Re: locked reads for atomics

Here's a v2 of the patch set in which I've attempted to address all
feedback. I've also added a pg_write_membarrier_u* pair of functions that
provide an easy way to write to an atomic variable with full barrier
semantics. In the generic implementation, these are just aliases for an
atomic exchange.

0002 demonstrates how these functions might be used to eliminate the
arch_lck spinlock, which is only ever used for one boolean variable. My
hope is that the membarrier functions make eliminating spinlocks for
non-performance-sensitive code easy to reason about.

(We might be able to use a pg_atomic_flag instead for 0002, but that code
seems intended for a slightly different use-case and has more complicated
barrier semantics.)

--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com