once, local-refs, and threads

[Stephen Weeks]

> Another, conservative, definition of a ThreadCopyPoint is any point in the
> function where the primitive Thread_copyCurrent is invoked or any non-tail
> call.  (Looking at Steve's recent check-in, I think this is the kind of
> approximation made in the App case of globalize in closure conversion.)

Yes.

> This could hurt you a little, especially if there is a clear division
> between start-up code (say one set of functions and the "first-part" of
> main) and then threaded-code (say, another set of functions and the rest
> of main).  But, it is certainly easier to compute.

Agreed.  I would go for a better analysis on SSA.  I only chose that
simple "analysis" because it's on SXML, where we don't like to do
complicated analyses.