[llvm-bugs] [Bug 44148] New: Missed vectorization opportunity

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=44148

            Bug ID: 44148
           Summary: Missed vectorization opportunity
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: david.bolvansky at gmail.com
                CC: llvm-bugs at lists.llvm.org

#define N 512

int a[N], b[N];

int
foo (int aval)
{
  int res = 0;
  for (int i=0; i<N; i++)
  {
    if (a[i] != 0)
      res = aval;
  }
  return res;
}

Clang -O3 -march=haswell: loop not vectorized [-Rpass-missed=loop-vectorize]

foo(int): # @foo(int)
  xor eax, eax
  mov rcx, -2048
.LBB0_1: # =>This Inner Loop Header: Depth=1
  mov edx, dword ptr [rcx + a+2052]
  or edx, dword ptr [rcx + a+2048]
  or edx, dword ptr [rcx + a+2056]
  or edx, dword ptr [rcx + a+2060]
  or edx, dword ptr [rcx + a+2064]
  or edx, dword ptr [rcx + a+2068]
  or edx, dword ptr [rcx + a+2072]
  or edx, dword ptr [rcx + a+2076]
  cmovne eax, edi
  add rcx, 32
  jne .LBB0_1
  ret
a:
  .zero 2048

b:
  .zero 2048


ICC -O3 -march=haswell:
foo(int):
        xor       eax, eax                                      #10.11
        vpxor     ymm1, ymm1, ymm1                              #13.17
        xor       edx, edx                                      #11.3
        vpcmpeqd  ymm0, ymm0, ymm0                              #10.11
..B1.2:                         # Preds ..B1.2 ..B1.1
        vpcmpeqd  ymm2, ymm1, YMMWORD PTR [a+rdx*4]             #13.17
        add       rdx, 8                                        #11.3
        vpxor     ymm3, ymm2, ymm0                              #10.11
        vmovmskps ecx, ymm3                                     #10.11
        or        eax, ecx                                      #10.11
        cmp       rdx, 512                                      #11.3
        jb        ..B1.2        # Prob 99%                      #11.3
        test      eax, eax                                      #16.10
        cmovne    eax, edi                                      #16.10
        vzeroupper                                              #16.10
        ret                                                     #16.10
a:
b:

https://godbolt.org/z/AjRQNY

This loop is similar, also not vectorized:
int
foo (int aval)
{
  int res = 0;
  for (int i=0; i<N; i++)
  {
    if (a[i] != 0)
      res = b[i];
  }
  return res;
}

https://godbolt.org/z/AU_5rf

-- 
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/20191126/656f74ae/attachment.html>