This is a patch implementing builtins for an atomic exchange with full, 
acquire, and release memory barrier semantics.  It is similar to 
__sync_lock_test_and_set(), but the target does not have the option of 
implementing a reduced functionality of only implementing a store of 1.  
Also, unlike __sync_lock_test_and_set(), we have all three memory 
barrier variants.

The compiler will fall back to a full barrier if the user requests an 
acquire/release and it is not available in the target.  Also, if no 
variant is available, we will fall back to a compare and swap loop with 
a full barrier at the end.

The real reason for this patch is to implement atomic stores in the C++ 
runtime library, which can currently incorrectly move prior stores past 
an atomic store, thus invalidating the happens-before promise for the 
sequentially consistent model.  I am attaching the corresponding patch 
to libstdc++ to show how I intend to use the builtin.  This is not an 
official submission for the C++ library bits, as I have not yet fully 
tested the library.  I will do so separately.

In a followup patch I will be implementing acq/rel/full variants for all 
the __sync_* builtins which we can use for the atomic loads and for some 
of the OpenMP atomics Jakub has been working on.

Oh yeah, I would gladly accept patterns/patches for other architectures :).

Tested on x86-64 Linux.

OK for mainline?