[llvm-bugs] [Bug 41545] New: Byte swap idioms loses optimization on AVX+

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=41545

            Bug ID: 41545
           Summary: Byte swap idioms loses optimization on AVX+
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: jed at 59a2.org
                CC: craig.topper at gmail.com, llvm-bugs at lists.llvm.org,
                    llvm-dev at redking.me.uk, spatel+llvm at rotateright.com

This code correctly optimizes to a simple load on x86-64:

static unsigned read_u32_le(const unsigned char arr[]) {
  return (arr[0] << 0)
    | (arr[1] << 8)
    | (arr[2] << 16)
    | (arr[3] << 24);
}

clang -O:

read_u32_le: # @read_u32_le
  mov eax, dword ptr [rdi]
  ret

However, when allowed to inline into code such as

unsigned sum_buf(int len, const unsigned char *arr) {
  unsigned sum = 0;
  for (int i=0; i<len; i+=4) {
    sum += read_u32_le(arr+i);
  }
  return sum;
}

on AVX/AVX2, the optimization is lost. For example, with -march=haswell:

.LBB0_5: # =>This Inner Loop Header: Depth=1
  vmovdqu xmm7, xmmword ptr [rsi + 4*rax]
  vmovdqu xmm0, xmmword ptr [rsi + 4*rax + 16]
  vmovdqu xmm1, xmmword ptr [rsi + 4*rax + 32]
  vmovdqu xmm2, xmmword ptr [rsi + 4*rax + 48]
  vpblendw xmm11, xmm7, xmm8, 170 # xmm11 =
xmm7[0],xmm8[1],xmm7[2],xmm8[3],xmm7[4],xmm8[5],xmm7[6],xmm8[7]
  vpblendw xmm12, xmm0, xmm8, 170 # xmm12 =
xmm0[0],xmm8[1],xmm0[2],xmm8[3],xmm0[4],xmm8[5],xmm0[6],xmm8[7]
  vpblendw xmm13, xmm1, xmm8, 170 # xmm13 =
xmm1[0],xmm8[1],xmm1[2],xmm8[3],xmm1[4],xmm8[5],xmm1[6],xmm8[7]
  vpblendw xmm14, xmm2, xmm8, 170 # xmm14 =
xmm2[0],xmm8[1],xmm2[2],xmm8[3],xmm2[4],xmm8[5],xmm2[6],xmm8[7]
  vpand xmm3, xmm7, xmm9
  vpor xmm11, xmm11, xmm3
  vpand xmm3, xmm0, xmm9
  vpor xmm12, xmm12, xmm3
  vpand xmm3, xmm1, xmm9
  vpor xmm13, xmm13, xmm3
  vpand xmm3, xmm2, xmm9
  vpor xmm3, xmm14, xmm3
  vpand xmm7, xmm7, xmm10
  vpor xmm7, xmm11, xmm7
  vpaddd xmm15, xmm7, xmm15
  vpand xmm0, xmm0, xmm10
  vpor xmm0, xmm12, xmm0
  vpaddd xmm4, xmm0, xmm4
  vpand xmm0, xmm1, xmm10
  vpor xmm0, xmm13, xmm0
  vpaddd xmm5, xmm0, xmm5
  vpand xmm0, xmm2, xmm10
  vpor xmm0, xmm3, xmm0
  vpaddd xmm6, xmm0, xmm6
  add rax, 16
  cmp rdi, rax
  jne .LBB0_5


Meanwhile, the same inner loop at x86-64 generic:

.LBB0_5: # =>This Inner Loop Header: Depth=1
  movdqu xmm2, xmmword ptr [rsi + 4*rax]
  paddd xmm0, xmm2
  movdqu xmm2, xmmword ptr [rsi + 4*rax + 16]
  paddd xmm1, xmm2
  add rax, 8
  cmp rdi, rax
  jne .LBB0_5

https://gcc.godbolt.org/z/Mop93_

-- 
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/20190420/527b0bd3/attachment-0001.html>