[PATCH] Partial specialization after class template instantiation.

[Richard Smith]

================
Comment at: ../llvm/tools/clang/include/clang/Basic/DiagnosticSemaKinds.td:3395
@@ -3394,1 +3394,3 @@
   "explicit specialization of %0 after instantiation">;
+def err_part_specialization_after_instantiation : Error<
+  "partial specialization of %0 after instantiation">;
----------------
Please use `partial_spec` or `partial_specialization` rather than `part_specialization`, like other diagnostics do.

================
Comment at: ../llvm/tools/clang/lib/Sema/SemaTemplate.cpp:6225
@@ +6224,3 @@
+  //   unit in which such a use occurs; no diagnostic is required.
+  if (isPartialSpecialization) {
+    auto *ThisPartialSpec =
----------------
We don't need these checks for a redeclaration:

  if (isPartialSpecialization && !PrevDecl) {

================
Comment at: ../llvm/tools/clang/lib/Sema/SemaTemplate.cpp:6227
@@ +6226,3 @@
+    auto *ThisPartialSpec =
+        static_cast<ClassTemplatePartialSpecializationDecl *>(Specialization);
+    for (const auto &S : ClassTemplate->specializations()) {
----------------
Use `cast`, not `static_cast`, here. (It'll assert if the type doesn't match.)

================
Comment at: ../llvm/tools/clang/lib/Sema/SemaTemplate.cpp:6232
@@ +6231,3 @@
+          S->hasDefinition() &&
+          !DeduceTemplateArguments(ThisPartialSpec, S->getTemplateArgs(), Info)) {
+        auto *InstantiatedFrom =
----------------
Hmm, what happens if we get an error outside of the immediate context during this deduction? Are we allowed to reject the program for that reason?

For instance:

  template<typename T> struct X { typedef typename T::type type; };
  template<typename T, typename U> struct A {};
  A<int, int> Aint;
  template<typename T> struct A<T, typename X<T>::type> {};

Do we provide a sufficiently useful template instantiation backtrace in this case?

http://reviews.llvm.org/D5744