Re: [boost] [intrusive_ptr] atomic ref counting?

From: Peter Dimov (pdimov_at_[hidden])
Date: 2006-04-15 07:58:43

Eric Niebler wrote:
> I'm thinking about replacing some uses of shared_ptr with
> intrusive_ptr, but I've got questions. One of the nice things about
> shared_ptr is the lock-free thread-safe ref counting. I want my
> intrusive_ptr to have the same safety and efficiency, but I don't see
> how to get it. I see $boost_root/detail/atomic_count.hpp, but
> shared_ptr doesn't use it. Why not? And why is it in detail/? Should
> I not consider using it?

shared_ptr doesn't use atomic_count because the addition of weak_ptr has 
made atomic_count's interface insufficient.

atomic_count is in detail:: because I didn't consider it ready for boost::. 
On the other hand, if it works for you, there is no reason to not use it.

> Taking in the bigger picture, intrusive_ptr requires people to write
> their own add_ref and release functions, most of which will probably
> end up using ++ and -- on a plain integer. This is totally broken in
> multi-threaded applications, right?

Using ++ and -- on a plain integer without synchronization is broken in MT 
apps, yes.

> Isn't it dangerous to not provide
> complementary utilities for correctly managing reference counts? It
> seems to me intrusive_ptr is impossible to use safely and portably.
> What have I missed?

It's possible to use intrusive_ptr safely and portably if you already have 
an existing library that provides intrusive reference counting without also 
supplying a corresponding pointer type, and one of the goals of 
boost::intrusive_ptr is to establish a standard "corresponding pointer type" 
for such libraries.

However, you are right that intrusive_ptr doesn't help the implementers of 
these libraries to get their reference counting correct, and that it 
probably should do that.