Re: [C++ bug] access control for types

["Martin v. Loewis" <martin at loewis dot home dot cs dot tu-berlin dot de>]

> Anyway, even with the explicit friendship, you get the same errors.

Hi Nathan,

I don't: With 20000210, g++ accepts my code (with the friend
declarations) as posted...

>       /* The type and its nested classes are implicitly friends, as
>          per core issue 45 (this is a change from the standard).  */
[...]
> The above code doesn't quite match the comment. The comment implies a
> bi-directional friendship -- the nested class is a friend of its
> container, and the container is a friend of the nested class.

The text of this issue is available at 

http://www.informatik.hu-berlin.de/~loewis/corer8.html#45

(among other places). The proposed resolution reads

# A member class should have access to the members of the enclosing
# class in the same manner as if it had been declared a friend of the
# enclosing class. See paper J16/99-0009 = WG21 N1186.

This doesn't imply biderectional friendship, so I'd say that g++
implements that part correctly. Indeed, is_friend(Berived, Nested)
returns true.

> So, I agree with you that explicit friendship is needed by the
> standard, but it is not needed according to the (intended) behaviour.
> Perhaps a pedwarn is necessary here. It depends on what we mean by
> pedantic. Adherance to the published standard plus any normative
> thingies (I've forgotten the proper name). Or adherence to the
> published standard plus defect reports as they are resolved.

I'd say g++ should consider defect reports according to the likely
resolution. Issue 45 is in drafting state, so it seems that future
corrigendum will most likely support the current implementation.

I didn't know this was an issue, so I earlier took the position that
your code is ill-formed. Now I think it should be accepted. I haven't
analysed this further; the problem appears to be that the protected
access doesn't take this implicit friendship into account. I see that
is_friend(Base, Nested) is being called, which (correctly) returns
false. I haven't found the place where friends get access to protected
members.

Regards,
Martin