[Bug tree-optimization/115531] vectorizer generates inefficient code for masked conditional update loops

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115531

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
AVX512 produces

.L3:
        vmovdqu8        (%rsi), %zmm9{%k1}
        kshiftrq        $32, %k1, %k5
        kshiftrq        $48, %k1, %k4
        movl    %r9d, %eax
        vmovdqu32       128(%rcx), %zmm7{%k5}
        subl    %esi, %eax
        movl    $64, %edi
        vmovdqu32       128(%rdx), %zmm3{%k5}
        kshiftrq        $16, %k1, %k6
        addl    %r10d, %eax
        vmovdqu32       192(%rcx), %zmm8{%k4}
        cmpl    %edi, %eax
        vmovdqu32       192(%rdx), %zmm4{%k4}
        cmova   %edi, %eax
        addq    $64, %rsi
        addq    $256, %rcx
        vmovdqu32       -256(%rcx), %zmm5{%k1}
        vmovdqu32       (%rdx), %zmm1{%k1}
        vmovdqu32       -192(%rcx), %zmm6{%k6}
        vmovdqu32       64(%rdx), %zmm2{%k6}
        vpcmpb  $4, %zmm14, %zmm9, %k2
        kshiftrq        $32, %k2, %k3
        vpblendmd       %zmm7, %zmm3, %zmm10{%k3}
        kshiftrd        $16, %k3, %k3
        vpblendmd       %zmm8, %zmm4, %zmm0{%k3}
        vpblendmd       %zmm5, %zmm1, %zmm12{%k2}
        vmovdqu32       %zmm10, 128(%rdx){%k5}
        kshiftrd        $16, %k2, %k2
        vmovdqu32       %zmm0, 192(%rdx){%k4}
        vpblendmd       %zmm6, %zmm2, %zmm11{%k2}
        vpbroadcastb    %eax, %zmm0
        movl    %r9d, %eax
        vmovdqu32       %zmm12, (%rdx){%k1}
        subl    %esi, %eax
        addl    %r8d, %eax
        vmovdqu32       %zmm11, 64(%rdx){%k6}
        addq    $256, %rdx
        vpcmpub $6, %zmm13, %zmm0, %k1
        cmpl    $64, %eax
        ja      .L3


The vectorizer sees

  <bb 4> [local count: 955630225]:
  # i_26 = PHI <i_23(9), 0(16)>
  _1 = (long unsigned int) i_26;
  _2 = _1 * 4;
  _3 = c_17(D) + _2;
  res_18 = *_3;
  _4 = stride_14(D) + i_26;
  _5 = (long unsigned int) _4;
  _6 = _5 * 4;
  _7 = b_19(D) + _6;
  t_20 = *_7;
  _8 = a_21(D) + _1;
  _9 = *_8;
  _34 = _9 != 0;
  res_11 = _34 ? t_20 : res_18;
  *_3 = res_11;
  i_23 = i_26 + 1;
  if (n_16(D) > i_23)

I believe that to get proper vectorizer costing we want to have an
optimization phase that can take into account whether we use a masked
loop or not.  Note that your intended transform relies on identifying
the open-coded "conditional store"

      int res = c[i];
      if (a[i] != 0)
        res = t;
      c[i] = res;

As Andrew says when that's a .MASK_STORE it's going to be easier to
identify the opportunity.  So yeah, if-conversion could recognize
this pattern and produce a .MASK_STORE from it as a first step.