[C++ Patch] PR 58609

[Paolo Carlini]

Hi,

in this regression we ICE during error recovery, much, much later, in 
the middle-end. In order to avoid that it seems to me that we can safely 
return NULL_TREE.

I'm also taking the occasion to change a pair of errors to error + 
inform (this is also consistent with eg, the explain-once type messages 
we have got in do_friend and in resolve_address_of_overloaded_function)

Tested x86_64-linux.

Thanks,
Paolo.

////////////////////
/cp
2014-03-07  Paolo Carlini  <paolo.carl...@oracle.com>

        PR c++/58609
        * decl.c (check_initializer): Return NULL_TREE after error;
        consistently use inform.

/testsuite
2014-03-07  Paolo Carlini  <paolo.carl...@oracle.com>

        PR c++/58609
        * g++.dg/cpp0x/constexpr-ice12.C: New.
Index: cp/decl.c
===================================================================
--- cp/decl.c   (revision 208406)
+++ cp/decl.c   (working copy)
@@ -5809,9 +5809,11 @@ check_initializer (tree decl, tree init, int flags
               "member %qD", decl);
       if (!explained)
        {
-         error ("(an out of class initialization is required)");
+         inform (input_location,
+                 "(an out of class initialization is required)");
          explained = 1;
        }
+      return NULL_TREE;
     }
 
   return init_code;
Index: testsuite/g++.dg/cpp0x/constexpr-ice12.C
===================================================================
--- testsuite/g++.dg/cpp0x/constexpr-ice12.C    (revision 0)
+++ testsuite/g++.dg/cpp0x/constexpr-ice12.C    (working copy)
@@ -0,0 +1,9 @@
+// PR c++/58609
+// { dg-do compile { target c++11 } }
+
+struct A
+{
+  static constexpr int&& i = 0;  // { dg-error "initialization" }
+};
+
+int j = A::i;