[PATCH] Delay lookup of simple default template arguments under -fms-compatibility

[Richard Smith]

================
Comment at: include/clang/AST/Type.h:4056-4057
@@ -4055,4 +4055,4 @@
 
 /// \brief Represents a qualified type name for which the type name is
 /// dependent.
 ///
----------------
You seem to be using this to represent an unqualified type in some cases? (You've made `NNS` optional.) A quick look indicates that downstream code will not do the right thing in this case. (`TreeTransform` will treat it as if an error has occurred, and bail out with no diagnostic, for instance.)

================
Comment at: lib/Sema/SemaDecl.cpp:361
@@ +360,3 @@
+                                       RD->getTypeForDecl());
+  return SubNNS;
+}
----------------
Reid Kleckner wrote:
> Richard Smith wrote:
> > Once we hit this, we should bail out entirely. (This can currently include local classes in the name specifier, which doesn't seem right.)
> Bail out entirely, as in, return nullptr, or as in, fail the lookup?
I think this has been made irrelevant by other changes.

================
Comment at: lib/Sema/SemaDecl.cpp:353-357
@@ +352,7 @@
+    auto *ND = dyn_cast<NamespaceDecl>(DC);
+    if (ND && !ND->isInline() && !ND->isAnonymousNamespace())
+      return NestedNameSpecifier::Create(Context, nullptr, ND);
+    else if (auto *RD = dyn_cast<CXXRecordDecl>(DC))
+      return NestedNameSpecifier::Create(Context, nullptr, RD->isTemplateDecl(),
+                                         RD->getTypeForDecl());
+  }
----------------
What about the translation unit? Should you create a `::` `NestedNameSpecifier` for it? (Once you do that, the `'return nullptr;'` below and the null `NNS` case in `DependentNameType` are unreachable.)

================
Comment at: test/SemaTemplate/ms-delayed-default-template-args.cpp:6-9
@@ +5,6 @@
+namespace test_basic {
+template <typename T = Baz> // expected-warning {{using the undeclared type 'Baz' as a default template argument is a Microsoft extension}}
+struct Foo { T x; };
+typedef int Baz;
+template struct Foo<>;
+}
----------------
Reid Kleckner wrote:
> Richard Smith wrote:
> > Maybe also add a test for
> > 
> >   template<typename T> struct Something {};
> >   template<typename T = Something<Bar>> struct SomethingElse;
> > 
> > ... and
> > 
> >   template<Something<Bar> *p> struct AnotherThing;
> I assume Bar is always undeclared in these tests.
> 
> MSVC accepts your first test case, but Clang doesn't with my patch.  I'd need to find a way to delay instantiating Something<Bar> until later, which feels beyond the scope of this patch.
> 
> The second test is a negative test, right?  Both we and MSVC need to know the type of the non-type template parameter.
I had meant for these to both be negative tests (maybe with a FIXME for the first case).

http://reviews.llvm.org/D3995