[PATCH] D19479: 26748 - clang-cl fails to compile atlctrlw.h header from WTL

[Reid Kleckner via cfe-commits]

rnk added a comment.

You've described the core issue with parsing C++ templates prior to instantiation, you have to know what identifiers are types, and that's why we need typename. :) MSVC only parses templates after instantiation, so it doesn't need typename.

To deal with the specific case in ATL, we can use the 'new' expression as a hint that the unknown identifier is actually a dependent type. There should be some bit of context we can look at in DiagnoseUnknownTypeName to know that we are in a new expression inside a dependent template, and then form a DependentNameType to delay checking until template instantiation.

In the general case, I don't think we can ever behave exactly like MSVC without implementing token-based template instantiation, and nobody wants to add an entirely new, untested, separate codepath just for MSVC-style template instantiation. It means giving up on compiling invalid C++ that MSVC accepts, such as this:

  template <typename T>
  struct A : T {
    void f() {
      g(TypeOrFunction());
    }
  };
  // In A<B>, it's a type
  struct B {
    struct TypeOrFunction { };
    void g(TypeOrFunction);
  };
  template struct A<B>;
  // In A<C>, it's a function
  struct C {
    int TypeOrFunction();
    void g(int);
  };
  template struct A<C>;


http://reviews.llvm.org/D19479