Hi Yufeng,
One alternative is to make sure all items are created / destructed in the main thread then handed over to the spawned threads, keeping track of which object is under the domain of which TID. Use reference counting, and post their imminent demise to a reaper list which the main thread acts upon. Any TID that is noted causes the main thread to go through it's outstanding object list to put all those items on the reap list -- note that reference counts may let them stay around for a few cycles until the things referring to them are destructed. Disallow stack based objects in threads.
May not be doable depending on all your developers and the third party code you use.
Or, you could try fixing those issues of pthread_cancel / pthread_exit abuse with QA and apoplectical management.
Or, try bribing the developers with cookies and donuts. If that fails, use thumbscrews.
Or, if you want proper garbage collecting in a multithreaded environment, use Java. (Multithreading is a "bolt on" for C and C++. Unfortunately. It's not going become part of the language proper any time soon.)
--Eljay
