Architecture-specific reference count management underneath shared_ptr is in boost/detail/sp_counted_base*.hpp. Note that it uses gcc where TBB confusingly uses linux (in tbb/machine). From a quick glance at gcc_x86 it seems decent enough?

(Added) A mutex is used in boost if there is no architecture-specific support for atomic access to a variable, at the cost of increasing its size and lowering its performance. TBB atomics are the size of the underlying type and always require a minimal set of supported architecture-specific operations, so it will not (be able to) fall back to using mutexes.