Hi,

I'm having a look at this ICE during error recovery regression and I 
have a couple of different proposals which both pass testing. In the 
first case, instead of reaching (in cp_parser_cache_defarg):

   default_argument = make_node (DEFAULT_ARG);
   DEFARG_TOKENS (default_argument)
     = cp_token_cache_new (first_token, token);
   DEFARG_INSTANTIATIONS (default_argument) = NULL;

   return default_argument;

we could simply return an error_mark_node right after:

error_at (token->location, "file ends in default argument");

and certainly avoid the ICE in merge_exception_specifiers much later. 
Alternately, maybe it's too heavy handed, maybe fatal_errors are really 
appropriate in a very restrict set of cases (not sure), it would work to 
simply change the error_at to fatal_error, which also cuts some 
redundant diagnostic we produce right now after the sensible one about 
file ends. The idea occurred to me in particular because the error_at 
above does *not* result (and never did to my best knowledge) in a actual 
location being printed, we issue:

cc1plus: error: file ends in default argument

which smells like a fatal-type error message?!?

What do you think?

Thanks,
Paolo.

/////////////////////////