These are available since GCC 4.7 and will eventually replace the __sync
operations. They support the memory model defined by C++11. For instance,
by using __ATOMIC_RELAXED for some operations on the reference counters we
can avoid memory barriers, which are required by __sync operations (whose
memory model essentially is __ATOMIC_SEQ_CST).
On many architectures it is safe to read the value directly (those
using cache coherency protocols, and with atomic loads for 32-bit
values) but it is not if that's not the case or if we ever decide to
make refcount_t 64-bit (load not atomic on x86).
So make sure the operation is actually atomic and that users do not
have to care about the size of refcount_t.
The cast to a bool* cut of the actual value on big-endian systems
if bool was shorter than int because the bool argument to printf gets
promoted to an int.
Fixes#479.