local refs

[Stephen Weeks]

> Out of 1101 top level SSA functions, 126 has flattenable refs.
...
Interestingly, we occasionally get things like:
> x_208258 LocalRefs: [plists_92: {reff = Some ((L_114603, list_85)),
> 				  assigns = [L_114600, L_114569],
> 				  derefs = [L_114601, L_114569],
> 				  locall = local,
> 				  threadCopyCurrent = {assign = false,
> 						       deref = false}}]
> x_208258 Violations: (plists_92)
> 
> which looks like something to do with property-lists.

Yeah, this is cool.  It is from a property that has the "destroy"
capability.  plists is a list ref maintaining the list of property
lists on which the property has been stored.  This is quite possibly
the extract function in ssa/shrink.fun, which is small enough that one
could believe that all the uses of the property are in a single
control-flow graph.

> This one is cute:
...
> It's basically turning nested List.foreach and Vector.foreach into the
> appropriate folds:

Neat.

> If you've made it this far, here's a little more good news.  I tried
> replacing the once pass in constantPropagation, by replacing
>       val once = Once.once program
> with
>       val {varIsMultiDefed, ...} = Multi.multi program
>       val once = not o varIsMultiDefed
> 
> No problems on any of the regressions.  I don't know if it's really doing
> anything different.

The old once pass assumed all functions other than main were
multiUsed, so there is the possibility that Multi.multi is doing
something different.  I vote for removing the old once pass, replacing
the use with this, and renaming once.sml to multi.sml.