[PATCH] Improve error recovery around colon.

[Serge Pavlov]

  Thank you for feedback!


================
Comment at: include/clang/Basic/DiagnosticParseKinds.td:342-343
@@ -341,1 +341,4 @@
   "unexpected ':' in nested name specifier; did you mean '::'?">;
+def err_unexected_token_in_nested_name_spec : Error<
+  "unexpected '%0' in nested name specifier; did you mean ':'?">;
+def err_nested_name_spec_is_not_class : Error<
----------------
Richard Smith wrote:
> Either the second thing here should be `%1` (and you pass in `tok::colon`), or this diagnostic should have a better name. `"did you mean ':'?"` is not an obvious thing to say about a nested name specifier =)
> 
> Also, using `%1` will let you merge this with the previous diagnostic.
> 
> While you're here, please change `unexected` to `unexpected` in both diagnostic names. :)
Yes, I see. Tried to make more descriptive message.
Also changed typos in diagnostic names :)

================
Comment at: lib/Parse/ParseExprCXX.cpp:470-483
@@ -452,1 +469,16 @@
+                                              ObjectType, EnteringContext, SS,
+                                              ErrMode, &IdDecl)) {
+        // Identifier is not recognized as a nested name, but we can have
+        // mistyped '::' instead of ':'.
+        if (ColonIsSacred && IdDecl) {
+          Diag(CCLoc, diag::err_nested_name_spec_is_not_class)
+              << &II << getLangOpts().CPlusPlus
+              << FixItHint::CreateReplacement(CCLoc, ":");
+          Diag(IdDecl->getLocation(), diag::note_declared_at);
+          ColonColon.setKind(tok::colon);
+          PP.EnterToken(Tok);
+          PP.EnterToken(ColonColon);
+          Tok = Identifier;
+          break;
+        }
         SS.SetInvalid(SourceRange(IdLoc, CCLoc));
----------------
Richard Smith wrote:
> I think it'd be cleaner to pass a `bool *CorrectToColon` into `ActOnCXXNestedNameSpecifier`, and issue the diagnostics there (so this diagnostic is not split into two disparate places). Then, if the flag is set, just fix up the token stream here.
This is good approach. Looks like patch becomes more compact.

================
Comment at: lib/Sema/SemaCXXScopeSpec.cpp:521
@@ +520,3 @@
+            << &Identifier << getLangOpts().CPlusPlus;
+      if (NamedDecl *ND = R.getAsSingle<NamedDecl>()) {
+        if (RecoveryMode == ReportAll)
----------------
Richard Smith wrote:
> This will silently ignore the error if you're in `DontReportWrongMember` mode and name lookup finds something other than a single result.
Remade this test.


http://llvm-reviews.chandlerc.com/D2870