[X86][AVX] Fold CONCAT(HOP(X,Y),HOP(Z,W)) -> HOP(CONCAT(X,Z),CONCAT(Y,W)) for float types
Only do this for AVX2+ targets as we still get some regressions on AVX1 without PERMPD/PERMQ
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 9ca4d33..3c1e8cb 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -48257,6 +48257,8 @@
break;
case X86ISD::HADD:
case X86ISD::HSUB:
+ case X86ISD::FHADD:
+ case X86ISD::FHSUB:
case X86ISD::PACKSS:
case X86ISD::PACKUS:
if (!IsSplat && VT.is256BitVector() && Subtarget.hasInt256()) {
diff --git a/llvm/test/CodeGen/X86/haddsub-2.ll b/llvm/test/CodeGen/X86/haddsub-2.ll
index f28ed59..ff5ce09 100644
--- a/llvm/test/CodeGen/X86/haddsub-2.ll
+++ b/llvm/test/CodeGen/X86/haddsub-2.ll
@@ -444,14 +444,21 @@
; SSE-NEXT: movapd %xmm2, %xmm1
; SSE-NEXT: retq
;
-; AVX-LABEL: avx_vhadd_pd_test:
-; AVX: # %bb.0:
-; AVX-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX-NEXT: vhaddpd %xmm2, %xmm1, %xmm1
-; AVX-NEXT: vextractf128 $1, %ymm0, %xmm2
-; AVX-NEXT: vhaddpd %xmm2, %xmm0, %xmm0
-; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX-NEXT: retq
+; AVX1-LABEL: avx_vhadd_pd_test:
+; AVX1: # %bb.0:
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT: vhaddpd %xmm2, %xmm1, %xmm1
+; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT: vhaddpd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: retq
+;
+; AVX2-LABEL: avx_vhadd_pd_test:
+; AVX2: # %bb.0:
+; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
+; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT: vhaddpd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT: retq
%vecext = extractelement <4 x double> %A, i32 0
%vecext1 = extractelement <4 x double> %A, i32 1
%add = fadd double %vecext, %vecext1
@@ -479,14 +486,21 @@
; SSE-NEXT: movapd %xmm2, %xmm1
; SSE-NEXT: retq
;
-; AVX-LABEL: avx_vhsub_pd_test:
-; AVX: # %bb.0:
-; AVX-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX-NEXT: vhsubpd %xmm2, %xmm1, %xmm1
-; AVX-NEXT: vextractf128 $1, %ymm0, %xmm2
-; AVX-NEXT: vhsubpd %xmm2, %xmm0, %xmm0
-; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX-NEXT: retq
+; AVX1-LABEL: avx_vhsub_pd_test:
+; AVX1: # %bb.0:
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT: vhsubpd %xmm2, %xmm1, %xmm1
+; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT: vhsubpd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: retq
+;
+; AVX2-LABEL: avx_vhsub_pd_test:
+; AVX2: # %bb.0:
+; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
+; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT: vhsubpd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT: retq
%vecext = extractelement <4 x double> %A, i32 0
%vecext1 = extractelement <4 x double> %A, i32 1
%sub = fsub double %vecext, %vecext1
diff --git a/llvm/test/CodeGen/X86/haddsub-undef.ll b/llvm/test/CodeGen/X86/haddsub-undef.ll
index f8379e6..7d797b7 100644
--- a/llvm/test/CodeGen/X86/haddsub-undef.ll
+++ b/llvm/test/CodeGen/X86/haddsub-undef.ll
@@ -1146,14 +1146,22 @@
; AVX-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX-SLOW-NEXT: retq
;
-; AVX-FAST-LABEL: PR34724_add_v4f64_u123:
-; AVX-FAST: # %bb.0:
-; AVX-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX-FAST-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX-FAST-NEXT: vhaddpd %xmm2, %xmm1, %xmm1
-; AVX-FAST-NEXT: vhaddpd %xmm0, %xmm0, %xmm0
-; AVX-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX-FAST-NEXT: retq
+; AVX1-FAST-LABEL: PR34724_add_v4f64_u123:
+; AVX1-FAST: # %bb.0:
+; AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm2
+; AVX1-FAST-NEXT: vhaddpd %xmm2, %xmm1, %xmm1
+; AVX1-FAST-NEXT: vhaddpd %xmm0, %xmm0, %xmm0
+; AVX1-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-FAST-NEXT: retq
+;
+; AVX512-FAST-LABEL: PR34724_add_v4f64_u123:
+; AVX512-FAST: # %bb.0:
+; AVX512-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; AVX512-FAST-NEXT: vblendpd {{.*#+}} ymm2 = ymm0[0,1],ymm1[2,3]
+; AVX512-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX512-FAST-NEXT: vhaddpd %ymm2, %ymm0, %ymm0
+; AVX512-FAST-NEXT: retq
%3 = shufflevector <4 x double> %0, <4 x double> %1, <2 x i32> <i32 2, i32 4>
%4 = shufflevector <4 x double> %0, <4 x double> %1, <2 x i32> <i32 3, i32 5>
%5 = fadd <2 x double> %3, %4
@@ -1196,13 +1204,19 @@
; AVX-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX-SLOW-NEXT: retq
;
-; AVX-FAST-LABEL: PR34724_add_v4f64_0u23:
-; AVX-FAST: # %bb.0:
-; AVX-FAST-NEXT: vhaddpd %xmm1, %xmm0, %xmm0
-; AVX-FAST-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX-FAST-NEXT: vhaddpd %xmm2, %xmm1, %xmm1
-; AVX-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX-FAST-NEXT: retq
+; AVX1-FAST-LABEL: PR34724_add_v4f64_0u23:
+; AVX1-FAST: # %bb.0:
+; AVX1-FAST-NEXT: vhaddpd %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm2
+; AVX1-FAST-NEXT: vhaddpd %xmm2, %xmm1, %xmm1
+; AVX1-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-FAST-NEXT: retq
+;
+; AVX512-FAST-LABEL: PR34724_add_v4f64_0u23:
+; AVX512-FAST: # %bb.0:
+; AVX512-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX512-FAST-NEXT: vhaddpd %ymm1, %ymm0, %ymm0
+; AVX512-FAST-NEXT: retq
%3 = shufflevector <4 x double> %0, <4 x double> %1, <2 x i32> <i32 0, i32 4>
%4 = shufflevector <4 x double> %0, <4 x double> %1, <2 x i32> <i32 1, i32 5>
%5 = fadd <2 x double> %3, %4
@@ -1245,14 +1259,21 @@
; AVX-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX-SLOW-NEXT: retq
;
-; AVX-FAST-LABEL: PR34724_add_v4f64_01u3:
-; AVX-FAST: # %bb.0:
-; AVX-FAST-NEXT: vextractf128 $1, %ymm0, %xmm2
-; AVX-FAST-NEXT: vhaddpd %xmm2, %xmm0, %xmm0
-; AVX-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX-FAST-NEXT: vhaddpd %xmm1, %xmm1, %xmm1
-; AVX-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX-FAST-NEXT: retq
+; AVX1-FAST-LABEL: PR34724_add_v4f64_01u3:
+; AVX1-FAST: # %bb.0:
+; AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm2
+; AVX1-FAST-NEXT: vhaddpd %xmm2, %xmm0, %xmm0
+; AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
+; AVX1-FAST-NEXT: vhaddpd %xmm1, %xmm1, %xmm1
+; AVX1-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-FAST-NEXT: retq
+;
+; AVX512-FAST-LABEL: PR34724_add_v4f64_01u3:
+; AVX512-FAST: # %bb.0:
+; AVX512-FAST-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
+; AVX512-FAST-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
+; AVX512-FAST-NEXT: vhaddpd %ymm2, %ymm0, %ymm0
+; AVX512-FAST-NEXT: retq
%3 = shufflevector <4 x double> %0, <4 x double> undef, <2 x i32> <i32 0, i32 2>
%4 = shufflevector <4 x double> %0, <4 x double> undef, <2 x i32> <i32 1, i32 3>
%5 = fadd <2 x double> %3, %4
@@ -1292,13 +1313,21 @@
; AVX-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX-SLOW-NEXT: retq
;
-; AVX-FAST-LABEL: PR34724_add_v4f64_012u:
-; AVX-FAST: # %bb.0:
-; AVX-FAST-NEXT: vextractf128 $1, %ymm0, %xmm2
-; AVX-FAST-NEXT: vhaddpd %xmm2, %xmm0, %xmm0
-; AVX-FAST-NEXT: vhaddpd %xmm1, %xmm1, %xmm1
-; AVX-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX-FAST-NEXT: retq
+; AVX1-FAST-LABEL: PR34724_add_v4f64_012u:
+; AVX1-FAST: # %bb.0:
+; AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm2
+; AVX1-FAST-NEXT: vhaddpd %xmm2, %xmm0, %xmm0
+; AVX1-FAST-NEXT: vhaddpd %xmm1, %xmm1, %xmm1
+; AVX1-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-FAST-NEXT: retq
+;
+; AVX512-FAST-LABEL: PR34724_add_v4f64_012u:
+; AVX512-FAST: # %bb.0:
+; AVX512-FAST-NEXT: vextractf128 $1, %ymm0, %xmm2
+; AVX512-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX512-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
+; AVX512-FAST-NEXT: vhaddpd %ymm1, %ymm0, %ymm0
+; AVX512-FAST-NEXT: retq
%3 = shufflevector <4 x double> %0, <4 x double> undef, <2 x i32> <i32 0, i32 2>
%4 = shufflevector <4 x double> %0, <4 x double> undef, <2 x i32> <i32 1, i32 3>
%5 = fadd <2 x double> %3, %4