Hi, I'm new to boost, so this post might seem silly to some of you, and I apologize if this is the case. I have a tiny performance suggestion: The c'tor template<class Y> shared_ptr(shared_ptr<Y> const & r, detail::dynamic_cast_tag): px(dynamic_cast<element_type *>(r.px)), pn(r.pn) { if(px == 0) // need to allocate new counter -- the cast failed { pn = detail::shared_count(); } } should be changed to: template<class Y> shared_ptr(shared_ptr<Y> const & r, detail::dynamic_cast_tag): px(dynamic_cast<element_type *>(r.px)), pn(px ? r.pn : detail::shared_count()) { } The difference is when the dynamic_cast fails and returns NULL. In such cases the existing code copies the shared_count object, including mutex locking, only to discover later that it was unnecessary (the 'if(px == 0)' statement). The suggested code saves this copy of the shared_count object, which also saves some mutex lockings. Hope I'm not talking nonsense, Yuval