[llvm] (Draft) [SCEV] forgetValue: support (extractvalue 0, (with-overflow-inst op0, op1)) (PR #98015)

[via llvm-commits]

================
@@ -1589,4 +1589,40 @@ TEST_F(ScalarEvolutionsTest, ApplyLoopGuards) {
   });
 }
 
+TEST_F(ScalarEvolutionsTest, ForgetValueWithOverflowInst) {
+  LLVMContext C;
+  SMDiagnostic Err;
+  std::unique_ptr<Module> M = parseAssemblyString(
+      "declare { i32, i1 } @llvm.smul.with.overflow.i32(i32, i32) "
+      "define void @foo(i32 %i) { "
+      "entry: "
+      "  br label %loop.body "
+      "loop.body: "
+      "  %iv = phi i32 [ %iv.next, %loop.body ], [ 0, %entry ] "
+      "  %iv.next = add nsw i32 %iv, 1 "
+      "  %call = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %iv, i32 -2) "
+      "  %extractvalue = extractvalue {i32, i1} %call, 0 "
+      "  %cmp = icmp eq i32 %iv.next, 16 "
+      "  br i1 %cmp, label %exit, label %loop.body "
+      "exit: "
+      "  ret void "
+      "} ",
+      Err, C);
----------------
v01dXYZ wrote:

If I only add a lit test, how can I be certain it will exactly test the eviction and not the fact that SCEV value was never created or the stale value used because the implementation of a pass has changed ?

I can add the case from the issue but it requires `LoopPeel` to produce a CFG that is out of sync with the SCEV value.

I don't know how to have a robust test for something as specific using an integration test.



https://github.com/llvm/llvm-project/pull/98015