GCC 4.9 regression

[gregory38]

Remainder until I manage to open a GCC issue (account creation is blocked due to spam).

Here a snippet of the bug:

The bug is a 4.9/5/6/7 regression on AVX2 handling.

// g++ -O2 -c ./intrin-avx2.c -std=c++11 -mavx2 -save-temps
#include <immintrin.h>

__m256i m;

__m128i extract(__m128i minmax)
{
    int lod = _mm_cvtsi128_si32(_mm256_castsi256_si128(m));
    return _mm_srli_epi16(minmax,lod);
}

// It will compile (GCC 4.9) to those 2 instructions (not intel syntax, destination is on the right).
vmovdqa m(%rip), %xmm1
// Bug is here %xmm1 ough to contains only 32 bits of data. [64:32] bits ought to be 0.
vpsrlw  %xmm1, %xmm0, %xmm0

[gregory38]

Note PR #1876 added a workaround to avoid the issue.

[turtleli]

Turtleli mentioned there 6.3 looks fine.

I mentioned the workaround worked fine.

[gregory38]

Bug confirmed :smile: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80286