combineVSelectWithAllOnesOrZeros - extend to support X86ISD::BLENDV

[RKSimon]

https://llvm.godbolt.org/z/e4hG7cE3q

define void @masked_add(<8 x i32>* %pa, <8 x i32>* %pb, i8 %bitmask) {
  %a = load <8 x i32>, <8 x i32>* %pa, align 32
  %b = load <8 x i32>, <8 x i32>* %pb, align 32
  %m = bitcast i8 %bitmask to <8 x i1>
  %s = select <8 x i1> %m, <8 x i32> %b, <8 x i32> zeroinitializer
  %r = add <8 x i32> %s, %a
  store <8 x i32> %r, <8 x i32>* %pa, align 32
  ret void
}

llc -mcpu=znver2

masked_add: # @masked_add
  vmovdqa .LCPI0_0(%rip), %ymm1 # ymm1 = [1,2,4,8,16,32,64,128]
  vmovd %edx, %xmm0
  vpbroadcastb %xmm0, %ymm0
  vpand %ymm1, %ymm0, %ymm0
  vpcmpeqd %ymm1, %ymm0, %ymm0
  vpxor %xmm1, %xmm1, %xmm1
  vblendvps %ymm0, (%rsi), %ymm1, %ymm0
  vpaddd (%rdi), %ymm0, %ymm0
  vmovdqa %ymm0, (%rdi)
  vzeroupper
  retq

We should be folding away the vblendvps to either and/andnot depending on whether any of the inputs are 0/-1

[llvmbot]

@llvm/issue-subscribers-backend-x86

[RKSimon]

This was fixed by f591231