Unreproducible bug

[Stephen Weeks]

> I've started getting some strange compiler errors:
> 
> x_0 has no num occurrences

This is caused by a bug in the Xml shrinker -- which hasn't changed since June
21.

> f_0 is not in scope

This is printed by the CPS typechecker when it gets an invalid program.  It
means some simplification pass screwed up.

> And, these are occuring using the
> SML/NJ version of MLton, not a self-compiled version.  I've gotten it a
> couple of times on the regression suite, but when I rerun it on the same
> file, there's no error.  

I find the irreproducibility very difficult to understand.  There is no
randomness in MLton.  The only thing I can think of (and this is a huge stretch) 
is some kind of state being kept in the filesystem.

Please send me any of the .sml files that are causing a problem.  I don't really
have anything to go on.  I haven't seen these sorts of bugs since June, and they
were always completely reproducible at the time.