On Fri, Jul 20, 2012 at 11:04 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> On Fri, Jul 20, 2012 at 10:37:32AM +0200, Richard Guenther wrote:
>> On Thu, Jul 19, 2012 at 6:53 PM, Jakub Jelinek <ja...@redhat.com> wrote:
>> Hum. How hard would it be to merge the attributes?
>
> IMHO hard and ugly. The thing is that you probably can do some hacks
> easily in handle_nonnull_attribute, so that multiple nonnull attributes
> on the same prototype get merged together (at the end of function when
> returning without *no_add_attrs = true; before it, do
> a = lookup_attribute ("nonnull", TYPE_ATTRIBUTES (type));
> if (a != NULL)
> {
> merge stuff into a;
> *no_add_attrs = true;
> }
> ), but that handles just one of the cases, where multiple nonnull attributes
> appear on the same prototype. But you can (and with builtins.def usually
> do) have also
> void foo (void *, void *, void *) __attribute__((nonnull (1)));
> void foo (void *, void *, void *) __attribute__((nonnull (2)));
> and for that case no attribute hook is called, so either merge_attributes
> would need to special case this attribute (which would be a layering
> violation, as nonnull is just C/C++ attribute), or each FE would need in its
> merge_decls and similar call lookup_attribute ("nonnull", TYPE_ATTRIBUTES
> (...));
> twice and do the merging manually. As there are just two users of the
> nonnull attribute, handling all of them there looked much shorter and easier
> to me.
In the end it asks for a rewrite of the attribute representation and
handling code ...
(a unified attributes.def file and a non-string-based
non-tree-list-based storage,
etc. ...).
Your patch is ok meanwhile.
Thanks,
Richard.
>> > 2012-07-19 Jakub Jelinek <ja...@redhat.com>
>> >
>> > PR c++/28656
>> > * tree-vrp.c (nonnull_arg_p): Handle all nonnull attributes instead
>> > of just the first one.
>> >
>> > * c-common.c (check_function_nonnull): Handle multiple nonnull
>> > attributes properly.
>> >
>> > * c-c++-common/pr28656.c: New test.
>
> Jakub
