[llvm-bugs] [Bug 51786] New: Failure to vectorize max element loop when it iterates over a long range

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=51786

            Bug ID: 51786
           Summary: Failure to vectorize max element loop when it iterates
                    over a long range
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: gabravier at gmail.com
                CC: llvm-bugs at lists.llvm.org

int test(int (&arr)[76])
{
    int *first = &arr[0];
    int *last = &arr[76];

    int *largest = first++;
    for (; first != last; ++first)
        if (*largest < *first)
            largest = first;
    return *largest;
}

With -O3 -mavx2, GCC optimizes the function to this:

test(int (&) [76]):
  vmovdqu ymm2, YMMWORD PTR [rdi+36]
  vpbroadcastd ymm0, DWORD PTR [rdi]
  vpmaxsd ymm1, ymm2, YMMWORD PTR [rdi+68]
  vpmaxsd ymm0, ymm0, YMMWORD PTR [rdi+4]
  vmovdqu ymm3, YMMWORD PTR [rdi+100]
  vmovdqu ymm4, YMMWORD PTR [rdi+196]
  vpmaxsd ymm1, ymm1, ymm0
  vpmaxsd ymm0, ymm3, YMMWORD PTR [rdi+132]
  vpmaxsd ymm0, ymm0, YMMWORD PTR [rdi+164]
  vpmaxsd ymm1, ymm1, ymm0
  vpmaxsd ymm0, ymm4, YMMWORD PTR [rdi+228]
  vpmaxsd ymm0, ymm0, YMMWORD PTR [rdi+260]
  vpmaxsd ymm1, ymm1, ymm0
  vextracti128 xmm0, ymm1, 0x1
  vpmaxsd xmm0, xmm0, xmm1
  vpsrldq xmm1, xmm0, 8
  vpmaxsd xmm0, xmm0, xmm1
  vpsrldq xmm1, xmm0, 4
  vpmaxsd xmm0, xmm0, xmm1
  vmovd xmm1, DWORD PTR [rdi+292]
  vinsertps xmm0, xmm0, xmm0, 0xe
  vpmaxsd xmm0, xmm0, xmm1
  vmovd xmm1, DWORD PTR [rdi+296]
  vpmaxsd xmm0, xmm0, xmm1
  vmovd xmm1, DWORD PTR [rdi+300]
  vpmaxsd xmm0, xmm0, xmm1
  vmovd eax, xmm0
  vzeroupper
  ret

LLVM instead outputs this:

test(int (&) [76]): # @test(int (&) [76])
  lea rax, [rdi + 304]
  mov rcx, rdi
  add rcx, 4
  jmp .LBB0_1
.LBB0_11: # in Loop: Header=BB0_1 Depth=1
  mov rdi, rdx
  add rcx, 20
  cmp rcx, rax
  je .LBB0_13
.LBB0_1: # =>This Inner Loop Header: Depth=1
  mov esi, dword ptr [rdi]
  mov rdx, rcx
  cmp esi, dword ptr [rcx]
  jge .LBB0_2
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 4]
  jl .LBB0_4
.LBB0_5: # in Loop: Header=BB0_1 Depth=1
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 8]
  jl .LBB0_6
.LBB0_7: # in Loop: Header=BB0_1 Depth=1
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 12]
  jl .LBB0_8
.LBB0_9: # in Loop: Header=BB0_1 Depth=1
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 16]
  jge .LBB0_11
  jmp .LBB0_10
.LBB0_2: # in Loop: Header=BB0_1 Depth=1
  mov rdx, rdi
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 4]
  jge .LBB0_5
.LBB0_4: # in Loop: Header=BB0_1 Depth=1
  lea rdx, [rcx + 4]
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 8]
  jge .LBB0_7
.LBB0_6: # in Loop: Header=BB0_1 Depth=1
  lea rdx, [rcx + 8]
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 12]
  jge .LBB0_9
.LBB0_8: # in Loop: Header=BB0_1 Depth=1
  lea rdx, [rcx + 12]
  mov esi, dword ptr [rdx]
  cmp esi, dword ptr [rcx + 16]
  jge .LBB0_11
.LBB0_10: # in Loop: Header=BB0_1 Depth=1
  lea rdi, [rcx + 16]
  add rcx, 20
  cmp rcx, rax
  jne .LBB0_1
.LBB0_13:
  mov eax, dword ptr [rdi]
  ret

For anything smaller, LLVM seems capable of doing this optimization, but as
soon as it gets to 76, it isn't vectorized anymore

Godbolt link: https://godbolt.org/z/zbKa8fjhj

-- 
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/20210907/ee22aceb/attachment.html>