Only suppress instance context if a member is actually
accessible in its declaring class;  otherwise we might
fail to apply [class.protected] when considering
accessibility in derived classes.

Noticed by inspection; <rdar://13270329>.

I had an existing test wrong.  Here's why it's wrong:

Follow the rules (and notation) of [class.access]p5.
The naming class (N) is B and the context (R) is D::getX.
- 'x' as a member of B is protected, but R does not occur
  in a member or friend of a class derived from B.
- There does exist a base class of B, A, which is accessible
  from R, and 'x' is accessible at R when named in A because
  'x' as a member of A is protected and R occurs in a member
  of a class, D, that is derived from A;  however, by
  [class.protected], the class of the object expression must
  be equal to or derived from that class, and A does not
  derive from D.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@175858 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Sema/SemaAccess.cpp b/lib/Sema/SemaAccess.cpp
index 229e6f4..55bd76b 100644
--- a/lib/Sema/SemaAccess.cpp
+++ b/lib/Sema/SemaAccess.cpp
@@ -1317,7 +1317,13 @@
     FinalAccess = Target->getAccess();
     switch (HasAccess(S, EC, DeclaringClass, FinalAccess, Entity)) {
     case AR_accessible:
+      // Target is accessible at EC when named in its declaring class.
+      // We can now hill-climb and simply check whether the declaring
+      // class is accessible as a base of the naming class.  This is
+      // equivalent to checking the access of a notional public
+      // member with no instance context.
       FinalAccess = AS_public;
+      Entity.suppressInstanceContext();
       break;
     case AR_inaccessible: break;
     case AR_dependent: return AR_dependent; // see above
@@ -1325,8 +1331,6 @@
 
     if (DeclaringClass == NamingClass)
       return (FinalAccess == AS_public ? AR_accessible : AR_inaccessible);
-
-    Entity.suppressInstanceContext();
   } else {
     FinalAccess = AS_public;
   }