[clang] Reland [clang] Handle instantiated members to determine visibility (#136128) (PR #136689)

Erich Keane via cfe-commits cfe-commits at lists.llvm.org
Wed Apr 23 06:24:46 PDT 2025


================
@@ -4787,8 +4787,10 @@ LinkageInfo LinkageComputer::computeTypeLinkageInfo(const Type *T) {
     return computeTypeLinkageInfo(cast<ReferenceType>(T)->getPointeeType());
   case Type::MemberPointer: {
     const auto *MPT = cast<MemberPointerType>(T);
-    LinkageInfo LV =
-        getDeclLinkageAndVisibility(MPT->getMostRecentCXXRecordDecl());
+    LinkageInfo LV;
+    if (CXXRecordDecl *D = MPT->getMostRecentCXXRecordDecl()) {
----------------
erichkeane wrote:

> > Should we be taking visibility from that `getClass` in this case? It is a dependent expression in this case, but perhaps we should be?
> 
> I don't know how to construct a test case where a dependent class has visibility or linkage. We can do the following, but probably shouldn't unless there is a test case:
> 
> ```diff
> diff --git a/clang/lib/AST/Type.cpp b/clang/lib/AST/Type.cpp
> index 5783f8701273..951249b8c495 100644
> --- a/clang/lib/AST/Type.cpp
> +++ b/clang/lib/AST/Type.cpp
> @@ -4790,6 +4790,8 @@ LinkageInfo LinkageComputer::computeTypeLinkageInfo(const Type *T) {
>      LinkageInfo LV;
>      if (CXXRecordDecl *D = MPT->getMostRecentCXXRecordDecl()) {
>        LV.merge(getDeclLinkageAndVisibility(D));
> +    } else if (const Type *Ty = MPT->getQualifier()->getAsType()) {
> +      LV.merge(computeTypeLinkageInfo(Ty));
>      }
>      LV.merge(computeTypeLinkageInfo(MPT->getPointeeType()));
>      return LV;
> ```

I'm sympathetic to the 'don't unless there is a test case', but I'm leaning the other way: we should get this as right as we can, even if we don't know what a reproducer looks like.  

https://github.com/llvm/llvm-project/pull/136689


More information about the cfe-commits mailing list