| Message ID | 500EAFFC.8030703@oracle.com |
|---|---|
| State | New |
| Headers | show |
On 07/24/2012 10:23 AM, Paolo Carlini wrote: > while looking into the re-opened c++/51213 (doesn't seem trivial: > details soon, I hope) I noticed things which look like old pastos: in > convert_template_argument and tsubst we are computing complain & > tf_error for the - tsubst_flags_t, not bool - fourth argument of > make_typename_type. That doesn't make much sense to me and the below, > as-is, passes testing. The point of this use of (complain & tf_error) is to mask out flags we don't want and leave only tf_error. It may or may not be needed, but it's not a pasto. Jason
On 07/24/2012 08:06 PM, Jason Merrill wrote: > On 07/24/2012 10:23 AM, Paolo Carlini wrote: >> while looking into the re-opened c++/51213 (doesn't seem trivial: >> details soon, I hope) I noticed things which look like old pastos: in >> convert_template_argument and tsubst we are computing complain & >> tf_error for the - tsubst_flags_t, not bool - fourth argument of >> make_typename_type. That doesn't make much sense to me and the below, >> as-is, passes testing. > > The point of this use of (complain & tf_error) is to mask out flags we > don't want and leave only tf_error. It may or may not be needed, but > it's not a pasto. Ok, thanks. Probably loooked to me like a pasto because we are using this kind of clearing in very few places, if at all (if I'm not mistaken). Anyway, in the meanwhile I had a closer look and I think isn't really needed because: 1- make_typename_type and the functions it calls (like perform_or_defer_access_check) only pay attention to tf_error (only make_typename_type itself to tf_keep_type_decl). 2- make_typename_type checks indeed tf_keep_type_decl but calling it with complain & tf_error (as convert_template_argument does) vs plain complain cannot make a difference, because tf_keep_type_decl cannot be "randomly" set in complain, because in the whole front-end we set tf_keep_type_decl *only* in the flags we pass to make_typename_type itself from tsubst. Thus, all in all, I propose doing the change anyway, of course with a different ChangeLog: if what I summarized above is correct the risk of breaking something is really tiny and the code would become a tad simpler. Paolo.
On 07/24/2012 03:05 PM, Paolo Carlini wrote: > Thus, all in all, I propose doing the change anyway, of course with a > different ChangeLog: if what I summarized above is correct the risk of > breaking something is really tiny and the code would become a tad simpler. OK. Jason
Index: pt.c =================================================================== --- pt.c (revision 189808) +++ pt.c (working copy) @@ -6139,7 +6139,7 @@ convert_template_argument (tree parm, orig_arg = make_typename_type (TREE_OPERAND (arg, 0), TREE_OPERAND (arg, 1), typename_type, - complain & tf_error); + complain); arg = orig_arg; is_type = 1; } @@ -11402,7 +11402,7 @@ tsubst (tree t, tree args, tsubst_flags_t complain } f = make_typename_type (ctx, f, typename_type, - (complain & tf_error) | tf_keep_type_decl); + complain | tf_keep_type_decl); if (f == error_mark_node) return f; if (TREE_CODE (f) == TYPE_DECL)