[llvm] Adding abdu/abds to canCreateUndefOrPoison (PR #149017)

[Simon Pilgrim via llvm-commits]

================
@@ -395,3 +395,51 @@ define i64 @freeze_array() {
   %t1 = add i64 %v1, %v2
   ret i64 %t1
 }
+
+define i32 @freeze_abdu(i8 %x, i8 %y) {
+; CHECK-SD-LABEL: freeze_abdu:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    and w8, w0, #0xff
+; CHECK-SD-NEXT:    sub w8, w8, w1, uxtb
+; CHECK-SD-NEXT:    cmp w8, #0
+; CHECK-SD-NEXT:    cneg w0, w8, mi
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: freeze_abdu:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    and w8, w0, #0xff
+; CHECK-GI-NEXT:    sub w8, w8, w1, uxtb
+; CHECK-GI-NEXT:    cmp w8, #0
+; CHECK-GI-NEXT:    cneg w0, w8, le
+; CHECK-GI-NEXT:    ret
+  %a   = zext   i8 %x to i32
+  %b   = zext   i8 %y to i32
+  %d   = sub    i32 %a, %b
+  %t   = call   i32 @llvm.abs.i32(i32 %d, i1 false)
+  %f   = freeze i32 %t
+  ret  i32 %f
+}
----------------
RKSimon wrote:

This isn't testing anything yet - we need to use the aarch64 `@llvm.aarch64.neon.?abd.*` intrinsics directly to ensure we start from a ABDU/S node (look at arm64-vabs.ll for examples). We then need to prove that the freeze has been moved - something similar to what I've proposed for the AVG nodes here: https://github.com/llvm/llvm-project/pull/148191#discussion_r2204332512

https://github.com/llvm/llvm-project/pull/149017