Robert Ramey wrote:

Gennadiy Rozental wrote:

...

This is about how I expect it to work. But from this description it's unclear:

...

1. Why would you need to save counter into archive? After all value 10 has nothing to do with resulting 5. 2. Why would you need direct access to shared_ptr counter? After all using regular shared_ptr copy would automatically bump it to proper value.

I remember my initial implementation just restored the original state. Thinking better of it I made a change to the current system. Now that you mention it, it wouldn't seem to need the counter anymore. Just for laughs I commented out the serialization of the counter and if failed to build the test. I'll have to look into it.

Seems promising! Thinking about smart pointer serialization and how to generalize it, ie. to allow serialization of any "my_ref_count_ptr", I got this idea. -How about extending the tracking system with a case something like: serialization::track_ref_count? (perhaps using a is_smart_ptr<> traits function). Based on this tag the serializer would at load time maintain a map<pointer_value, my_ref_count_ptr<T> > to enable the ref-copy "bumping" of all the subsequent my_ref_count_ptrs. Where pointer_value is used as key and obtained by boost::get_pointer (assumed to be specialized for my_ref_ptr). This is essentially the same approach I outlined earlier in my reviews and seems similar to both what Peter Dimov uses and Gennadiy mentions above. It would of course require extensions to the code but I think it would possible to use/adapt the existing pointer code-path to some degree. Hope this sounds reasonable // Fredrik Blomqvist