I'm not very proud of this solution, but unless anyone has a better
ideas this is how I plan to add the atomic operations for shared_ptr.

I used __gnu_cxx::__mutex instead of std::mutex because it has fewer
dependencies, so the atomic operations should always be available for
shared_ptr even when the target doesn't support C++11 threads.

The operations are also defined for the __shared_ptr class template.

Tested x86_64-linux.