[X86] Unnecessary repeated costly comparisons in multiple blocks

[RKSimon]

CGP's sinkCmpExpression is duplicating a costly CmpInst in multiple blocks instead of just reusing the result

using Word = unsigned _BitInt(128);
void doit(const Word &a, const Word &b, Word *p0, Word *p1) {
    bool eq = a == b;
    *p0++ |= Word((unsigned)eq);
    if (eq) {
        *p1++ |= Word((unsigned)eq);
    }
}

define void @doit(ptr %a, ptr %b, ptr %p0, ptr %p1) {
entry:
  %0 = load i128, ptr %a, align 8
  %1 = load i128, ptr %b, align 8
  %cmp = icmp eq i128 %0, %1
  %conv1 = zext i1 %cmp to i128
  %2 = load i128, ptr %p0, align 8
  %or = or i128 %2, %conv1
  store i128 %or, ptr %p0, align 8
  br i1 %cmp, label %if.then, label %if.end

if.then:
  %3 = load i128, ptr %p1, align 8
  %or7 = or i128 %3, %conv1
  store i128 %or7, ptr %p1, align 8
  br label %if.end

if.end:
  ret void
}

doit:
# %bb.0:                                # %entry
        movq    (%rdi), %rax
        movq    8(%rdi), %r8
        vmovdqu (%rdi), %xmm0
        movq    (%rsi), %r9
        movq    8(%rsi), %rdi
        vpxor   (%rsi), %xmm0, %xmm0
        xorl    %esi, %esi
        vptest  %xmm0, %xmm0  ;  FIRST COMPARISON
        sete    %sil
        orq     %rsi, (%rdx)
        vptest  %xmm0, %xmm0 ;  SECOND COMPARISON - use TEST ESI,ESI?
        jne     .LBB0_2
# %bb.1:                                # %if.then
        xorq    %r9, %rax
        xorq    %rdi, %r8
        xorl    %edx, %edx
        orq     %rax, %r8  ;  THIRD COMPARISON - ARGH!
        sete    %dl
        orq     %rdx, (%rcx)
.LBB0_2:                                # %if.end
        retq

The third comparison in particular is annoying as we can no longer fold to a VPTEST as the i128 are split into i64 pairs.

It looks like some targets have already started overriding TLI::hasMultipleConditionRegisters to hack this - we might just want to do the same on x86.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

CGP's sinkCmpExpression is duplicating a costly CmpInst in multiple blocks instead of just reusing the result ```cpp using Word = unsigned _BitInt(128); void doit(const Word &a, const Word &b, Word *p0, Word *p1) { bool eq = a == b; *p0++ |= Word((unsigned)eq); if (eq) { *p1++ |= Word((unsigned)eq); } } ``` ```ll define void @doit(ptr %a, ptr %b, ptr %p0, ptr %p1) { entry: %0 = load i128, ptr %a, align 8 %1 = load i128, ptr %b, align 8 %cmp = icmp eq i128 %0, %1 %conv1 = zext i1 %cmp to i128 %2 = load i128, ptr %p0, align 8 %or = or i128 %2, %conv1 store i128 %or, ptr %p0, align 8 br i1 %cmp, label %if.then, label %if.end

if.then:
%3 = load i128, ptr %p1, align 8
%or7 = or i128 %3, %conv1
store i128 %or7, ptr %p1, align 8
br label %if.end

if.end:
ret void
}

```asm
doit:
# %bb.0:                                # %entry
        movq    (%rdi), %rax
        movq    8(%rdi), %r8
        vmovdqu (%rdi), %xmm0
        movq    (%rsi), %r9
        movq    8(%rsi), %rdi
        vpxor   (%rsi), %xmm0, %xmm0
        xorl    %esi, %esi
        vptest  %xmm0, %xmm0  ;  FIRST COMPARISON
        sete    %sil
        orq     %rsi, (%rdx)
        vptest  %xmm0, %xmm0 ;  SECOND COMPARISON - use TEST ESI,ESI?
        jne     .LBB0_2
# %bb.1:                                # %if.then
        xorq    %r9, %rax
        xorq    %rdi, %r8
        xorl    %edx, %edx
        orq     %rax, %r8  ;  THIRD COMPARISON - ARGH!
        sete    %dl
        orq     %rdx, (%rcx)
.LBB0_2:                                # %if.end
        retq

The third comparison in particular is annoying as we can no longer fold to a VPTEST as the i128 are split into i64 pairs.

It looks like some targets have already started overriding TLI::hasMultipleConditionRegisters to hack this - we might just want to do the same on x86.