[llvm-bugs] [Bug 50555] New: Suboptimal SLP vectorization

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=50555

            Bug ID: 50555
           Summary: Suboptimal SLP vectorization
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: david.bolvansky at gmail.com
                CC: llvm-bugs at lists.llvm.org

#define N 16

void f (unsigned short *restrict a, unsigned char *restrict b,
   unsigned char *restrict c)
{
  for (int i = 0; i < N; ++i)
    {
      unsigned short res = b[i] + c[i];
      a[i] = (res + (res >> 1)) >> 2;
    }
}



https://godbolt.org/z/1x8hPbo8d

LLVM -O3 -mavx:
f:                                      # @f
        vpmovzxbd       xmm0, dword ptr [rsi + 4] # xmm0 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpmovzxbd       xmm1, dword ptr [rsi]   # xmm1 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpmovzxbd       xmm2, dword ptr [rdx + 4] # xmm2 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpmovzxbd       xmm3, dword ptr [rdx]   # xmm3 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpaddd  xmm0, xmm2, xmm0
        vpaddd  xmm1, xmm3, xmm1
        vpsrld  xmm2, xmm0, 1
        vpsrld  xmm3, xmm1, 1
        vpaddd  xmm1, xmm3, xmm1
        vpaddd  xmm0, xmm2, xmm0
        vpsrld  xmm0, xmm0, 2
        vpsrld  xmm1, xmm1, 2
        vpackusdw       xmm0, xmm1, xmm0
        vmovdqu xmmword ptr [rdi], xmm0
        vpmovzxbd       xmm0, dword ptr [rsi + 12] # xmm0 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpmovzxbd       xmm1, dword ptr [rsi + 8] # xmm1 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpmovzxbd       xmm2, dword ptr [rdx + 12] # xmm2 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpaddd  xmm0, xmm2, xmm0
        vpmovzxbd       xmm2, dword ptr [rdx + 8] # xmm2 =
mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
        vpaddd  xmm1, xmm2, xmm1
        vpsrld  xmm2, xmm0, 1
        vpsrld  xmm3, xmm1, 1
        vpaddd  xmm1, xmm3, xmm1
        vpaddd  xmm0, xmm2, xmm0
        vpsrld  xmm0, xmm0, 2
        vpsrld  xmm1, xmm1, 2
        vpackusdw       xmm0, xmm1, xmm0
        vmovdqu xmmword ptr [rdi + 16], xmm0
        ret


GCC -O3 -mavx:
f:
        vmovdqu xmm1, XMMWORD PTR [rsi]
        vmovdqu xmm0, XMMWORD PTR [rdx]
        vpmovzxbw       xmm3, xmm1
        vpsrldq xmm1, xmm1, 8
        vpmovzxbw       xmm2, xmm0
        vpsrldq xmm0, xmm0, 8
        vpaddw  xmm2, xmm2, xmm3
        vpmovzxbw       xmm1, xmm1
        vpmovzxbw       xmm0, xmm0
        vpaddw  xmm0, xmm0, xmm1
        vpsrlw  xmm1, xmm2, 1
        vpaddw  xmm1, xmm1, xmm2
        vpsrlw  xmm1, xmm1, 2
        vmovdqu XMMWORD PTR [rdi], xmm1
        vpsrlw  xmm1, xmm0, 1
        vpaddw  xmm0, xmm1, xmm0
        vpsrlw  xmm0, xmm0, 2
        vmovdqu XMMWORD PTR [rdi+16], xmm0
        ret


Plain -O3 codegen is suboptimal too. AVX512 codegen looks good.

-- 
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/20210602/a41efe11/attachment-0001.html>