x86: Improve spinlocks using ISO C++11 intrinsic atomics

Make some improvements to the x86 spinlock implementation using the ISO
C++11 intrinsic atomic wrappers.  Note that since the spinlocks use
cmpxchg(), xadd() and __add(), it already makes some use of this.

 (1) Use acquire variants in spin_lock() and spin_trylock() paths.  On x86
     this shouldn't actually make a difference.

 (2) Use try_cmpxchg*() rather than cmpxchg*().  This should produce
     slightly more efficient code as a comparison can be discarded in
     favour of using the value in the Z flag.

Note as mentioned before, the LOCK prefix is currently mandatory, but a gcc
ticket is open to try and get these listed for suppression.

Signed-off-by: David Howells <dhowells@redhat.com>