[llvm-bugs] [Bug 42674] [X86][SSE] Missed vXi8 sum reduction optimization

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=42674

Joel Yliluoma <bisqwit at iki.fi> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|FIXED                       |---
             Status|RESOLVED                    |REOPENED

--- Comment #10 from Joel Yliluoma <bisqwit at iki.fi> ---
I am sorry about reopening this once again, but Clang still generates
suboptimal code for vector size 8 and totally forfeits SIMD for 128:

On 8 bytes, Clang trunk:

        vpmovzxbw xmm0, qword ptr [rip + bytes]
        vpshufd xmm1, xmm0, 78          # xmm1 = xmm0[2,3,0,1]
        vpaddw  xmm0, xmm0, xmm1
        vpshufd xmm1, xmm0, 229         # xmm1 = xmm0[1,1,2,3]
        vpaddw  xmm0, xmm0, xmm1
        vpsrld  xmm1, xmm0, 16
        vpaddw  xmm0, xmm0, xmm1
        vpextrb eax, xmm0, 0
        ret

On 128 bytes, Clang trunk:

        mov     rcx, -128
        xor     eax, eax
    .LBB0_1:
        add     al, byte ptr [rcx + bytes+128]
        add     al, byte ptr [rcx + bytes+129]
        add     al, byte ptr [rcx + bytes+130]
        add     al, byte ptr [rcx + bytes+131]
        add     al, byte ptr [rcx + bytes+132]
        add     al, byte ptr [rcx + bytes+133]
        add     al, byte ptr [rcx + bytes+134]
        add     al, byte ptr [rcx + bytes+135]
        add     rcx, 8
        jne     .LBB0_1
        ret

Test code:

        #define SIZE 128
        unsigned char bytes[SIZE];
        unsigned char sum (void)
        {
          unsigned char r = 0;
          const unsigned char *p = (const unsigned char *) bytes;
          int n;
          for (n = 0; n < sizeof (bytes); ++n)
            r += p[n];
          return r;
        }

For reference, here’s GCC (trunk) for 8:

        vmovq   xmm0, QWORD PTR bytes[rip]
        vpxor   xmm1, xmm1, xmm1
        vpsadbw xmm2, xmm0, xmm1
        vmovq   rax, xmm2
        ret

and GCC (trunk) for 128 (AVX-512BW):

        vmovdqu8      zmm2, ZMMWORD PTR bytes[rip]
        vpaddb  zmm1, zmm2, ZMMWORD PTR bytes[rip+64]
        vextracti64x4 ymm0, zmm1, 0x1
        vpaddb  ymm1, ymm0, ymm1
        vextracti128  xmm0, ymm1, 0x1
        vpaddb  xmm0, xmm0, xmm1
        vpsrldq xmm1, xmm0, 8
        vpaddb  xmm0, xmm0, xmm1
        vpxor   xmm1, xmm1, xmm1
        vpsadbw xmm0, xmm0, xmm1
        vpextrb eax, xmm0, 0
        vzeroupper
        ret

or GCC (trunk) for 128 (non-AVX):

        movdqa  xmm0, XMMWORD PTR bytes[rip]
        paddb   xmm0, XMMWORD PTR bytes[rip+16]
        paddb   xmm0, XMMWORD PTR bytes[rip+48]
        paddb   xmm0, XMMWORD PTR bytes[rip+32]
        paddb   xmm0, XMMWORD PTR bytes[rip+80]
        paddb   xmm0, XMMWORD PTR bytes[rip+64]
        paddb   xmm0, XMMWORD PTR bytes[rip+96]
        paddb   xmm0, XMMWORD PTR bytes[rip+112]
        movdqa  xmm1, xmm0
        psrldq  xmm1, 8
        paddb   xmm0, xmm1
        pxor    xmm1, xmm1
        movdqa  xmm2, xmm0
        psadbw  xmm2, xmm1
        movaps  XMMWORD PTR [rsp-40], xmm2
        movzx   eax, BYTE PTR [rsp-40]
        ret

On 64 and 256 Clang does fine now. On 96 it does sort of fine, but GCC does
better.

Clang (trunk) on 96, with AVX2 and no AVX512BW:

        vmovdqu ymm0, ymmword ptr [rip + bytes+32]
        vpaddb  ymm0, ymm0, ymmword ptr [rip + bytes+64]
        vextracti128    xmm1, ymm0, 1
        vpaddb  xmm0, xmm0, xmm1
        vpshufd xmm1, xmm0, 78          # xmm1 = xmm0[2,3,0,1]
        vpaddb  xmm0, xmm0, xmm1
        vpxor   xmm1, xmm1, xmm1
        vpsadbw xmm0, xmm0, xmm1
        vpextrb ecx, xmm0, 0
        vmovdqa xmm0, xmmword ptr [rip + bytes]
        vpaddb  xmm0, xmm0, xmmword ptr [rip + bytes+16]
        vpshufd xmm2, xmm0, 78          # xmm2 = xmm0[2,3,0,1]
        vpaddb  xmm0, xmm0, xmm2
        vpsadbw xmm0, xmm0, xmm1
        vpextrb eax, xmm0, 0
        add     al, cl
        vzeroupper
        ret

GCC (trunk) on 96, with AVX2 and no AVX512BW:

        vmovdqu ymm2, YMMWORD PTR bytes[rip]
        vpaddb  ymm0, ymm2, YMMWORD PTR bytes[rip+32]
        vpaddb  ymm1, ymm0, YMMWORD PTR bytes[rip+64]
        vextracti128    xmm0, ymm1, 0x1
        vpaddb  xmm0, xmm0, xmm1
        vpsrldq xmm1, xmm0, 8
        vpaddb  xmm0, xmm0, xmm1
        vpxor   xmm1, xmm1, xmm1
        vpsadbw xmm0, xmm0, xmm1
        vpextrb eax, xmm0, 0
        vzeroupper
        ret

-- 
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/20190801/3e8bca9f/attachment-0001.html>