXT:    movaps 1808(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2000(%rdi), %xmm0
; SSE-NEXT:    movaps 1936(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2128(%rdi), %xmm0
; SSE-NEXT:    movaps 2064(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2256(%rdi), %xmm0
; SSE-NEXT:    movaps 2192(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2384(%rdi), %xmm0
; SSE-NEXT:    movaps 2320(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2512(%rdi), %xmm0
; SSE-NEXT:    movaps 2448(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2640(%rdi), %xmm0
; SSE-NEXT:    movaps 2576(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2768(%rdi), %xmm0
; SSE-NEXT:    movaps 2704(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2896(%rdi), %xmm0
; SSE-NEXT:    movaps 2832(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3024(%rdi), %xmm0
; SSE-NEXT:    movaps 2960(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3152(%rdi), %xmm0
; SSE-NEXT:    movaps 3088(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3280(%rdi), %xmm0
; SSE-NEXT:    movaps 3216(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3408(%rdi), %xmm0
; SSE-NEXT:    movaps 3344(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3536(%rdi), %xmm0
; SSE-NEXT:    movaps 3472(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3664(%rdi), %xmm0
; SSE-NEXT:    movaps 3600(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3792(%rdi), %xmm0
; SSE-NEXT:    movaps 3728(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3920(%rdi), %xmm0
; SSE-NEXT:    movaps 3856(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 4048(%rdi), %xmm0
; SSE-NEXT:    movaps 3984(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 96(%rdi), %xmm0
; SSE-NEXT:    movaps 32(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 224(%rdi), %xmm0
; SSE-NEXT:    movaps 160(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 352(%rdi), %xmm0
; SSE-NEXT:    movaps 288(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 480(%rdi), %xmm0
; SSE-NEXT:    movaps 416(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 608(%rdi), %xmm0
; SSE-NEXT:    movaps 544(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 736(%rdi), %xmm0
; SSE-NEXT:    movaps 672(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 864(%rdi), %xmm0
; SSE-NEXT:    movaps 800(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 992(%rdi), %xmm0
; SSE-NEXT:    movaps 928(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1120(%rdi), %xmm0
; SSE-NEXT:    movaps 1056(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1248(%rdi), %xmm0
; SSE-NEXT:    movaps 1184(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1376(%rdi), %xmm0
; SSE-NEXT:    movaps 1312(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1504(%rdi), %xmm0
; SSE-NEXT:    movaps 1440(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1632(%rdi), %xmm0
; SSE-NEXT:    movaps 1568(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1760(%rdi), %xmm0
; SSE-NEXT:    movaps 1696(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1888(%rdi), %xmm0
; SSE-NEXT:    movaps 1824(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2016(%rdi), %xmm0
; SSE-NEXT:    movaps 1952(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2144(%rdi), %xmm0
; SSE-NEXT:    movaps 2080(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2272(%rdi), %xmm0
; SSE-NEXT:    movaps 2208(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2400(%rdi), %xmm0
; SSE-NEXT:    movaps 2336(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2528(%rdi), %xmm0
; SSE-NEXT:    movaps 2464(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2656(%rdi), %xmm0
; SSE-NEXT:    movaps 2592(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2784(%rdi), %xmm0
; SSE-NEXT:    movaps 2720(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2912(%rdi), %xmm0
; SSE-NEXT:    movaps 2848(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3040(%rdi), %xmm0
; SSE-NEXT:    movaps 2976(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3168(%rdi), %xmm0
; SSE-NEXT:    movaps 3104(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3296(%rdi), %xmm0
; SSE-NEXT:    movaps 3232(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3424(%rdi), %xmm0
; SSE-NEXT:    movaps 3360(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3552(%rdi), %xmm0
; SSE-NEXT:    movaps 3488(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3680(%rdi), %xmm0
; SSE-NEXT:    movaps 3616(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3808(%rdi), %xmm0
; SSE-NEXT:    movaps 3744(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 3936(%rdi), %xmm0
; SSE-NEXT:    movaps 3872(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 4064(%rdi), %xmm0
; SSE-NEXT:    movaps 4000(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 112(%rdi), %xmm0
; SSE-NEXT:    movaps 48(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 240(%rdi), %xmm0
; SSE-NEXT:    movaps 176(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 368(%rdi), %xmm0
; SSE-NEXT:    movaps 304(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 496(%rdi), %xmm0
; SSE-NEXT:    movaps 432(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 624(%rdi), %xmm0
; SSE-NEXT:    movaps 560(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 752(%rdi), %xmm0
; SSE-NEXT:    movaps 688(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 880(%rdi), %xmm0
; SSE-NEXT:    movaps 816(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1008(%rdi), %xmm0
; SSE-NEXT:    movaps 944(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1136(%rdi), %xmm0
; SSE-NEXT:    movaps 1072(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1264(%rdi), %xmm0
; SSE-NEXT:    movaps 1200(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1392(%rdi), %xmm0
; SSE-NEXT:    movaps 1328(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1520(%rdi), %xmm0
; SSE-NEXT:    movaps 1456(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1648(%rdi), %xmm0
; SSE-NEXT:    movaps 1584(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1776(%rdi), %xmm0
; SSE-NEXT:    movaps 1712(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 1904(%rdi), %xmm0
; SSE-NEXT:    movaps 1840(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2032(%rdi), %xmm0
; SSE-NEXT:    movaps 1968(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2160(%rdi), %xmm0
; SSE-NEXT:    movaps 2096(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2288(%rdi), %xmm0
; SSE-NEXT:    movaps 2224(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2416(%rdi), %xmm0
; SSE-NEXT:    movaps 2352(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, (%rsp) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2544(%rdi), %xmm0
; SSE-NEXT:    movaps 2480(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2672(%rdi), %xmm0
; SSE-NEXT:    movaps 2608(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2800(%rdi), %xmm0
; SSE-NEXT:    movaps 2736(%rdi), %xmm1
; SSE-NEXT:    movaps %xmm1, %xmm2
; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movaps 2928(%rdi), %xmm0
; SSE-NEXT:    movaps 2864(%rdi), %xmm14
; SSE-NEXT:    movaps %xmm14, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
; SSE-NEXT:    movaps 3056(%rdi), %xmm0
; SSE-NEXT:    movaps 2992(%rdi), %xmm11
; SSE-NEXT:    movaps %xmm11, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1]
; SSE-NEXT:    movaps 3184(%rdi), %xmm0
; SSE-NEXT:    movaps 3120(%rdi), %xmm15
; SSE-NEXT:    movaps %xmm15, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]
; SSE-NEXT:    movaps 3312(%rdi), %xmm0
; SSE-NEXT:    movaps 3248(%rdi), %xmm9
; SSE-NEXT:    movaps %xmm9, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
; SSE-NEXT:    movaps 3440(%rdi), %xmm0
; SSE-NEXT:    movaps 3376(%rdi), %xmm12
; SSE-NEXT:    movaps %xmm12, %xmm1
; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]
; SSE-NEXT:    movaps 3568(%rdi), %xmm0
; SSE-NEXT:    movaps 3504(%rdi), %xmm8
; SSE-NEXT:    movaps %xmm8, %xmm13
; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
; SSE-NEXT:    movaps 4016(%rdi), %xmm4
; SSE-NEXT:    movaps 3952(%rdi), %xmm3
; SSE-NEXT:    movaps 3696(%rdi), %xmm0
; SSE-NEXT:    movaps 3632(%rdi), %xmm2
; SSE-NEXT:    movaps %xmm2, %xmm10
; SSE-NEXT:    movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]
; SSE-NEXT:    movaps 4080(%rdi), %xmm1
; SSE-NEXT:    movaps 3888(%rdi), %xmm5
; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
; SSE-NEXT:    movaps 3824(%rdi), %xmm6
; SSE-NEXT:    movaps 3760(%rdi), %xmm0
; SSE-NEXT:    movaps %xmm7, 496(%rsi)
; SSE-NEXT:    movaps %xmm0, %xmm7
; SSE-NEXT:    movlhps {{.*#+}} xmm7 = xmm7[0],xmm6[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm6[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; SSE-NEXT:    movaps %xmm6, 480(%rsi)
; SSE-NEXT:    movaps %xmm5, %xmm6
; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm3[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
; SSE-NEXT:    movaps %xmm4, %xmm3
; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]
; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 464(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 448(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 432(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 416(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 400(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 384(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 368(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 352(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 336(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 320(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 304(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 288(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 272(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 256(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 240(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 224(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 208(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 192(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 176(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 160(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 144(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 128(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 112(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 96(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 80(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 64(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 496(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 480(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 464(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 448(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 432(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 416(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 400(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 384(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 368(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 352(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 336(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 320(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 304(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 288(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 272(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 256(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 240(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 224(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 208(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 192(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 176(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 160(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 144(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 128(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 112(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 96(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 80(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 64(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 496(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 480(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 464(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 448(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 432(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 416(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 400(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 384(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 368(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 352(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 336(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 320(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 304(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 288(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 272(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 256(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 240(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 224(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 208(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 192(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 176(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 160(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 144(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 128(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 112(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 96(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 80(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 64(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 496(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 480(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 464(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 448(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 432(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 416(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 400(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 384(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 368(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 352(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 336(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 320(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 304(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 288(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 272(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 256(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 240(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 224(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 208(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 192(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 176(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 160(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 144(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 128(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 112(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 96(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 80(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 64(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 496(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 480(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 464(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 448(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 432(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 416(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 400(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 384(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 368(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 352(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 336(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 320(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 304(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 288(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 272(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 256(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 240(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 224(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 208(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 192(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 176(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 160(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 144(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 128(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 112(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 96(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 80(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 64(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 48(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 32(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 496(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 480(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 464(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 448(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 432(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 416(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 400(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 384(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 368(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 352(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 336(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 320(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 304(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 288(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 272(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 256(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 240(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 224(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 208(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 192(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 176(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 160(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 144(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 128(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm3, 496(%rax)
; SSE-NEXT:    movaps %xmm6, 480(%rax)
; SSE-NEXT:    movaps %xmm7, 464(%rax)
; SSE-NEXT:    movaps %xmm10, 448(%rax)
; SSE-NEXT:    movaps %xmm13, 432(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 416(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 400(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 384(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 368(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 352(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 336(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 320(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 304(%rax)
; SSE-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 288(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 272(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 256(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 240(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 224(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 208(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 192(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 176(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 160(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 144(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 128(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movaps %xmm4, 496(%rax)
; SSE-NEXT:    movaps %xmm5, 480(%rax)
; SSE-NEXT:    movaps %xmm0, 464(%rax)
; SSE-NEXT:    movaps %xmm2, 448(%rax)
; SSE-NEXT:    movaps %xmm8, 432(%rax)
; SSE-NEXT:    movaps %xmm12, 416(%rax)
; SSE-NEXT:    movaps %xmm9, 400(%rax)
; SSE-NEXT:    movaps %xmm15, 384(%rax)
; SSE-NEXT:    movaps %xmm11, 368(%rax)
; SSE-NEXT:    movaps %xmm14, 352(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 336(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 320(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 304(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 288(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 272(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 256(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 240(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 224(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 208(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 192(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 176(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 160(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 144(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 128(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 112(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 96(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 80(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 64(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, 16(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movaps %xmm0, (%rax)
; SSE-NEXT:    addq $3720, %rsp # imm = 0xE88
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i64_stride8_vf64:
; AVX:       # %bb.0:
; AVX-NEXT:    subq $5016, %rsp # imm = 0x1398
; AVX-NEXT:    vmovaps 2496(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2432(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3008(%rdi), %xmm2
; AVX-NEXT:    vmovaps 2944(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3520(%rdi), %xmm4
; AVX-NEXT:    vmovaps 3456(%rdi), %xmm5
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2880(%rdi), %xmm6
; AVX-NEXT:    vmovaps 3392(%rdi), %xmm7
; AVX-NEXT:    vmovaps 3904(%rdi), %xmm8
; AVX-NEXT:    vmovaps 3840(%rdi), %xmm9
; AVX-NEXT:    vmovaps 4032(%rdi), %xmm10
; AVX-NEXT:    vmovaps 3968(%rdi), %xmm11
; AVX-NEXT:    vmovlhps {{.*#+}} xmm12 = xmm11[0],xmm10[0]
; AVX-NEXT:    vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm10 = xmm11[1],xmm10[1]
; AVX-NEXT:    vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm10 = xmm9[0],xmm8[0]
; AVX-NEXT:    vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm8 = xmm9[1],xmm8[1]
; AVX-NEXT:    vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3328(%rdi), %xmm8
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm8[0],xmm7[0]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm4 = xmm8[1],xmm7[1]
; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2816(%rdi), %xmm4
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm4[0],xmm6[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm4[1],xmm6[1]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2368(%rdi), %xmm2
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2304(%rdi), %xmm0
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm2[0]
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1984(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1920(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1856(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1792(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1472(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1408(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1344(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1280(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 448(%rdi), %xmm0
; AVX-NEXT:    vmovaps 384(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 960(%rdi), %xmm0
; AVX-NEXT:    vmovaps 896(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 832(%rdi), %xmm0
; AVX-NEXT:    vmovaps 768(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3776(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3712(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3648(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3584(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3264(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3200(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3136(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3072(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2752(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2688(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2624(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2560(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1728(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1664(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1600(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1536(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 192(%rdi), %xmm0
; AVX-NEXT:    vmovaps 128(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX-NEXT:    vmovaps (%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 704(%rdi), %xmm0
; AVX-NEXT:    vmovaps 640(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX-NEXT:    vmovaps 512(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2240(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2176(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2112(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2048(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1216(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1152(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1088(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1024(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 448(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 384(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 336(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 272(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 960(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 848(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, (%rsp) # 16-byte Spill
; AVX-NEXT:    vmovaps 784(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1408(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1360(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1296(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1872(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1808(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2496(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2432(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 2384(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2320(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3008(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2944(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 2896(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2832(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3520(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3456(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 3408(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3344(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 4032(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3968(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 3920(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3856(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 192(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 128(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX-NEXT:    vmovaps 80(%rdi), %xmm0
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 16(%rdi), %xmm8
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm8[0],xmm0[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 704(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 640(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX-NEXT:    vmovaps 592(%rdi), %xmm0
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 528(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm1[0],xmm0[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm2[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1216(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1152(%rdi), %ymm2
; AVX-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm0[0],ymm2[2],ymm0[2]
; AVX-NEXT:    vmovaps 1104(%rdi), %xmm10
; AVX-NEXT:    vmovaps 1040(%rdi), %xmm2
; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm2[0],xmm10[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1728(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1664(%rdi), %ymm3
; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm0[0],ymm3[2],ymm0[2]
; AVX-NEXT:    vmovaps 1616(%rdi), %xmm11
; AVX-NEXT:    vmovaps 1552(%rdi), %xmm3
; AVX-NEXT:    vmovlhps {{.*#+}} xmm5 = xmm3[0],xmm11[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm4[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2240(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2176(%rdi), %ymm4
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm4[0],ymm0[0],ymm4[2],ymm0[2]
; AVX-NEXT:    vmovaps 2128(%rdi), %xmm12
; AVX-NEXT:    vmovaps 2064(%rdi), %xmm4
; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm12[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2752(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2688(%rdi), %ymm5
; AVX-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm0[0],ymm5[2],ymm0[2]
; AVX-NEXT:    vmovaps 2640(%rdi), %xmm13
; AVX-NEXT:    vmovaps 2576(%rdi), %xmm5
; AVX-NEXT:    vmovlhps {{.*#+}} xmm7 = xmm5[0],xmm13[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm6[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3264(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3200(%rdi), %ymm6
; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm6[0],ymm0[0],ymm6[2],ymm0[2]
; AVX-NEXT:    vmovaps 3152(%rdi), %xmm14
; AVX-NEXT:    vmovaps 3088(%rdi), %xmm6
; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm6[0],xmm14[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm7[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3776(%rdi), %ymm7
; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3712(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[2],ymm7[2]
; AVX-NEXT:    vmovaps 3664(%rdi), %xmm15
; AVX-NEXT:    vmovaps 3600(%rdi), %xmm7
; AVX-NEXT:    vmovlhps {{.*#+}} xmm9 = xmm7[0],xmm15[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX-NEXT:    # xmm8 = xmm8[1],mem[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX-NEXT:    # xmm8 = mem[0,1],xmm8[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = xmm1[1],mem[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm10[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm11[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm12[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm13[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm14[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm15[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 224(%rdi), %xmm0
; AVX-NEXT:    vmovaps 160(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 480(%rdi), %xmm0
; AVX-NEXT:    vmovaps 416(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 736(%rdi), %xmm0
; AVX-NEXT:    vmovaps 672(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 992(%rdi), %xmm0
; AVX-NEXT:    vmovaps 928(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
; AVX-NEXT:    vmovaps 1248(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1184(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1504(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1440(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1760(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1696(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2016(%rdi), %xmm0
; AVX-NEXT:    vmovaps 1952(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2144(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2080(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2272(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2208(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2400(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2336(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2528(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2464(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2656(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2592(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2784(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2720(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2912(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2848(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3040(%rdi), %xmm0
; AVX-NEXT:    vmovaps 2976(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3168(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3104(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3296(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3232(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3424(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3360(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3552(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3488(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3680(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3616(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3808(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3744(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 3936(%rdi), %xmm0
; AVX-NEXT:    vmovaps 3872(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 4064(%rdi), %xmm0
; AVX-NEXT:    vmovaps 4000(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 112(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 48(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 368(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 304(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 624(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 560(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 880(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 816(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1248(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1184(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1136(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1072(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1504(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1440(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1392(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1328(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1760(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1696(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1648(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1584(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2016(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 1952(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 1904(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 1840(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2272(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2208(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 2160(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2096(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2528(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2464(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 2416(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2352(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2784(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2720(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 2672(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2608(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3040(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 2976(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 2928(%rdi), %xmm2
; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovaps 2864(%rdi), %xmm1
; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3296(%rdi), %ymm1
; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3232(%rdi), %ymm0
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX-NEXT:    vmovaps 3184(%rdi), %xmm14
; AVX-NEXT:    vmovaps 3120(%rdi), %xmm13
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm13[0],xmm14[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3552(%rdi), %ymm12
; AVX-NEXT:    vmovaps 3488(%rdi), %ymm11
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX-NEXT:    vmovaps 3440(%rdi), %xmm10
; AVX-NEXT:    vmovaps 3376(%rdi), %xmm9
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm9[0],xmm10[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 3808(%rdi), %ymm8
; AVX-NEXT:    vmovaps 3744(%rdi), %ymm7
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX-NEXT:    vmovaps 3696(%rdi), %xmm6
; AVX-NEXT:    vmovaps 3632(%rdi), %xmm5
; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm6[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovaps 4064(%rdi), %ymm4
; AVX-NEXT:    vmovaps 4000(%rdi), %ymm3
; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX-NEXT:    vmovaps 3952(%rdi), %xmm2
; AVX-NEXT:    vmovaps 3888(%rdi), %xmm1
; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm1[0],xmm2[0]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vmovlps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload
; AVX-NEXT:    # xmm15 = mem[0,1],xmm15[2,3]
; AVX-NEXT:    vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm13 = xmm13[1],xmm14[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm9 = xmm9[1],xmm10[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm5 = xmm5[1],xmm6[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 464(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 448(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 256(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 384(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 320(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 272(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 400(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 336(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 496(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 480(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 416(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 352(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 288(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 432(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 368(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 304(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 256(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 272(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 320(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 336(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 384(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 400(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 448(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 464(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 288(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 304(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 352(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 368(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 416(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 432(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 480(%rdx)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 496(%rdx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 448(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 384(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 320(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 256(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 480(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 416(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 352(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 288(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 480(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 448(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 416(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 384(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 352(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 320(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 288(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 256(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r8)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 496(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 480(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 464(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 448(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 432(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 416(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 400(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 384(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 368(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 352(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 336(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 320(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 304(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 288(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 272(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 256(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%r9)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%r9)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 496(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 480(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 464(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 448(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 432(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 416(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 400(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 384(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 368(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 352(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 336(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 320(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 304(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 288(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 272(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 256(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 240(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 224(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 208(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 192(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 176(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 160(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 144(%rax)
; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 128(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 112(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 96(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 80(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 64(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 48(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 32(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, 16(%rax)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm1, (%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 480(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 448(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 416(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 384(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 320(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 288(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 256(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX-NEXT:    vmovaps %ymm5, 448(%rax)
; AVX-NEXT:    vmovaps %ymm9, 416(%rax)
; AVX-NEXT:    vmovaps %ymm13, 384(%rax)
; AVX-NEXT:    vmovaps %ymm15, 352(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm0, (%rax)
; AVX-NEXT:    addq $5016, %rsp # imm = 0x1398
; AVX-NEXT:    vzeroupper
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i64_stride8_vf64:
; AVX2:       # %bb.0:
; AVX2-NEXT:    subq $5064, %rsp # imm = 0x13C8
; AVX2-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 448(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 832(%rdi), %xmm2
; AVX2-NEXT:    vinsertf128 $1, 960(%rdi), %ymm2, %ymm2
; AVX2-NEXT:    vmovaps 768(%rdi), %xmm3
; AVX2-NEXT:    vinsertf128 $1, 896(%rdi), %ymm3, %ymm3
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1344(%rdi), %xmm4
; AVX2-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm4, %ymm4
; AVX2-NEXT:    vmovaps 1280(%rdi), %xmm5
; AVX2-NEXT:    vinsertf128 $1, 1408(%rdi), %ymm5, %ymm5
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1856(%rdi), %xmm6
; AVX2-NEXT:    vinsertf128 $1, 1984(%rdi), %ymm6, %ymm6
; AVX2-NEXT:    vmovaps 1792(%rdi), %xmm7
; AVX2-NEXT:    vinsertf128 $1, 1920(%rdi), %ymm7, %ymm7
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2368(%rdi), %xmm8
; AVX2-NEXT:    vinsertf128 $1, 2496(%rdi), %ymm8, %ymm8
; AVX2-NEXT:    vmovaps 2304(%rdi), %xmm9
; AVX2-NEXT:    vinsertf128 $1, 2432(%rdi), %ymm9, %ymm9
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2880(%rdi), %xmm10
; AVX2-NEXT:    vinsertf128 $1, 3008(%rdi), %ymm10, %ymm10
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2816(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2944(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3392(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 3520(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3328(%rdi), %xmm2
; AVX2-NEXT:    vinsertf128 $1, 3456(%rdi), %ymm2, %ymm2
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3904(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 4032(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3840(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 3968(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 192(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps (%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 128(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 512(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 640(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1088(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1216(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1024(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1152(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1600(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1728(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1536(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2112(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2240(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 2048(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 2176(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2624(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2752(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 2560(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 2688(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3136(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 3264(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 3072(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 3200(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3648(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 3776(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 3584(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 3712(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 320(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 256(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 448(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 384(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 832(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 768(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 960(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1344(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1280(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1408(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1856(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1792(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2368(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2304(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2496(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2432(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2880(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2816(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3008(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2944(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3392(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3328(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3520(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3456(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3904(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3840(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 4032(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3968(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps (%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 128(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 576(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 512(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 704(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 640(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1088(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1024(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1216(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1152(%rdi), %ymm15
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1600(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1536(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1728(%rdi), %ymm14
; AVX2-NEXT:    vmovaps 1664(%rdi), %ymm11
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm14[0],ymm11[2],ymm14[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2112(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2048(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2240(%rdi), %ymm10
; AVX2-NEXT:    vmovaps 2176(%rdi), %ymm8
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2624(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2560(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2752(%rdi), %ymm7
; AVX2-NEXT:    vmovaps 2688(%rdi), %ymm5
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm7[0],ymm5[2],ymm7[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3136(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3072(%rdi), %ymm9
; AVX2-NEXT:    vmovaps 3264(%rdi), %ymm4
; AVX2-NEXT:    vmovaps 3200(%rdi), %ymm3
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm9[0],ymm1[0],ymm9[2],ymm1[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm2[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3648(%rdi), %ymm12
; AVX2-NEXT:    vmovaps 3584(%rdi), %ymm6
; AVX2-NEXT:    vmovaps 3776(%rdi), %ymm2
; AVX2-NEXT:    vmovaps 3712(%rdi), %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm6[0],ymm12[0],ymm6[2],ymm12[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm14[1],ymm11[3],ymm14[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm8 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm8[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm8 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm8[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm7[1],ymm5[3],ymm7[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm3 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm3 = ymm9[1],mem[1],ymm9[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm3 = ymm3[1],mem[1],ymm3[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm6[1],ymm12[1],ymm6[3],ymm12[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1184(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 1760(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 1952(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2144(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2272(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 2080(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 2208(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2400(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2528(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 2336(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 2464(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2656(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 2784(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 2592(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 2720(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2912(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 3040(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 2848(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 2976(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3168(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 3296(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 3104(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 3232(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3424(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 3552(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 3360(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 3488(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3680(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 3808(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 3616(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 3744(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3936(%rdi), %xmm0
; AVX2-NEXT:    vinsertf128 $1, 4064(%rdi), %ymm0, %ymm0
; AVX2-NEXT:    vmovaps 3872(%rdi), %xmm1
; AVX2-NEXT:    vinsertf128 $1, 4000(%rdi), %ymm1, %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 32(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 352(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX2-NEXT:    vmovaps 288(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 608(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 544(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 864(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 800(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1120(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1056(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1248(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1184(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1376(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1312(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1504(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1440(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1632(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1568(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1760(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1696(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1888(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1824(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2016(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 1952(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2144(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2080(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2272(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2208(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2400(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2336(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2528(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2464(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2656(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2592(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2784(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2720(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2912(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2848(%rdi), %ymm3
; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3040(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 2976(%rdi), %ymm0
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3168(%rdi), %ymm2
; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3104(%rdi), %ymm1
; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3296(%rdi), %ymm14
; AVX2-NEXT:    vmovaps 3232(%rdi), %ymm13
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm14[0],ymm13[2],ymm14[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3424(%rdi), %ymm12
; AVX2-NEXT:    vmovaps 3360(%rdi), %ymm11
; AVX2-NEXT:    vmovaps 3552(%rdi), %ymm10
; AVX2-NEXT:    vmovaps 3488(%rdi), %ymm9
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3680(%rdi), %ymm8
; AVX2-NEXT:    vmovaps 3616(%rdi), %ymm7
; AVX2-NEXT:    vmovaps 3808(%rdi), %ymm6
; AVX2-NEXT:    vmovaps 3744(%rdi), %ymm5
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovaps 3936(%rdi), %ymm4
; AVX2-NEXT:    vmovaps 3872(%rdi), %ymm3
; AVX2-NEXT:    vmovaps 4064(%rdi), %ymm2
; AVX2-NEXT:    vmovaps 4000(%rdi), %ymm1
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd (%rsp), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm15 = ymm15[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm13[1],ymm14[1],ymm13[3],ymm14[3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm9 = ymm9[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm5[2,3],ymm0[2,3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 448(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 384(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 320(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 256(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 480(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 416(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 352(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 288(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 448(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 384(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 320(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 256(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 480(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 416(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 352(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 288(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 448(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 384(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 320(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 256(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 480(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 416(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 352(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 288(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 480(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 448(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 416(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 384(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 352(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 320(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 288(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 256(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 480(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 448(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 416(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 384(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 352(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 320(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 288(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 256(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 480(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 448(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 416(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 384(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 320(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 288(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 256(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 480(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 448(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 416(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 384(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 320(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 288(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 256(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX2-NEXT:    vmovaps %ymm5, 448(%rax)
; AVX2-NEXT:    vmovaps %ymm9, 416(%rax)
; AVX2-NEXT:    vmovaps %ymm13, 384(%rax)
; AVX2-NEXT:    vmovaps %ymm15, 352(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-NEXT:    addq $5064, %rsp # imm = 0x13C8
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i64_stride8_vf64:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    subq $5064, %rsp # imm = 0x13C8
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 832(%rdi), %xmm2
; AVX2-FP-NEXT:    vinsertf128 $1, 960(%rdi), %ymm2, %ymm2
; AVX2-FP-NEXT:    vmovaps 768(%rdi), %xmm3
; AVX2-FP-NEXT:    vinsertf128 $1, 896(%rdi), %ymm3, %ymm3
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1344(%rdi), %xmm4
; AVX2-FP-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm4, %ymm4
; AVX2-FP-NEXT:    vmovaps 1280(%rdi), %xmm5
; AVX2-FP-NEXT:    vinsertf128 $1, 1408(%rdi), %ymm5, %ymm5
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1856(%rdi), %xmm6
; AVX2-FP-NEXT:    vinsertf128 $1, 1984(%rdi), %ymm6, %ymm6
; AVX2-FP-NEXT:    vmovaps 1792(%rdi), %xmm7
; AVX2-FP-NEXT:    vinsertf128 $1, 1920(%rdi), %ymm7, %ymm7
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2368(%rdi), %xmm8
; AVX2-FP-NEXT:    vinsertf128 $1, 2496(%rdi), %ymm8, %ymm8
; AVX2-FP-NEXT:    vmovaps 2304(%rdi), %xmm9
; AVX2-FP-NEXT:    vinsertf128 $1, 2432(%rdi), %ymm9, %ymm9
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2880(%rdi), %xmm10
; AVX2-FP-NEXT:    vinsertf128 $1, 3008(%rdi), %ymm10, %ymm10
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2816(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2944(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3392(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 3520(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3328(%rdi), %xmm2
; AVX2-FP-NEXT:    vinsertf128 $1, 3456(%rdi), %ymm2, %ymm2
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3904(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 4032(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3840(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 3968(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 512(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 640(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1088(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1216(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1024(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1152(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1600(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1728(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1536(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2112(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2240(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 2048(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 2176(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2624(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2752(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 2560(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 2688(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3136(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 3264(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 3072(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 3200(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3648(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 3776(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 3584(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 3712(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 320(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 256(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 448(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 384(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 832(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 768(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 960(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1344(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1280(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1408(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1856(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1792(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2368(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2304(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2496(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2432(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2880(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2816(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3008(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2944(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3392(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3328(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3520(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3456(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3904(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3840(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 4032(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3968(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 576(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 512(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 704(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 640(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1088(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1024(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1216(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1152(%rdi), %ymm15
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1600(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1536(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1728(%rdi), %ymm14
; AVX2-FP-NEXT:    vmovaps 1664(%rdi), %ymm11
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm14[0],ymm11[2],ymm14[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2112(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2048(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2240(%rdi), %ymm10
; AVX2-FP-NEXT:    vmovaps 2176(%rdi), %ymm8
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2624(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2560(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2752(%rdi), %ymm7
; AVX2-FP-NEXT:    vmovaps 2688(%rdi), %ymm5
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm7[0],ymm5[2],ymm7[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3136(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3072(%rdi), %ymm9
; AVX2-FP-NEXT:    vmovaps 3264(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovaps 3200(%rdi), %ymm3
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm9[0],ymm1[0],ymm9[2],ymm1[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm2[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3648(%rdi), %ymm12
; AVX2-FP-NEXT:    vmovaps 3584(%rdi), %ymm6
; AVX2-FP-NEXT:    vmovaps 3776(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps 3712(%rdi), %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm6[0],ymm12[0],ymm6[2],ymm12[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm14[1],ymm11[3],ymm14[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm8 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm8[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm8 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm8[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm7[1],ymm5[3],ymm7[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm3 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm3 = ymm9[1],mem[1],ymm9[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm3 = ymm3[1],mem[1],ymm3[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm6[1],ymm12[1],ymm6[3],ymm12[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1184(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 1760(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 1952(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2144(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2272(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 2080(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 2208(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2400(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2528(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 2336(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 2464(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2656(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 2784(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 2592(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 2720(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2912(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 3040(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 2848(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 2976(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3168(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 3296(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 3104(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 3232(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3424(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 3552(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 3360(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 3488(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3680(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 3808(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 3616(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 3744(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3936(%rdi), %xmm0
; AVX2-FP-NEXT:    vinsertf128 $1, 4064(%rdi), %ymm0, %ymm0
; AVX2-FP-NEXT:    vmovaps 3872(%rdi), %xmm1
; AVX2-FP-NEXT:    vinsertf128 $1, 4000(%rdi), %ymm1, %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 352(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 288(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 608(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 544(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 864(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 800(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1120(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1056(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1248(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1184(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1376(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1312(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1504(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1440(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1632(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1568(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1760(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1696(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1888(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1824(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2016(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 1952(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2144(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2080(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2272(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2208(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2400(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2336(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2528(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2464(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2656(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2592(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2784(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2720(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2912(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2848(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3040(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 2976(%rdi), %ymm0
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3168(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3104(%rdi), %ymm1
; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3296(%rdi), %ymm14
; AVX2-FP-NEXT:    vmovaps 3232(%rdi), %ymm13
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm14[0],ymm13[2],ymm14[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3424(%rdi), %ymm12
; AVX2-FP-NEXT:    vmovaps 3360(%rdi), %ymm11
; AVX2-FP-NEXT:    vmovaps 3552(%rdi), %ymm10
; AVX2-FP-NEXT:    vmovaps 3488(%rdi), %ymm9
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3680(%rdi), %ymm8
; AVX2-FP-NEXT:    vmovaps 3616(%rdi), %ymm7
; AVX2-FP-NEXT:    vmovaps 3808(%rdi), %ymm6
; AVX2-FP-NEXT:    vmovaps 3744(%rdi), %ymm5
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovaps 3936(%rdi), %ymm4
; AVX2-FP-NEXT:    vmovaps 3872(%rdi), %ymm3
; AVX2-FP-NEXT:    vmovaps 4064(%rdi), %ymm2
; AVX2-FP-NEXT:    vmovaps 4000(%rdi), %ymm1
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd (%rsp), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm15 = ymm15[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm13[1],ymm14[1],ymm13[3],ymm14[3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm9 = ymm9[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm5[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-FP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 480(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 448(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 416(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 384(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 320(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 288(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 256(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm5, 448(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm9, 416(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm13, 384(%rax)
; AVX2-FP-NEXT:    vmovaps %ymm15, 352(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FP-NEXT:    addq $5064, %rsp # imm = 0x13C8
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i64_stride8_vf64:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    subq $5064, %rsp # imm = 0x13C8
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 448(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 384(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %xmm2
; AVX2-FCP-NEXT:    vinsertf128 $1, 960(%rdi), %ymm2, %ymm2
; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %xmm3
; AVX2-FCP-NEXT:    vinsertf128 $1, 896(%rdi), %ymm3, %ymm3
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1344(%rdi), %xmm4
; AVX2-FCP-NEXT:    vinsertf128 $1, 1472(%rdi), %ymm4, %ymm4
; AVX2-FCP-NEXT:    vmovaps 1280(%rdi), %xmm5
; AVX2-FCP-NEXT:    vinsertf128 $1, 1408(%rdi), %ymm5, %ymm5
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1856(%rdi), %xmm6
; AVX2-FCP-NEXT:    vinsertf128 $1, 1984(%rdi), %ymm6, %ymm6
; AVX2-FCP-NEXT:    vmovaps 1792(%rdi), %xmm7
; AVX2-FCP-NEXT:    vinsertf128 $1, 1920(%rdi), %ymm7, %ymm7
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
; AVX2-FCP-NEXT:    vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2368(%rdi), %xmm8
; AVX2-FCP-NEXT:    vinsertf128 $1, 2496(%rdi), %ymm8, %ymm8
; AVX2-FCP-NEXT:    vmovaps 2304(%rdi), %xmm9
; AVX2-FCP-NEXT:    vinsertf128 $1, 2432(%rdi), %ymm9, %ymm9
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
; AVX2-FCP-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2880(%rdi), %xmm10
; AVX2-FCP-NEXT:    vinsertf128 $1, 3008(%rdi), %ymm10, %ymm10
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2816(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2944(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm10[0],ymm0[2],ymm10[2]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3392(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 3520(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm7[1],ymm6[1],ymm7[3],ymm6[3]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3328(%rdi), %xmm2
; AVX2-FCP-NEXT:    vinsertf128 $1, 3456(%rdi), %ymm2, %ymm2
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm10[1],ymm0[3],ymm10[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3904(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 4032(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3840(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 3968(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 192(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 128(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 704(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 640(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1088(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1216(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1024(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1152(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1600(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1728(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1536(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1664(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2112(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2240(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 2048(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 2176(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2624(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2752(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 2560(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 2688(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3136(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 3264(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 3072(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 3200(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3648(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 3776(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 3584(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 3712(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 320(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 256(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 448(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 384(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 832(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 768(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 960(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 896(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1344(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1280(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1472(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1408(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1856(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1792(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1984(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1920(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2368(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2304(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2496(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2432(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2880(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2816(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3008(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2944(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3392(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3328(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3520(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3456(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3904(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3840(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 4032(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3968(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 576(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 512(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 704(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 640(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1088(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1024(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1216(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1152(%rdi), %ymm15
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1600(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1536(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1728(%rdi), %ymm14
; AVX2-FCP-NEXT:    vmovaps 1664(%rdi), %ymm11
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm11[0],ymm14[0],ymm11[2],ymm14[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2112(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2048(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2240(%rdi), %ymm10
; AVX2-FCP-NEXT:    vmovaps 2176(%rdi), %ymm8
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm8[0],ymm10[0],ymm8[2],ymm10[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2624(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2560(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2752(%rdi), %ymm7
; AVX2-FCP-NEXT:    vmovaps 2688(%rdi), %ymm5
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm7[0],ymm5[2],ymm7[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3136(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3072(%rdi), %ymm9
; AVX2-FCP-NEXT:    vmovaps 3264(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovaps 3200(%rdi), %ymm3
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm9[0],ymm1[0],ymm9[2],ymm1[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm2[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3648(%rdi), %ymm12
; AVX2-FCP-NEXT:    vmovaps 3584(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovaps 3776(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps 3712(%rdi), %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm13 = ymm6[0],ymm12[0],ymm6[2],ymm12[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm11[1],ymm14[1],ymm11[3],ymm14[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm11 = ymm11[1],mem[1],ymm11[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm11[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm10[1],ymm8[3],ymm10[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm8 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm8[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm8 = ymm8[1],mem[1],ymm8[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm8[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm7[1],ymm5[3],ymm7[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm3 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm3 = ymm9[1],mem[1],ymm9[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm3 = ymm3[1],mem[1],ymm3[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm6[1],ymm12[1],ymm6[3],ymm12[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm1 = ymm1[1],mem[1],ymm1[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 224(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 480(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 416(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 736(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 672(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 992(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 928(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1120(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1248(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1056(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1184(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1376(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1504(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1312(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1440(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1632(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 1760(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1568(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1696(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1888(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2016(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 1824(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 1952(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2144(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2272(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 2080(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 2208(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2400(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2528(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 2336(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 2464(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2656(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 2784(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 2592(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 2720(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2912(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 3040(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 2848(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 2976(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3168(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 3296(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 3104(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 3232(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3424(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 3552(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 3360(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 3488(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3680(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 3808(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 3616(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 3744(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3936(%rdi), %xmm0
; AVX2-FCP-NEXT:    vinsertf128 $1, 4064(%rdi), %ymm0, %ymm0
; AVX2-FCP-NEXT:    vmovaps 3872(%rdi), %xmm1
; AVX2-FCP-NEXT:    vinsertf128 $1, 4000(%rdi), %ymm1, %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 352(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 288(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 480(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 416(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 608(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 544(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 736(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 672(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 864(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 800(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 992(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 928(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1120(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1056(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1248(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1184(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1376(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1312(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1504(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1440(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1632(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1568(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1760(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1696(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1888(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1824(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2016(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 1952(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2144(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2080(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2272(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2208(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2400(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2336(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2528(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2464(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2656(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2592(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2784(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2720(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2912(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2848(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3040(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 2976(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3168(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3104(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3296(%rdi), %ymm14
; AVX2-FCP-NEXT:    vmovaps 3232(%rdi), %ymm13
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm14[0],ymm13[2],ymm14[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3424(%rdi), %ymm12
; AVX2-FCP-NEXT:    vmovaps 3360(%rdi), %ymm11
; AVX2-FCP-NEXT:    vmovaps 3552(%rdi), %ymm10
; AVX2-FCP-NEXT:    vmovaps 3488(%rdi), %ymm9
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[2],ymm10[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3680(%rdi), %ymm8
; AVX2-FCP-NEXT:    vmovaps 3616(%rdi), %ymm7
; AVX2-FCP-NEXT:    vmovaps 3808(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovaps 3744(%rdi), %ymm5
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovaps 3936(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovaps 3872(%rdi), %ymm3
; AVX2-FCP-NEXT:    vmovaps 4064(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovaps 4000(%rdi), %ymm1
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd (%rsp), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm15 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm15 = ymm15[1],mem[1],ymm15[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm15 = ymm15[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm13[1],ymm14[1],ymm13[3],ymm14[3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-FCP-NEXT:    vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    # ymm13 = ymm13[1],mem[1],ymm13[3],mem[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm9[1],ymm10[1],ymm9[3],ymm10[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm9 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm9 = ymm9[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm5[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
; AVX2-FCP-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 480(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 448(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 416(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 384(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 352(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 320(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 288(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 256(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 192(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 160(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 128(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 64(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovaps %ymm0, 480(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm5, 448(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm9, 416(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm13, 384(%rax)
; AVX2-FCP-NEXT:    vmovaps %ymm15, 352(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 320(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 288(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 256(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 224(%rax)
; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 192(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 160(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 128(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 96(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 64(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)
; AVX2-FCP-NEXT:    addq $5064, %rsp # imm = 0x13C8
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i64_stride8_vf64:
; AVX512:       # %bb.0:
; AVX512-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    movb $-64, %al
; AVX512-NEXT:    kmovw %eax, %k1
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i64_stride8_vf64:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512-FCP-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512-FCP-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512-FCP-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    movb $-64, %al
; AVX512-FCP-NEXT:    kmovw %eax, %k1
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512-FCP-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512-FCP-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512-FCP-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512-FCP-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512-FCP-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512-FCP-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512-FCP-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512-FCP-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512-FCP-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512-FCP-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512-FCP-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512-FCP-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512-FCP-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512-FCP-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512-FCP-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512-FCP-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512-FCP-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512-FCP-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512-FCP-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512-FCP-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512-FCP-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512-FCP-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512-FCP-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512-FCP-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512-FCP-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512-FCP-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512-FCP-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512-FCP-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512-FCP-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512-FCP-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512-FCP-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512-FCP-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512-FCP-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512-FCP-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512-FCP-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512-FCP-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512-FCP-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512-FCP-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512-FCP-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i64_stride8_vf64:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512DQ-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512DQ-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512DQ-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512DQ-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512DQ-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512DQ-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512DQ-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512DQ-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    movb $-64, %al
; AVX512DQ-NEXT:    kmovw %eax, %k1
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512DQ-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512DQ-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512DQ-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512DQ-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512DQ-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512DQ-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512DQ-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512DQ-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512DQ-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512DQ-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512DQ-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512DQ-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512DQ-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512DQ-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512DQ-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512DQ-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512DQ-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512DQ-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512DQ-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512DQ-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512DQ-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512DQ-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512DQ-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512DQ-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512DQ-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512DQ-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQ-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512DQ-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512DQ-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512DQ-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512DQ-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512DQ-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512DQ-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512DQ-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512DQ-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512DQ-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512DQ-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512DQ-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512DQ-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512DQ-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512DQ-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512DQ-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512DQ-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512DQ-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512DQ-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512DQ-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512DQ-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512DQ-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512DQ-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512DQ-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512DQ-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512DQ-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512DQ-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512DQ-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512DQ-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512DQ-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512DQ-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512DQ-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512DQ-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512DQ-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512DQ-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512DQ-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512DQ-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512DQ-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512DQ-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512DQ-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512DQ-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512DQ-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512DQ-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512DQ-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512DQ-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512DQ-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512DQ-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i64_stride8_vf64:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512DQ-FCP-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512DQ-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    movb $-64, %al
; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQ-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-FCP-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512DQ-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-FCP-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512DQ-FCP-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512DQ-FCP-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512DQ-FCP-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512DQ-FCP-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512DQ-FCP-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512DQ-FCP-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512DQ-FCP-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512DQ-FCP-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512DQ-FCP-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQ-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512DQ-FCP-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQ-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512DQ-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512DQ-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512DQ-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512DQ-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512DQ-FCP-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512DQ-FCP-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512DQ-FCP-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512DQ-FCP-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512DQ-FCP-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512DQ-FCP-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512DQ-FCP-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-FCP-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i64_stride8_vf64:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512BW-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512BW-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512BW-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512BW-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512BW-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512BW-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    movb $-64, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512BW-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512BW-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512BW-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512BW-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512BW-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512BW-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512BW-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512BW-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512BW-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512BW-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512BW-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512BW-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512BW-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512BW-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512BW-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512BW-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512BW-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512BW-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512BW-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512BW-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512BW-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512BW-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512BW-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512BW-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512BW-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512BW-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512BW-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512BW-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512BW-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512BW-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512BW-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512BW-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512BW-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512BW-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512BW-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512BW-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512BW-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512BW-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512BW-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512BW-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512BW-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512BW-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512BW-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512BW-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512BW-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512BW-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512BW-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512BW-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512BW-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512BW-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512BW-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512BW-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512BW-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512BW-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512BW-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512BW-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512BW-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512BW-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512BW-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512BW-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512BW-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512BW-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512BW-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512BW-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i64_stride8_vf64:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512BW-FCP-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512BW-FCP-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512BW-FCP-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512BW-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    movb $-64, %al
; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-FCP-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-FCP-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512BW-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512BW-FCP-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512BW-FCP-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512BW-FCP-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512BW-FCP-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512BW-FCP-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512BW-FCP-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512BW-FCP-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512BW-FCP-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512BW-FCP-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512BW-FCP-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512BW-FCP-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512BW-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512BW-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512BW-FCP-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512BW-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512BW-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512BW-FCP-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512BW-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512BW-FCP-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512BW-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512BW-FCP-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512BW-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512BW-FCP-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512BW-FCP-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512BW-FCP-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512BW-FCP-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512BW-FCP-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512BW-FCP-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512BW-FCP-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512BW-FCP-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512BW-FCP-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512BW-FCP-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512BW-FCP-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512BW-FCP-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512BW-FCP-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512BW-FCP-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512BW-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512BW-FCP-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i64_stride8_vf64:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512DQ-BW-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512DQ-BW-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512DQ-BW-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512DQ-BW-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    movb $-64, %al
; AVX512DQ-BW-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512DQ-BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-BW-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512DQ-BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512DQ-BW-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512DQ-BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-BW-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512DQ-BW-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512DQ-BW-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-BW-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512DQ-BW-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512DQ-BW-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512DQ-BW-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512DQ-BW-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512DQ-BW-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512DQ-BW-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512DQ-BW-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512DQ-BW-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512DQ-BW-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512DQ-BW-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-BW-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-BW-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512DQ-BW-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512DQ-BW-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512DQ-BW-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512DQ-BW-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512DQ-BW-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-BW-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512DQ-BW-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512DQ-BW-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512DQ-BW-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512DQ-BW-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512DQ-BW-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512DQ-BW-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512DQ-BW-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512DQ-BW-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512DQ-BW-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512DQ-BW-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512DQ-BW-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512DQ-BW-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512DQ-BW-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512DQ-BW-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512DQ-BW-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512DQ-BW-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512DQ-BW-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512DQ-BW-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512DQ-BW-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512DQ-BW-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512DQ-BW-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512DQ-BW-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512DQ-BW-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512DQ-BW-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-BW-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i64_stride8_vf64:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    subq $6664, %rsp # imm = 0x1A08
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3392(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3328(%rdi), %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3520(%rdi), %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3456(%rdi), %zmm28
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1856(%rdi), %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1984(%rdi), %zmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 832(%rdi), %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 768(%rdi), %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 960(%rdi), %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 896(%rdi), %zmm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %zmm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 448(%rdi), %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 384(%rdi), %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    movb $-64, %al
; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3264(%rdi), %ymm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3200(%rdi), %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3136(%rdi), %ymm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3072(%rdi), %ymm7
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm7[0],ymm3[0],ymm7[2],ymm3[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 704(%rdi), %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 640(%rdi), %ymm20
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm20[0],ymm0[0],ymm20[2],ymm0[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 576(%rdi), %ymm22
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %ymm19
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm19[0],ymm22[0],ymm19[2],ymm22[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm12, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %ymm21
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm13[0],ymm21[0],ymm13[2],ymm21[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1920(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1792(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1728(%rdi), %ymm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1664(%rdi), %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1600(%rdi), %ymm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1536(%rdi), %ymm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1472(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1408(%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1344(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1280(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1216(%rdi), %ymm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1152(%rdi), %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1088(%rdi), %ymm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1024(%rdi), %ymm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3008(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2944(%rdi), %zmm29
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2880(%rdi), %zmm24
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2816(%rdi), %zmm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2752(%rdi), %ymm27
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2688(%rdi), %ymm26
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm26[0],ymm27[0],ymm26[2],ymm27[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2624(%rdi), %ymm30
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2560(%rdi), %ymm18
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm18[0],ymm30[0],ymm18[2],ymm30[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2496(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2432(%rdi), %zmm31
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2368(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2304(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 2240(%rdi), %ymm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 2176(%rdi), %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm0 = ymm11[0],ymm12[0],ymm11[2],ymm12[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 2112(%rdi), %ymm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 2048(%rdi), %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm3[0],ymm10[0],ymm3[2],ymm10[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm5[2,3],ymm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 4032(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3968(%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3904(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3840(%rdi), %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm1, %zmm4, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3776(%rdi), %ymm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3712(%rdi), %ymm23
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm23[0],ymm17[0],ymm23[2],ymm17[2]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3648(%rdi), %ymm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3584(%rdi), %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm2, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm28
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm2, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm5 = ymm4[1],mem[1],ymm4[3],mem[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm20, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm5 = ymm20[1],mem[1],ymm20[3],mem[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm19[1],ymm22[1],ymm19[3],ymm22[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm2, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm22, %zmm2, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm13[1],ymm21[1],ymm13[3],ymm21[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm15, %zmm2, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm2, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm5 = ymm5[1],mem[1],ymm5[3],mem[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm7 = ymm7[1],mem[1],ymm7[3],mem[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm6, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm29 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm25 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm26[1],ymm27[1],ymm26[3],ymm27[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm18[1],ymm30[1],ymm18[3],ymm30[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm5 = ymm7[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm25, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm2, %zmm31
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm6 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm12[1],ymm11[3],ymm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm3[1],ymm10[1],ymm3[3],ymm10[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm5[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm6, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm11, %zmm12, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm3 = ymm23[1],ymm17[1],ymm23[3],ymm17[3]
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; AVX512DQ-BW-FCP-NEXT:    vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3264(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3200(%rdi), %zmm10
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3136(%rdi), %zmm23
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3072(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm8[0],zmm16[0],zmm8[2],zmm16[2],zmm8[4],zmm16[4],zmm8[6],zmm16[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 704(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 640(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 576(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm9[0],zmm4[0],zmm9[2],zmm4[2],zmm9[4],zmm4[4],zmm9[6],zmm4[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm19, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm3[0],zmm22[0],zmm3[2],zmm22[2],zmm3[4],zmm22[4],zmm3[6],zmm22[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1728(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1664(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1600(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1536(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm15, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm21[0],zmm13[0],zmm21[2],zmm13[2],zmm21[4],zmm13[4],zmm21[6],zmm13[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1216(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1152(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1088(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 1024(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm27 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # zmm3 = zmm27[0],mem[0],zmm27[2],mem[2],zmm27[4],mem[4],zmm27[6],mem[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2752(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2688(%rdi), %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2624(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2560(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # zmm3 = zmm7[0],mem[0],zmm7[2],mem[2],zmm7[4],mem[4],zmm7[6],mem[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2240(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2176(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2112(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2048(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm3 = zmm26[0],zmm4[0],zmm26[2],zmm4[2],zmm26[4],zmm4[4],zmm26[6],zmm4[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3776(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3712(%rdi), %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3648(%rdi), %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3584(%rdi), %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm18, %zmm29, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 = zmm12[0],zmm11[0],zmm12[2],zmm11[2],zmm12[4],zmm11[4],zmm12[6],zmm11[6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm19[1],zmm25[1],zmm19[3],zmm25[3],zmm19[5],zmm25[5],zmm19[7],zmm25[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm29
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm30, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm15[1],zmm28[1],zmm15[3],zmm28[3],zmm15[5],zmm28[5],zmm15[7],zmm28[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm12[1],zmm18[1],zmm12[3],zmm18[3],zmm12[5],zmm18[5],zmm12[7],zmm18[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm23, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm13[1],zmm3[3],zmm13[3],zmm3[5],zmm13[5],zmm3[7],zmm13[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # zmm3 = zmm27[1],mem[1],zmm27[3],mem[3],zmm27[5],mem[5],zmm27[7],mem[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # zmm3 = zmm7[1],mem[1],zmm7[3],mem[3],zmm7[5],mem[5],zmm7[7],mem[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm5 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm3 = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm10, %zmm9, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # zmm2 = zmm2[1],mem[1],zmm2[3],mem[3],zmm2[5],mem[5],zmm2[7],mem[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm0 = [6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm26, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # zmm2 {%k1} = zmm25[0],mem[0],zmm25[2],mem[2],zmm25[4],mem[4],zmm25[6],mem[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm22, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm24, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm28, %zmm0, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm15 {%k1} = zmm29[0],zmm30[0],zmm29[2],zmm30[2],zmm29[4],zmm30[4],zmm29[6],zmm30[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm15, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm20, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm30[0],zmm26[0],zmm30[2],zmm26[2],zmm30[4],zmm26[4],zmm30[6],zmm26[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm17, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm16, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm29, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm24[0],zmm23[0],zmm24[2],zmm23[2],zmm24[4],zmm23[4],zmm24[6],zmm23[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm13, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm14, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm23[0],zmm18[0],zmm23[2],zmm18[2],zmm23[4],zmm18[4],zmm23[6],zmm18[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm31, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm2 {%k1} # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # zmm2 {%k1} = zmm11[0],mem[0],zmm11[2],mem[2],zmm11[4],mem[4],zmm11[6],mem[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm27, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm0, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm2 {%k1} = zmm7[0],zmm28[0],zmm7[2],zmm28[2],zmm7[4],zmm28[4],zmm7[6],zmm28[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm5, %zmm0, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm0, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} zmm0 {%k1} = zmm9[0],zmm10[0],zmm9[2],zmm10[2],zmm9[4],zmm10[4],zmm9[6],zmm10[6]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm5 = [7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm5, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm19 {%k1} = zmm17[1],zmm2[1],zmm17[3],zmm2[3],zmm17[5],zmm2[5],zmm17[7],zmm2[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm12 = [0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm2, %zmm12, %zmm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vbroadcasti32x4 {{.*#+}} zmm2 = [1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm13[1],zmm11[1],zmm13[3],zmm11[3],zmm13[5],zmm11[5],zmm13[7],zmm11[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm12, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm11, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm12, %zmm20
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm4, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm30[1],zmm26[1],zmm30[3],zmm26[3],zmm30[5],zmm26[5],zmm30[7],zmm26[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm26
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm12, %zmm26
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm10, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm12, %zmm17
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm9, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm31
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm29, %zmm5, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm16 {%k1} = zmm24[1],zmm6[1],zmm24[3],zmm6[3],zmm24[5],zmm6[5],zmm24[7],zmm6[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm29
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm12, %zmm29
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm6, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm12, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm3, %zmm2, %zmm31
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm23[1],zmm18[1],zmm23[3],zmm18[3],zmm23[5],zmm18[5],zmm23[7],zmm18[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm24
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm12, %zmm24
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm18, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm12, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm21, %zmm2, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm30
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm1 {%k1} = zmm10[1],zmm0[1],zmm10[3],zmm0[3],zmm10[5],zmm0[5],zmm10[7],zmm0[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm18
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm30
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm23
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm5, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm21[1],zmm28[1],zmm21[3],zmm28[3],zmm21[5],zmm28[5],zmm21[7],zmm28[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, %zmm28
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm12, %zmm21
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm23
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm22
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm12, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm8, %zmm2, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermi2q %zmm25, %zmm15, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpermt2q %zmm25, %zmm5, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} zmm15 {%k1} = zmm14[1],zmm8[1],zmm14[3],zmm8[3],zmm14[5],zmm8[5],zmm14[7],zmm8[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm0 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm8 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm13 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm9 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm4 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm6 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm15 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm10 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpermt2q {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm14 # 64-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm13[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm11[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpblendd $15, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm0 = mem[0,1,2,3],ymm15[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm28, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm25, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm14[0,1,2,3],ymm10[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm17 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 192(%rdi), %ymm8, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 128(%rdi), %ymm0, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm28 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm28, %zmm17, %zmm26
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm20 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 576(%rdi), %xmm13
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 704(%rdi), %ymm13, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 512(%rdi), %xmm28
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 640(%rdi), %ymm28, %ymm28
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm4 = ymm28[0],ymm13[0],ymm28[2],ymm13[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm20, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm16 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1088(%rdi), %xmm4
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 1216(%rdi), %ymm4, %ymm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1024(%rdi), %xmm7
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 1152(%rdi), %ymm7, %ymm7
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm5 = ymm7[0],ymm4[0],ymm7[2],ymm4[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm16, %zmm24
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm19 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1600(%rdi), %xmm5
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 1728(%rdi), %ymm5, %ymm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 1536(%rdi), %xmm11
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 1664(%rdi), %ymm11, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm29 = ymm11[0],ymm5[0],ymm11[2],ymm5[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm29, %zmm19, %zmm19
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm23 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 2112(%rdi), %xmm6
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 2240(%rdi), %ymm6, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2048(%rdi), %xmm29
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 2176(%rdi), %ymm29, %ymm29
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm3 = ymm29[0],ymm6[0],ymm29[2],ymm6[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm23, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm18 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 2624(%rdi), %xmm10
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 2752(%rdi), %ymm10, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 2560(%rdi), %xmm23
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 2688(%rdi), %ymm23, %ymm23
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm20 = ymm23[0],ymm10[0],ymm23[2],ymm10[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm20, %zmm18, %zmm18
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3136(%rdi), %xmm17
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 3264(%rdi), %ymm17, %ymm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 3072(%rdi), %xmm20
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $1, 3200(%rdi), %ymm20, %ymm20
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm15 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm15, %zmm1, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm12 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3648(%rdi), %xmm9
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 3776(%rdi), %ymm9, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 3584(%rdi), %xmm14
; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, 3712(%rdi), %ymm14, %ymm14
; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} ymm16 = ymm14[0],ymm9[0],ymm14[2],ymm9[2]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm16 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm16, %zmm12, %zmm16
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm13 = ymm28[1],ymm13[1],ymm28[3],ymm13[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm13, %zmm12, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, %zmm12 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm5, %zmm31, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm27 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm4 = ymm7[1],ymm4[1],ymm7[3],ymm4[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm27, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm30 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm7 = ymm23[1],ymm10[1],ymm23[3],ymm10[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm7, %zmm30, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm22 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm6 = ymm29[1],ymm6[1],ymm29[3],ymm6[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm6, %zmm22, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpunpckhqdq {{.*#+}} ymm8 = ymm14[1],ymm9[1],ymm14[3],ymm9[3]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm8, %zmm2, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 448(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, 384(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, 320(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, 256(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, 192(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, 128(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm1, 64(%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, 256(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, 320(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, 128(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, 192(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, 64(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 448(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 256(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 320(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 384(%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 448(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 256(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 320(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 384(%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 448(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 256(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 320(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 384(%r9)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 448(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 384(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, 448(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 256(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 320(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 128(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 192(%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm0, 64(%rax)
; AVX512DQ-BW-FCP-NEXT:    addq $6664, %rsp # imm = 0x1A08
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <512 x i64>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120, i32 128, i32 136, i32 144, i32 152, i32 160, i32 168, i32 176, i32 184, i32 192, i32 200, i32 208, i32 216, i32 224, i32 232, i32 240, i32 248, i32 256, i32 264, i32 272, i32 280, i32 288, i32 296, i32 304, i32 312, i32 320, i32 328, i32 336, i32 344, i32 352, i32 360, i32 368, i32 376, i32 384, i32 392, i32 400, i32 408, i32 416, i32 424, i32 432, i32 440, i32 448, i32 456, i32 464, i32 472, i32 480, i32 488, i32 496, i32 504>
  %strided.vec1 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121, i32 129, i32 137, i32 145, i32 153, i32 161, i32 169, i32 177, i32 185, i32 193, i32 201, i32 209, i32 217, i32 225, i32 233, i32 241, i32 249, i32 257, i32 265, i32 273, i32 281, i32 289, i32 297, i32 305, i32 313, i32 321, i32 329, i32 337, i32 345, i32 353, i32 361, i32 369, i32 377, i32 385, i32 393, i32 401, i32 409, i32 417, i32 425, i32 433, i32 441, i32 449, i32 457, i32 465, i32 473, i32 481, i32 489, i32 497, i32 505>
  %strided.vec2 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122, i32 130, i32 138, i32 146, i32 154, i32 162, i32 170, i32 178, i32 186, i32 194, i32 202, i32 210, i32 218, i32 226, i32 234, i32 242, i32 250, i32 258, i32 266, i32 274, i32 282, i32 290, i32 298, i32 306, i32 314, i32 322, i32 330, i32 338, i32 346, i32 354, i32 362, i32 370, i32 378, i32 386, i32 394, i32 402, i32 410, i32 418, i32 426, i32 434, i32 442, i32 450, i32 458, i32 466, i32 474, i32 482, i32 490, i32 498, i32 506>
  %strided.vec3 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123, i32 131, i32 139, i32 147, i32 155, i32 163, i32 171, i32 179, i32 187, i32 195, i32 203, i32 211, i32 219, i32 227, i32 235, i32 243, i32 251, i32 259, i32 267, i32 275, i32 283, i32 291, i32 299, i32 307, i32 315, i32 323, i32 331, i32 339, i32 347, i32 355, i32 363, i32 371, i32 379, i32 387, i32 395, i32 403, i32 411, i32 419, i32 427, i32 435, i32 443, i32 451, i32 459, i32 467, i32 475, i32 483, i32 491, i32 499, i32 507>
  %strided.vec4 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124, i32 132, i32 140, i32 148, i32 156, i32 164, i32 172, i32 180, i32 188, i32 196, i32 204, i32 212, i32 220, i32 228, i32 236, i32 244, i32 252, i32 260, i32 268, i32 276, i32 284, i32 292, i32 300, i32 308, i32 316, i32 324, i32 332, i32 340, i32 348, i32 356, i32 364, i32 372, i32 380, i32 388, i32 396, i32 404, i32 412, i32 420, i32 428, i32 436, i32 444, i32 452, i32 460, i32 468, i32 476, i32 484, i32 492, i32 500, i32 508>
  %strided.vec5 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125, i32 133, i32 141, i32 149, i32 157, i32 165, i32 173, i32 181, i32 189, i32 197, i32 205, i32 213, i32 221, i32 229, i32 237, i32 245, i32 253, i32 261, i32 269, i32 277, i32 285, i32 293, i32 301, i32 309, i32 317, i32 325, i32 333, i32 341, i32 349, i32 357, i32 365, i32 373, i32 381, i32 389, i32 397, i32 405, i32 413, i32 421, i32 429, i32 437, i32 445, i32 453, i32 461, i32 469, i32 477, i32 485, i32 493, i32 501, i32 509>
  %strided.vec6 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126, i32 134, i32 142, i32 150, i32 158, i32 166, i32 174, i32 182, i32 190, i32 198, i32 206, i32 214, i32 222, i32 230, i32 238, i32 246, i32 254, i32 262, i32 270, i32 278, i32 286, i32 294, i32 302, i32 310, i32 318, i32 326, i32 334, i32 342, i32 350, i32 358, i32 366, i32 374, i32 382, i32 390, i32 398, i32 406, i32 414, i32 422, i32 430, i32 438, i32 446, i32 454, i32 462, i32 470, i32 478, i32 486, i32 494, i32 502, i32 510>
  %strided.vec7 = shufflevector <512 x i64> %wide.vec, <512 x i64> poison, <64 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127, i32 135, i32 143, i32 151, i32 159, i32 167, i32 175, i32 183, i32 191, i32 199, i32 207, i32 215, i32 223, i32 231, i32 239, i32 247, i32 255, i32 263, i32 271, i32 279, i32 287, i32 295, i32 303, i32 311, i32 319, i32 327, i32 335, i32 343, i32 351, i32 359, i32 367, i32 375, i32 383, i32 391, i32 399, i32 407, i32 415, i32 423, i32 431, i32 439, i32 447, i32 455, i32 463, i32 471, i32 479, i32 487, i32 495, i32 503, i32 511>
  store <64 x i64> %strided.vec0, ptr %out.vec0, align 64
  store <64 x i64> %strided.vec1, ptr %out.vec1, align 64
  store <64 x i64> %strided.vec2, ptr %out.vec2, align 64
  store <64 x i64> %strided.vec3, ptr %out.vec3, align 64
  store <64 x i64> %strided.vec4, ptr %out.vec4, align 64
  store <64 x i64> %strided.vec5, ptr %out.vec5, align 64
  store <64 x i64> %strided.vec6, ptr %out.vec6, align 64
  store <64 x i64> %strided.vec7, ptr %out.vec7, align 64
  ret void
}