[llvm-bugs] [Bug 42305] New: [X86][AVX] Regression in subvector scatter due to splitting ymm loads

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=42305

            Bug ID: 42305
           Summary: [X86][AVX] Regression in subvector scatter due to
                    splitting ymm loads
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: llvm-dev at redking.me.uk
                CC: andrea.dibiagio at gmail.com, craig.topper at gmail.com,
                    greg.bedwell at sony.com, llvm-bugs at lists.llvm.org,
                    llvm-dev at redking.me.uk, spatel+llvm at rotateright.com

https://godbolt.org/z/8Cjr7X

#include <x86intrin.h>

void scatter_subvectors( int count, const float* src, const int* idx, float*
dst )
{
  for( int i = 0; i != count; ++i )
  {
        __m256 j01 = _mm256_loadu_ps((const float*)(src + 0));
        __m256 j23 = _mm256_loadu_ps((const float*)(src + 8));
        src += 16;

        _mm_storeu_ps(dst + idx[0],  _mm256_extractf128_ps(j01, 0));
        _mm_storeu_ps(dst + idx[1],  _mm256_extractf128_ps(j01, 1));
        _mm_storeu_ps(dst + idx[2],  _mm256_extractf128_ps(j23, 0));
        _mm_storeu_ps(dst + idx[3],  _mm256_extractf128_ps(j23, 1));
        idx += 4;
  }
}

We're seeing a regression as the ymm loads are each being split into 2 xmm
loads because we extract+store the subvectors independently. But because
vextractf128 can store directly we're increasing instruction count completely
unnecessarily.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20190618/7d523630/attachment-0001.html>