[clang] [Clang][Sema] Fix the lambda call expression inside of a type alias declaration (PR #82310)

via cfe-commits cfe-commits at lists.llvm.org
Tue Mar 5 03:31:40 PST 2024


================
@@ -13905,6 +13911,31 @@ TreeTransform<Derived>::TransformLambdaExpr(LambdaExpr *E) {
                                     /*IsInstantiation*/ true);
   SavedContext.pop();
 
+  // Recompute the dependency of the lambda so that we can defer the lambda call
+  // construction until after we have sufficient template arguments. For
+  // example, template <class> struct S {
+  //   template <class U>
+  //   using Type = decltype([](U){}(42.0));
+  // };
+  // void foo() {
+  //   using T = S<int>::Type<float>;
+  //             ^~~~~~
+  // }
+  // We would end up here from instantiating the S<int> as we're ensuring the
+  // completeness. That would make us transform the lambda call expression
+  // despite the fact that we don't see the argument for U yet. We have a
+  // mechanism that circumvents the semantic checking if the CallExpr is
+  // dependent. We can harness that by recomputing the lambda dependency from
+  // the instantiation arguments. I'm putting it here rather than the above
+  // since we can see transformed lambda parameters in case that they're
+  // useful for calculation.
+  DependencyKind = getDerived().ComputeLambdaDependency(&LSICopy);
+  Class->setLambdaDependencyKind(DependencyKind);
+  // Clean up the type cache created previously. Then, we re-create a type for
+  // such Decl with the new DependencyKind.
+  Class->setTypeForDecl(nullptr);
+  getSema().Context.getTypeDeclType(Class);
+
----------------
cor3ntin wrote:

This is not great, but I am not sure how we can improve. 
I think it's worth a fixme
@erichkeane 

https://github.com/llvm/llvm-project/pull/82310


More information about the cfe-commits mailing list