shrinker checked in

[Matthew Fluet]

> Playing around with mlprof, I think there is at least one problem with the
> flattener.  Check out tryrec_0 and rew_loop_0 under CPS and SSA
> optimizations.
> 
> More clues: in CPS, the Failure exception carries string, while in SSA it
> carries (string, unit ref).  So, something in useless (that's the only
> pass that might figure out that the unit ref is the same everywhere?).

Alright, working backwards;  something is up with constant propagation,
which isn't globalizing all of the Failure_0 allocations, so useless isn't
dropping the unit ref.