https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106704
Bug ID: 106704
Summary: avx intrinsic no generating vblendvps instruction with
-mavx
Product: gcc
Version: 12.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: mindmark at gmail dot com
Target Milestone: ---
I'm getting some really strange code generation on x86_64 when using the
_mm256_blendv_ps intrinsic in gcc 12 with just the -mavx flag.
https://godbolt.org/z/YbEK4nvEd
It should generate a single vblendvps instruction, but instead creates a lot of
branchy scaler code.
gcc 11 doesn't appear to exhibit this behavior.
Also using -mavx2 on gcc 12 generates the single instruction, just not -mavx
for whatever reason.
#include <immintrin.h>
__m256 bend_stuff( __m256 a, __m256 b, __m256 mask)
{
return _mm256_blendv_ps(a, b, mask);
}
gcc -O2 -mavx -c blend.c -masm=intel -S -o out.s
in GCC 12 it generates:
bend_stuff:
vmovd eax, xmm2
vmovaps ymm3, ymm0
vmovdqa xmm4, xmm2
test eax, eax
jns .L3
vmovaps xmm0, xmm1
.L3:
vpextrd eax, xmm4, 1
vshufps xmm9, xmm3, xmm3, 85
test eax, eax
jns .L5
vshufps xmm9, xmm1, xmm1, 85
.L5:
vpextrd eax, xmm4, 2
vunpckhps xmm5, xmm3, xmm3
test eax, eax
jns .L7
vunpckhps xmm5, xmm1, xmm1
.L7:
vpextrd eax, xmm4, 3
vshufps xmm7, xmm3, xmm3, 255
test eax, eax
jns .L9
vshufps xmm7, xmm1, xmm1, 255
.L9:
vextractf128 xmm2, ymm2, 0x1
vextractf128 xmm4, ymm3, 0x1
vmovd eax, xmm2
test eax, eax
jns .L11
vextractf128 xmm4, ymm1, 0x1
.L11:
vextractf128 xmm8, ymm3, 0x1
vpextrd eax, xmm2, 1
vshufps xmm8, xmm8, xmm8, 85
test eax, eax
jns .L13
vextractf128 xmm8, ymm1, 0x1
vshufps xmm8, xmm8, xmm8, 85
.L13:
vextractf128 xmm6, ymm3, 0x1
vpextrd eax, xmm2, 2
vunpckhps xmm6, xmm6, xmm6
test eax, eax
jns .L15
vextractf128 xmm6, ymm1, 0x1
vunpckhps xmm6, xmm6, xmm6
.L15:
vextractf128 xmm3, ymm3, 0x1
vpextrd eax, xmm2, 3
vshufps xmm3, xmm3, xmm3, 255
test eax, eax
jns .L17
vextractf128 xmm1, ymm1, 0x1
vshufps xmm3, xmm1, xmm1, 255
.L17:
vunpcklps xmm6, xmm6, xmm3
vunpcklps xmm4, xmm4, xmm8
vunpcklps xmm5, xmm5, xmm7
vunpcklps xmm0, xmm0, xmm9
vmovlhps xmm4, xmm4, xmm6
vmovlhps xmm0, xmm0, xmm5
vinsertf128 ymm0, ymm0, xmm4, 0x1
ret
in GCC 11.3 it generates:
bend_stuff:
vblendvps ymm0, ymm0, ymm1, ymm2
ret
