Switch clang over to using fence/atomicrmw/cmpxchg instead of the intrinsics...

Switch clang over to using fence/atomicrmw/cmpxchg instead of the intrinsics (which will go away).  LLVM CodeGen does almost exactly the same thing with these and the old intrinsics, so I'm reasonably confident this will not break anything.

There are still a few issues which need to be resolved with code generation for atomic load and store, so I'm not converting the places which need those for now.

I'm not entirely sure what to do about __builtin_llvm_memory_barrier: the fence instruction doesn't expose all the possibilities which can be expressed by __builtin_llvm_memory_barrier.  I would appreciate hearing from anyone who is using this intrinsic.

llvm-svn: 139216