On Thu, Mar 22, 2012 at 11:58 AM, Paolo Carlini
> On 03/22/2012 03:49 PM, Jason Merrill wrote:
>> It's ill-formed to have a field with incomplete type. The best thing
>> would be to complain about that before we get to literal_type_p so that
>> errorcount is set, if that's not too complicated.
> Agreed. The problem is that if we just change check_field_decls to produce
> an error about the incomplete field, we produce also another later: that is,
Can't we set a bit saying that the field has already gone through
like we do when trying to avoid duplicate warnings?
> considering cp_parser_lambda_expression, we get to check_field_decls from
> finish_struct, but we eventually also produce an error with
> cxx_incomplete_type_diagnostic from build_lambda_object (-> force_rvalue ->
> build_special_member_call -> complete_type_or_maybe_complain)
> Anyway, I also think not calling literal_type_p from check_field_decls if
> the type is incomplete is pretty ugly, but I'm not sure which is the best
> way to make progress: I could try returning a boolean from check_field_decls
> if something goes wrong in order to bail out early from
> cp_parser_lambda_expression (at the moment, finish_struct_1,
> check_bases_and_members, all return void). Or I could try to catch the
> incomplete field even *before* check_field_decls.
> What do you suggest?