[MLton] inferring getAndSet via an Atomic{Begin, End} optimization

[Henry Cejtin]

I remember you (Stephen) talking about how it used to be that Atomic{Begin,End}
pairs that weren't needed (because there were no limit checks between them)
were previously removed, but that it had been a real problem (although I don't
remember why).

Other than that, it sounds like a good move.  It would mean that all cases
that should be really cheap would be automatically.
Mind you, it also means that it wouldn't be clear from the source code when
things were cheap and when they weren't.