Larry Evans wrote:
But if the problematic roots are known in advance, why doesn't the programmer simply use weak_ptr's to them? I guess it's because the programmer doesn't know he's connecting to a problematic root, although he knows what they are.
The general scenario is that you have a class X that contains a shared_ptr<Y>, where the Y's can also contain shared_ptr<Y>s to each other, but the user never sees an Y directly. You know that all your roots are in the instances of X, but you can't use weak_ptr to link the Y's because this won't keep a cyclic structure alive until the last X to it is gone. However you can reset_and_collect the shared_ptr<Y> in ~X. (The problem that prevented Eric from using it is that reset_and_collect may throw.) Alternatively, you can also keep a {vector, set}< weak_ptr<Y> > global root structure, add to it in X::X and scan and prune it periodically.