Steven Watanabe wrote:
Given the relative "distance" between this initialization code and a possible access of either manager or invoker member in the vtable, I think the problem is of a rather academical nature. But the code is still not strictly correct from my understanding.
I'm afraid that it is a very real problem. In pseudo-code
if(!vtable_initialized) { stored_vtable_initialized = true; stored_vtable.vtable_type(f); } if(stored_vtable.assign_to(f, functor)) vtable = &stored_vtable; else vtable = 0;
Even with a single CPU if the thread is interrupted immidiately after the assignment to stored_vtable_initialized, another thread can reach stored_vtable.assign_to(f, functor) before stored_vtable is initialized.
Right, that would be a legal implementation I just did not consider. This is a little sad since it could probably be easily fixed by a compiler vendor if the standard would cover it (with a proper memory model even for the SMP case). But as it is, this nice little optimization probably has to go. Regards Timmo Stange