[X86] Bugfix for rL349334 adding a check function for two intrinsics
_mm_mmask_i64gather_epi32 and _mm_mmask_i64gather_ps operates the
lower 64 bits and zeroes the higher 64 bits of the return value.
The old test suite steps 64 bits in the do_intrin_loop, which will
overlap higher 64 bits of other data address in the last iteration.
Especially when compiler malloc dst128_f and mask128 's address
adjacent, this test will run fail. A new check function is added
to check these two intrinsics.
Differential Revision: https://reviews.llvm.org/D79158
diff --git a/SingleSource/UnitTests/Vector/AVX512VL/i64gather_32.c b/SingleSource/UnitTests/Vector/AVX512VL/i64gather_32.c
index b510ab6..49941bf 100644
--- a/SingleSource/UnitTests/Vector/AVX512VL/i64gather_32.c
+++ b/SingleSource/UnitTests/Vector/AVX512VL/i64gather_32.c
@@ -46,22 +46,22 @@
void do_mm_mmask_i64gather_epi32() {
int i;
- for (i = 0; i < NUM; i += 2) {
+ for (i = 0; i < NUM; i += 4) {
__m128i ind = _mm_loadu_si128((const __m128i *)(g_index + i));
__m128i old_dst = _mm_loadu_si128((__m128i const *)(dst_i + i));
__m128i gtr =
- _mm_mmask_i64gather_epi32(old_dst, mask128[i / 2], ind, src_i, SCALE);
+ _mm_mmask_i64gather_epi32(old_dst, mask128[i / 4], ind, src_i, SCALE);
_mm_storeu_si128((__m128i *)(dst128_i + i), gtr);
}
}
void do_mm_mmask_i64gather_ps() {
int i;
- for (i = 0; i < NUM; i += 2) {
+ for (i = 0; i < NUM; i += 4) {
__m128i ind = _mm_loadu_si128((const __m128i *)(g_index + i));
__m128 old_dst = _mm_loadu_ps(dst_f + i);
__m128 gtr =
- _mm_mmask_i64gather_ps(old_dst, mask128[i / 2], ind, src_f, SCALE);
+ _mm_mmask_i64gather_ps(old_dst, mask128[i / 4], ind, src_f, SCALE);
_mm_storeu_ps(dst128_f + i, gtr);
}
}
@@ -110,16 +110,42 @@
return 0;
}
+int checkh(int id, int *res_dst, int *pass_thru_vals, int *mask, int *src,
+ int elems_in_vector) {
+ int i;
+ for (i = 0; i < NUM; i++) {
+ int kmask = mask[i / elems_in_vector];
+ int kmask_bit = kmask & (1 << (i % elems_in_vector));
+
+ int v;
+ if (i % elems_in_vector >= elems_in_vector / 2)
+ v = 0;
+ else
+ v = kmask_bit ? src[g_index[i]] : pass_thru_vals[i];
+ // printf("v= %d, g_index[i] = %ld, src[g_index[i]]= %d, res_dst[i]=%d\n ",
+ // v, g_index[i], src[g_index[i]], res_dst[i]);
+
+ if (v != res_dst[i]) {
+ printf("The testcase #%d FAILed at %d iteration\n", id, i);
+
+ printf("Expected value %d, actual %d\n", v, res_dst[i]);
+
+ return -1;
+ }
+ }
+ return 0;
+}
+
int main() {
int error = 0;
init_data();
do_mm_mmask_i64gather_epi32();
- error |= check(1, dst128_i, dst_i, mask128, src_i, 2);
+ error |= checkh(1, dst128_i, dst_i, mask128, src_i, 4);
do_mm_mmask_i64gather_ps();
- error |= check(2, (int *)dst128_f, (int *)dst_f, mask128, (int *)src_f, 2);
+ error |= checkh(2, (int *)dst128_f, (int *)dst_f, mask128, (int *)src_f, 4);
do_mm256_mmask_i64gather_epi32();
error |= check(3, dst256_i, dst_i, mask256, src_i, 4);
