[llvm] [AArch64] Use isKnownNonZero to optimize eligible compares to cmn (PR #96349)

[via llvm-commits]

================
@@ -258,3 +258,34 @@ define i32 @neg_range_int(i32 %a, i32 %b, i32 %c) {
   ret i32 %retval.0
 }
 
+; (b > -3 || a < -(c | 1))
+define i32 @neg_range_int_cmn(i32 %a, i32 %b, i32 %c) {
----------------
AtariDreams wrote:

> In the new code in this test:
> 
> ```
> define i32 @or_neg(i32 %x, i32 %y) {
> ; CHECK-LABEL: or_neg:
> ; CHECK:       // %bb.0:
> ; CHECK-NEXT:    orr w8, w0, #0x1
> ; CHECK-NEXT:    cmn w8, w1
> ; CHECK-NEXT:    cset w0, gt
> ; CHECK-NEXT:    ret
>   %3 = or i32 %x, 1
>   %4 = sub i32 0, %3
>   %5 = icmp sgt i32 %4, %y
>   %6 = zext i1 %5 to i32
>   ret i32 %6
> }
> ```
> 
> If, say, x==0, y==0 then is the result not backwards?

Fixed

https://github.com/llvm/llvm-project/pull/96349