x86 codegen, AVX2: missing optimization -- the compiler could swap args of commutative intrinsic `_mm256_testz_si256`

[AlexGuteniev]

The below function has better codegen when MASK_FIRST is defined:

#include <immintrin.h>

bool f2(const char* p)
{
    for (const char* e = p + 256; p != e; p+=32)
    {
        const __m256i m = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(p));
        if (_mm256_testz_si256(
#ifdef MASK_FIRST
                _mm256_set1_epi8(~0x40), m
#else            
                m, _mm256_set1_epi8(~0x40)
#endif
                ))
        {
            return true;
        }
    }
    
return false;
}

See https://godbolt.org/z/qjqdo5x69

I want the compiler to swap arguments automatically for commutative two-args operations.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Alex Guteniev (AlexGuteniev)

The below function has better codegen when `MASK_FIRST` is defined: ```C++ #include <immintrin.h>

bool f2(const char* p)
{
for (const char* e = p + 256; p != e; p+=32)
{
const __m256i m = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(p));
if (_mm256_testz_si256(
#ifdef MASK_FIRST
_mm256_set1_epi8(~0x40), m
#else
m, _mm256_set1_epi8(~0x40)
#endif
))
{
return true;
}
}

return false;
}

See https://godbolt.org/z/qjqdo5x69

I want the compiler to swap arguments automatically for commutative two-args operations.

</details>

[topperc]

I think the tricky bit is that it isn't commutable if the C flag is used.

[AlexGuteniev]

I think the tricky bit is that it isn't commutable if the C flag is used.

Yes, but I think the only reliable way to use the C flag is to have _mm256_testc_si256 nearby, so that they share the instruction.

[topperc]

I think the tricky bit is that it isn't commutable if the C flag is used.

Yes, but I think the only reliable way to use the C flag is to have _mm256_testc_si256 nearby, so that they share the instruction.

I meant it as a compiler implementation problem. We lose knowledge of whether _mm256_testc_si256 or _mm256_testz_si256 was used before we get to the place where commuting decisions happen. So we have to scan the users to check what flags they use.