[llvm] [X86] Prefer carry to overflow flag when using multiplication (PR #221611)

[via llvm-commits]

AZero13 wrote:

> Add a test case for the folding?

I had to force-push to add it to precommit. Here it is:

define i32 @mul_overflow_apx(i32 %a, i8 %b, i8 %c) {
; CHECK-LABEL: mul_overflow_apx:
; CHECK:       # %bb.0:
; CHECK-NEXT:    movl %esi, %eax # encoding: [0x89,0xf0]
; CHECK-NEXT:    # kill: def $al killed $al killed $eax
; CHECK-NEXT:    mulb %dl # encoding: [0xf6,0xe2]
; CHECK-NEXT:    adcl $0, %edi, %eax # encoding: [0x62,0xf4,0x7c,0x18,0x83,0xd7,0x00]
; CHECK-NEXT:    retq # encoding: [0xc3]
  %umul = tail call { i8, i1 } @llvm.umul.with.overflow.i8(i8 %b, i8 %c)
  %umul.overflow = extractvalue { i8, i1 } %umul, 1
  %conv2 = zext i1 %umul.overflow to i32
  %add = add i32 %a, %conv2
  ret i32 %add
}

https://github.com/llvm/llvm-project/pull/221611