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

[Stephen Weeks]

I was thinking about the suggestion to add a getAndSet primitive to
MLton and came up with the following idea. 

Atomic{Begin,End} are MLton primitives that are implemented in the
ssaToRssa pass.  Limit-check insertion runs near the beginning of the
Rssa optimizer right after this pass.  Rather than add a primitive
that gets implemented somewhere after limit-check insertion, why not
leave Atomic{Begin,End} around as primitives until after limit-check
insertion, and have an optimization that runs immediately after
limit-limit check insertion and removes Atomic{Begin,End} pairs that
are unnecessary, e.g. because they are in the same basic block.  Then,
immediately after this optimization, one can implement
Atomic{Begin,End} in terms of lower level memory ops.

This seems like a win because it applies to getAndSet as well as other
situations.  It also means that we don't have to add a new primitive
figure out how it interacts with all MLton's optimizations.

Thoughts?