Need to investigate the best way to call atomic operations on different
platforms. For example, on single-cpu x86, need to study whether we
can remove the lock prefix. Some atomic::membar() calls are guarded by
a os::is_MP() test, some aren't. Better make that consistent.