[clang] [C++20] [Modules] Don't set modules owner ship information for builtin declarations (PR #102115)

[Matheus Izvekov via cfe-commits]

================
@@ -2376,6 +2376,12 @@ NamedDecl *Sema::LazilyCreateBuiltin(IdentifierInfo *II, unsigned ID,
   FunctionDecl *New = CreateBuiltin(II, R, ID, Loc);
   RegisterLocallyScopedExternCDecl(New, S);
 
+  // Builtin functions shouldn't be owned by any module.
+  if (New->hasOwningModule()) {
+    New->setLocalOwningModule(nullptr);
----------------
mizvekov wrote:

> This is somewhat coincidence. We would create the builtin TypedefDecls in `Sema::Initialize()`. But we haven't assign a module for the TU that time. So it works more or less surprisingly. The difference between the builtin typedef declaration and the current builtin function decl is that, the builtin typedef declaration are created eagerly and the builtin function decls is created lazily.

I see. This eagerness of creating these TypedefDecls is something that annoys me a little as well.
I think whatever outcome we end up with, should be the same for both.


> BTW, another idea may be, attach the builtin function decl to the global module. But now, the implementation may only create the global module (in implementation, it is actually the global module fragment in the specification) conditionally, in another word, we will only create the global module fragment after we see `module;`. So the idea may be more complex.

Yeah, attaching these things to the global module seems like a clean approach as well, I like that :)
What are the difficulties involved in always creating the GMF in this case?


https://github.com/llvm/llvm-project/pull/102115