[llvm-bugs] [Bug 46968] New: Missed SLP vectorization with umin/umax

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=46968

            Bug ID: 46968
           Summary: Missed SLP vectorization with umin/umax
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Common Code Generator Code
          Assignee: unassignedbugs at nondot.org
          Reporter: david.bolvansky at gmail.com
                CC: llvm-bugs at lists.llvm.org

Hot code - LightPixel - from Firefox (rasterflood-svg benchmark could be
faster):


#include<stdint.h>
#include<stddef.h>

const ptrdiff_t B8G8R8A8_COMPONENT_BYTEOFFSET_B = 0;
const ptrdiff_t B8G8R8A8_COMPONENT_BYTEOFFSET_G = 1;
const ptrdiff_t B8G8R8A8_COMPONENT_BYTEOFFSET_R = 2;
const ptrdiff_t B8G8R8A8_COMPONENT_BYTEOFFSET_A = 3;

static const int sInputIntPrecisionBits = 15;
static const int sOutputIntPrecisionBits = 15;
static const int sCacheIndexPrecisionBits = 7;



static inline unsigned umax(unsigned a, unsigned b) {
  return a > b ? a : b;
}

static inline unsigned umin(unsigned a, unsigned b) {
  return a > b ? b : a;
}

void foo(uint8_t components[4], uint32_t specularNHi, uint32_t aColor) {

  components[B8G8R8A8_COMPONENT_BYTEOFFSET_B] =
      umin((specularNHi * components[B8G8R8A8_COMPONENT_BYTEOFFSET_B]) >>
               sOutputIntPrecisionBits,
           255U);
  components[B8G8R8A8_COMPONENT_BYTEOFFSET_G] =
      umin((specularNHi *components[B8G8R8A8_COMPONENT_BYTEOFFSET_G]) >>
               sOutputIntPrecisionBits,
           255U);
  components[B8G8R8A8_COMPONENT_BYTEOFFSET_R] =
      umin((specularNHi * components[B8G8R8A8_COMPONENT_BYTEOFFSET_R]) >>
               sOutputIntPrecisionBits,
           255U);

  components[B8G8R8A8_COMPONENT_BYTEOFFSET_A] =
      umax(components[B8G8R8A8_COMPONENT_BYTEOFFSET_B],
           umax(components[B8G8R8A8_COMPONENT_BYTEOFFSET_G],
                components[B8G8R8A8_COMPONENT_BYTEOFFSET_R]));
}

-O3 -mavx2 - We got: List vectorization was possible but not beneficial with
cost 0 >= 0 


foo(unsigned char*, unsigned int, unsigned int):                             #
@foo(unsigned char*, unsigned int, unsigned int)
        movzx   eax, byte ptr [rdi]
        imul    eax, esi
        shr     eax, 15
        cmp     eax, 255
        mov     r8d, 255
        cmovae  eax, r8d
        mov     byte ptr [rdi], al
        movzx   edx, byte ptr [rdi + 1]
        imul    edx, esi
        shr     edx, 15
        cmp     edx, 255
        cmovae  edx, r8d
        mov     byte ptr [rdi + 1], dl
        movzx   ecx, byte ptr [rdi + 2]
        imul    ecx, esi
        shr     ecx, 15
        cmp     ecx, 255
        cmovae  ecx, r8d
        mov     byte ptr [rdi + 2], cl
        cmp     edx, ecx
        cmova   ecx, edx
        cmp     eax, ecx
        cmova   ecx, eax
        mov     byte ptr [rdi + 3], cl
        ret


-O3 -mavx512f - Clang partially vectorizes it:
.LCPI0_0:
        .long   255                             # 0xff
foo(unsigned char*, unsigned int, unsigned int):                             #
@foo(unsigned char*, unsigned int, unsigned int)
        movzx   eax, byte ptr [rdi]
        imul    eax, esi
        shr     eax, 15
        cmp     eax, 255
        mov     ecx, 255
        cmovb   ecx, eax
        mov     byte ptr [rdi], cl
        movzx   eax, word ptr [rdi + 1]
        vmovd   xmm0, eax
        vpmovzxbd       xmm0, xmm0              # xmm0 =
xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero
        vmovd   xmm1, esi
        vpbroadcastd    xmm1, xmm1
        vpmulld xmm0, xmm1, xmm0
        vpbroadcastd    xmm1, dword ptr [rip + .LCPI0_0] # xmm1 =
[255,255,255,255]
        vpsrld  xmm0, xmm0, 15
        vpminud xmm0, xmm0, xmm1
        vmovd   eax, xmm0
        mov     byte ptr [rdi + 1], al
        vpextrd edx, xmm0, 1
        mov     byte ptr [rdi + 2], dl
        cmp     eax, edx
        cmova   edx, eax
        cmp     ecx, edx
        cmova   edx, ecx
        mov     byte ptr [rdi + 3], dl
        ret


Godbolt:  -Rpass-missed=vec*

-- 
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/20200803/26f75423/attachment.html>