[MLton] cvs commit: improved analysis for ref flattening

[Matthew Fluet]

>   I am more worried about space safety with this analysis, since it lets
>   a ref's containing object be kept alive quite a bit longer than the
>   old analysis.  This will be worth revisiting at some point.

There is a tangential downside of ref flattening keeping around containing
objects longer than before; you might end up wasting space by both
projecting a (non-ref) component of the tuple and also keeping the tuple
live to the point of use (because a ref component is keeping the object
alive).