[PATCH] MSVC compat: Allow lookup of friend types in enclosing namespaces

[Reid Kleckner]

================
Comment at: include/clang/Basic/DiagnosticSemaKinds.td:1000-1002
@@ -999,2 +999,5 @@
   InGroup<UnsupportedFriend>;
-  
+def ext_friend_tag_redecl_outside_namespace : ExtWarn<
+  "unqualified friend declarations referring to types outside of the nearest "
+  "enclosing namespace is a Microsoft extension; add a nested name specifier to fix">,
+  InGroup<Microsoft>;
----------------
Richard Smith wrote:
> Please add a bit more information here: which NNS should be added? What is the enclosing namespace that we looked in? Which type did we actually find? A fixit adding the NNS would help.
Done.  It's a very nice diagnostic now, although maybe this is overkill for a fairly pedantic portion of the standard where implementations disagree.  :)

================
Comment at: include/clang/Basic/DiagnosticSemaKinds.td:1001
@@ +1000,3 @@
+def ext_friend_tag_redecl_outside_namespace : ExtWarn<
+  "unqualified friend declarations referring to types outside of the nearest "
+  "enclosing namespace is a Microsoft extension; add a nested name specifier to fix">,
----------------
Richard Smith wrote:
> Singular/plural mismatch ("declarations ... is a MS extension"). Please make this singular.
done

================
Comment at: include/clang/Basic/DiagnosticSemaKinds.td:1002
@@ +1001,3 @@
+  "unqualified friend declarations referring to types outside of the nearest "
+  "enclosing namespace is a Microsoft extension; add a nested name specifier to fix">,
+  InGroup<Microsoft>;
----------------
Richard Smith wrote:
> Maybe drop the "to fix"?
done

================
Comment at: lib/Sema/SemaDecl.cpp:11467
@@ -11453,3 +11466,3 @@
   if (TUK == TUK_Friend)
     New->setObjectOfFriendDecl(!FriendSawTagOutsideEnclosingNamespace &&
                                getLangOpts().MicrosoftExt);
----------------
Richard Smith wrote:
> This was a prior attempt to get ~half of the same effect. In particular, for:
> 
>     struct X {};
>     namespace N {
>       struct Y { friend struct X; };
>       X x;
>     }
> 
> ... we need to not inject `X` into `N`, or the lookup for `N::x`'s type will find the wrong struct.
> 
> Perhaps we should remove this hack (for `MicrosoftExt`) now that we have a more faithful hack (for `MSVCCompat`)? It's not a conforming extension. That is, replace this with just
> 
>     New->setObjectOfFriendDecl(getLangOpts().MSVCCompat);
I agree, I'd rather have just one faithful hack, but presumably Doug put this under -fms-extensions for a reason.  Hopefully we can raise him.  :)

I'll make the change, since I think we both agree this shouldn't be under -fms-extensions.

http://reviews.llvm.org/D4443