http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55962
Manuel López-Ibáñez <manu at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|UNCONFIRMED |NEW Last reconfirmed| |2013-01-13 CC| |manu at gcc dot gnu.org Ever Confirmed|0 |1 --- Comment #1 from Manuel López-Ibáñez <manu at gcc dot gnu.org> 2013-01-13 19:17:05 UTC --- Confirmed. If you run GCC under GDB, you'll notice that the call if (expr == error_mark_node || TREE_CODE (expr) == TREE_LIST) { if (complain & tf_error) qualified_name_lookup_error (scope, TREE_OPERAND (qualified_id, 1), => expr, input_location); return error_mark_node; } uses input_location. I think it should use an explicit location. The problem is how to get the location of 'value' to this point. Up one-level input_location is "hacked" to be loc = input_location; if (EXPR_HAS_LOCATION (t)) input_location = EXPR_LOCATION (t); but in this case t is just: <scope_ref 0x7ffff753e640 tree_0 arg 0 <template_type_parm 0x7ffff75439d8 T type_0 type_6 VOID align 8 symtab 0 alias set -1 canonical type 0x7ffff75439d8 index 0 level 1 orig_level 1 chain <type_decl 0x7ffff753f958 T>> arg 1 <identifier_node 0x7ffff7552d10 value bindings <(nil)> local bindings <(nil)>>> so it doesn't have a location, so it uses input_location. The issue seems to be that input_location has somehow been smashed by something else. For the testcase: void assert(int, const char *); template <typename T> int foo(int b) { assert(T::value, "toto"); return b; } int main() { foo<int>(12); } we get: test.cc:6:26: error: ‘value’ is not a member of ‘int’ assert(T::value, "toto"); ^ which is slightly better and the only difference seems to be the value of input_location in the condition above. This requires much more investigation.