Hi,
in this regression we ICE during error recovery after an additional
redundant error message. I think it's one of those cases (we have got
quite a few elsewhere, in semantics.c too) when it's better to
immediately return error_mark_node when mark_used returns false, even if
we aren't in a SFINAE context. To be sure, I double checked that in
those cases mark_used certainly issues an error, thus we aren't risking
creating accepts-invalid bugs, it's only matter of fine tuning error
recovery. Tested x86_64-linux.
Thanks,
Paolo.
/////////////////////////////
/cp
2016-04-14 Paolo Carlini <paolo.carl...@oracle.com>
PR c++/70540
* semantics.c (process_outer_var_ref): Unconditionally return
error_mark_node when mark_used returns false.
/testsuite
2016-04-14 Paolo Carlini <paolo.carl...@oracle.com>
PR c++/70540
* g++.dg/cpp0x/auto47.C: New.
Index: cp/semantics.c
===================================================================
--- cp/semantics.c (revision 234970)
+++ cp/semantics.c (working copy)
@@ -3276,7 +3276,7 @@ process_outer_var_ref (tree decl, tsubst_flags_t c
tree initializer = convert_from_reference (decl);
/* Mark it as used now even if the use is ill-formed. */
- if (!mark_used (decl, complain) && !(complain & tf_error))
+ if (!mark_used (decl, complain))
return error_mark_node;
bool saw_generic_lambda = false;
Index: testsuite/g++.dg/cpp0x/auto47.C
===================================================================
--- testsuite/g++.dg/cpp0x/auto47.C (revision 0)
+++ testsuite/g++.dg/cpp0x/auto47.C (working copy)
@@ -0,0 +1,8 @@
+// PR c++/70540
+// { dg-do compile { target c++11 } }
+
+void
+foo ()
+{
+ auto f = [&] { return f; }; // { dg-error "before deduction" }
+}
