Le 08/09/12 22:15, Oliver Kowalke a écrit :
Am 08.09.2012 21:57, schrieb Vicente J. Botet Escriba:
The suggested Allocator is just there to responds to your need (predictability) :)
Vicente - I looked at packaged_task as you suggested. I found following code:
packaged_task() : { typedef typename Allocator::template rebind<detail::task_object<R,FR>
::other A2; A2 a2(a); // allocated on the stack, after return from ctor destroyed typedef thread_detail::allocator_destructor<A2> D; task = task_ptr(::new(a2.allocate(1)) detail::task_object<R,FR>(f), D(a2, 1) ); // a2 passed to allocator_destructor }
template <class _Alloc> class allocator_destructor { ... _Alloc& alloc_; // reference to A2 size_type s_; public: allocator_destructor(_Alloc& a, size_type s)BOOST_NOEXCEPT : alloc_(a), s_(s) {} ... };
allocator_destructor hold only a reference to the allocator A2 but the instance was allocated on the stack. Doesn't this mean that the alloc_ member of allocator_destructor will point to a non-existing object?
Yes, you are right. The code is using something not very clean that works only for stateless allocators. I guess that storing the allocator on allocator_destructor should solve the issue. Please could you create a ticket to track the issue? Best, Vicente