[SLP] Missing horizontal reduction when stored to dereferenced pointer

[davidbolvansky]

Bugzilla Link	51978
Version	trunk
OS	Linux
CC	@alexey-bataev,@anton-afanasyev,@fhahn,@kamleshbhalui,@RKSimon,@rotateright

Extended Description

void foo_notvectorized(unsigned char *p, unsigned char *t) {
  char sum = 0;
  for (int i = 0; i != 4; i++)
    sum += p[i];
  *t = sum;
}

char foo_vectorized(char *p) {
  char sum = 0;
  for (int i = 0; i != 4; i++)
    sum += p[i];
  return sum;
}

foo_notvectorized:
        mov     al, byte ptr [rdi + 1]
        add     al, byte ptr [rdi]
        add     al, byte ptr [rdi + 2]
        add     al, byte ptr [rdi + 3]
        mov     byte ptr [rsi], al
        ret
foo_vectorized:
        movd    xmm0, dword ptr [rdi]           # xmm0 = mem[0],zero,zero,zero
        pxor    xmm1, xmm1
        psadbw  xmm1, xmm0
        movd    eax, xmm1
        ret

[davidbolvansky]

Maybe a better examples.. N = 8

void foo8_notvectorized(unsigned char *p, unsigned char *t) {
  char sum = 0;
  for (int i = 0; i != 8; i++)
    sum += p[i];
  *t = sum;
}

char foo8_vectorized(char *p) {
  char sum = 0;
  for (int i = 0; i != 8; i++)
    sum += p[i];
  return sum;
}

GCC produces:

foo8_notvectorized:
        movq    xmm1, QWORD PTR [rdi]
        pxor    xmm0, xmm0
        movdqa  xmm2, xmm1
        psadbw  xmm2, xmm0
        movq    rax, xmm2
        mov     BYTE PTR [rsi], al
        ret
foo8_vectorized:
        movq    xmm1, QWORD PTR [rdi]
        pxor    xmm0, xmm0
        movdqa  xmm2, xmm1
        psadbw  xmm2, xmm0
        movq    rax, xmm2
        ret

LLVM:

foo8_notvectorized:
        mov     al, byte ptr [rdi + 1]
        add     al, byte ptr [rdi]
        add     al, byte ptr [rdi + 2]
        add     al, byte ptr [rdi + 3]
        add     al, byte ptr [rdi + 4]
        add     al, byte ptr [rdi + 5]
        add     al, byte ptr [rdi + 6]
        add     al, byte ptr [rdi + 7]
        mov     byte ptr [rsi], al
        ret

foo8_vectorized:
        movq    xmm0, qword ptr [rdi]           # xmm0 = mem[0],zero
        pxor    xmm1, xmm1
        psadbw  xmm1, xmm0
        movd    eax, xmm1
        ret

Or also same story for:

void foo32_notvectorized(int *p, int *t) {
  int sum = 0;
  for (int i = 0; i != 32; i++)
    sum += p[i];
  *t = sum;
}

char foo32_vectorized(int *p) {
  int sum = 0;
  for (int i = 0; i != 32; i++)
    sum += p[i];
  return sum;
}

https://godbolt.org/z/3Gqrxc6nj

[davidbolvansky]

void foo32_notvectorized_def_in_loop(int *p, int *t) {
  int sum = 0;
  for (int i = 0; i != 32; i++)
    *t += p[i];
}

Another case. ICC is able to vectorize this code, emits runtime check.

https://reviews.llvm.org/D102834 will probably help here.

[RKSimon]

Reduction to stores can be enabled with the -slp-vectorize-hor-store llvm command line flag

[RKSimon]

Candidate Patch: https://reviews.llvm.org/D131894