[llvm] [InstCombine] Preserve the nsw flag for (X | Op01C) + Op1C --> X + (Op01C + Op1C) (PR #94586)

[Nikita Popov via llvm-commits]

================
@@ -4137,5 +4137,25 @@ define i32 @fold_zext_nneg_add_const_fail2(i8 %x) {
   ret i32 %r
 }
 
+define i8 @fold_add_constant_preserve_nsw(i8 %x) {
+; CHECK-LABEL: @fold_add_constant_preserve_nsw(
+; CHECK-NEXT:    [[ADD:%.*]] = add nsw i8 [[X:%.*]], -120
+; CHECK-NEXT:    ret i8 [[ADD]]
+;
+  %or = or disjoint i8 %x, -128
+  %add = add nsw i8 %or, 8
+  ret i8 %add
+}
+
+define i8 @fold_add_constant_no_nsw(i8 %x) {
+; CHECK-LABEL: @fold_add_constant_no_nsw(
+; CHECK-NEXT:    [[ADD:%.*]] = add i8 [[X:%.*]], 120
+; CHECK-NEXT:    ret i8 [[ADD]]
+;
+  %or = or disjoint i8 %x, -128
+  %add = add nsw i8 %or, -8
+  ret i8 %add
+}
----------------
nikic wrote:

Can you please find where the existing tests for this transform are (e.g. by commenting it out and seeing what fails), and move these tests there, together with the sdiv_to_udiv test you added below)?

https://github.com/llvm/llvm-project/pull/94586