[clang] [clang][modules] giving the __stddef_ headers their own modules can cause redeclaration errors with -fbuiltin-headers-in-system-modules (PR #84127)

[Ian Anderson via cfe-commits]

================
@@ -301,10 +301,9 @@ bool Module::directlyUses(const Module *Requested) {
     if (Requested->isSubModuleOf(Use))
       return true;
 
-  // Anyone is allowed to use our builtin stdarg.h and stddef.h and their
-  // accompanying modules.
-  if (Requested->getTopLevelModuleName() == "_Builtin_stdarg" ||
-      Requested->getTopLevelModuleName() == "_Builtin_stddef")
+  // Anyone is allowed to use our builtin stddef.h and its accompanying modules.
+  if (Requested->fullModuleNameIs({"_Builtin_stddef", "max_align_t"}) ||
+      Requested->fullModuleNameIs({"_Builtin_stddef_wint_t"}))
----------------
ian-twilightcoder wrote:

We're taking `[no_undeclared_includes]` to imply `-fbuiltin-headers-in-system-modules`, where `_Builtin_stdarg` is empty. The assumption is that `[no_undeclared_includes]` is a workaround for the C++ headers not layering when the C stdlib headers are all in the same module, which is also what `-fbuiltin-headers-in-system-modules` is for. If we don't think that's a good assumption, then maybe we should add all of the clang C stdlib modules here (but it's been fine so far keeping the assumption).

https://github.com/llvm/llvm-project/pull/84127