Re: [PATCH,c++] convert cp_parser.unparsed_function_queues &co. to VECs

[Jason Merrill <jason at redhat dot com>]

On 06/30/2010 03:55 PM, Nathan Froyd wrote:
> > >        /* Friends must all go through the overload machinery,
> > >           even though they may not technically be overloaded.
> > >
> > >           Note that because classes all wind up being top-level
> > >           in their scope, their friend wind up in top-level scope as well.  */
> > >        if (funcdef_flag)
> > >          SET_DECL_FRIEND_CONTEXT (decl, current_class_type);
> >
> > What happens if you remove the if?
>
> If I set funcdef_flag to true at the start of do_friend when f is
> getting defined, things work as expected.  grokdeclarator attempts to
> grok f in FIELD context, which explains why funcdef_flag doesn't get set
> in grokdeclarator.

And it shouldn't get set, because f is not being defined, only declared.

> What's the right fix?  Presumably the test shouldn't be removed in
> do_friend...should grokdeclarator be checking declarator->kind?

Why presumably?  The comment for DECL_FRIEND_CONTEXT doesn't specify 
that it's only set for definitions.

But I would be sure to test that we give the appropriate errors if you 
have a friend declaration followed by a global definition that refers to 
static members of the type without explicit qualification.

Jason