From: Peter Dimov (pdimov_at_[hidden])
Date: 2002-10-01 12:53:39
From: "David Abrahams" <dave_at_[hidden]>
>
> I don't know how long I'll be able to hold this position in the face of
> move semantics, but so far I've been successful at promoting the idea that
> a "destructible but otherwise unusable" state is misguided for most
> classes. One thing I don't like about this arrangement is that is tends to
> encourage that sort of design. I think it might be avoidable by having the
> user inherit from something like this:
>
> struct intrusive_count
> {
> intrusive_count() : count(new counted_base) {}
> ~intrusive_count() { if (count->weak_count == 0) delete count; }
> counted_base* count;
> };
>
> Just a thought.
The current situation... leaves something to be desired. It is acceptable;
while the object is in a "zombie state" there should be no references left
to it (unless someone has stored a plain pointer/reference) as weak pointers
no longer hand out a shared_ptr, and it's usually pretty obvious what the
dispose() overload should do. Still, it doesn't "just work".
I think I can make a variation of the above work (interestingly the
destructor has to be empty AFAICS) but OTOH having an embedded count
eliminates one heap allocation.
Subject needs further research.