[X86] AVX512 integer compare fails to use icmp_sle(x,0) instead of icmp_slt(x,1)

[RKSimon]

Noticed while investigating SIGN patterns:

define i16 @cmp_sle_1(<16 x i8> %x) {
  %c = icmp sle <16 x i8> %x, splat (i8 0)
  %x = bitcast <16 x i1> %c to i16
  ret i16 %x
}

Both InstCombine and DAG fold this to:

define i16 @cmp_sle_1(<16 x i8> %x) {
  %c = icmp slt <16 x i8> %x, splat (i8 1)
  %x = bitcast <16 x i1> %c to i16
  ret i16 %x
}

-mcpu=x86-64-v3 - no ICMP_SLE handling so uses NOT(ICMP_SGT)

cmp_sle_1: # @cmp_sle_1
# %bb.0:
  vpxor %xmm0, %xmm0, %xmm0
  vpcmpgtb %xmm0, %xmm1, %xmm0
  vpmovmskb %xmm0, %eax
  notl %eax
  # kill: def $ax killed $ax killed $eax
  retq
  # -- End function

-mcpu=x86-64-v4 - VPCMP can use all condcodes, but chooses to load a constant:

.LCPI0_0:
  .zero 16,1
cmp_sle_1: # @cmp_sle_1
# %bb.0:
  vpcmpltb .LCPI0_0(%rip), %xmm1, %k0
  kmovd %k0, %eax
  # kill: def $ax killed $ax killed $eax
  retq
  # -- End function

[llvmorg-github-actions]

@llvm/issue-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Details

Noticed while investigating SIGN patterns: ``` define i16 @cmp_sle_1(<16 x i8> %x) { %c = icmp sle <16 x i8> %x, splat (i8 0) %x = bitcast <16 x i1> %c to i16 ret i16 %x } ``` Both InstCombine and DAG fold this to: ``` define i16 @cmp_sle_1(<16 x i8> %x) { %c = icmp slt <16 x i8> %x, splat (i8 1) %x = bitcast <16 x i1> %c to i16 ret i16 %x } ``` -mcpu=x86-64-v3 - no ICMP_SLE handling so uses NOT(ICMP_SGT) ``` cmp_sle_1: # @cmp_sle_1 # %bb.0: vpxor %xmm0, %xmm0, %xmm0 vpcmpgtb %xmm0, %xmm1, %xmm0 vpmovmskb %xmm0, %eax notl %eax # kill: def $ax killed $ax killed $eax retq # -- End function ``` -mcpu=x86-64-v4 - VPCMP can use all condcodes, but chooses to load a constant: ``` .LCPI0_0: .zero 16,1 cmp_sle_1: # @cmp_sle_1 # %bb.0: vpcmpltb .LCPI0_0(%rip), %xmm1, %k0 kmovd %k0, %eax # kill: def $ax killed $ax killed $eax retq # -- End function ```

[Shaurya2k06]

This issue is real; and I’d like to take this up. Plan is to keep the InstCombine canonical form and, in X86 DAG combining, rewrite AVX512 mask slt/sgt vs splat(±1) back to sle/sge vs zero so VPCMP doesn’t load a constant.

Let me know if you'd like to have it another way.
cc @RKSimon