reusing pseudo-regs

[Matthew Fluet]

> > Unfortunately, I still haven't worked out how to dynamically adapt
> > enough.
> ...
> You could have the bit-vector functor use an int ref that is the
> vector length, and set that on entry to each SSA function.

That's a decent idea.  I'll think some more about it.

> > Both of those would be nice; I think 1 has the potential to be the bigger
> > win, although I'd be wary of forcing a variable that could live in a
> > psuedo-reg to live in a stack slot just because it is moved to or from
> > there.  (I think you had this in mind with your description.)
> 
> I didn't mean for that to happen, and agree it looks risky.  All I
> meant was to target registers to registers and stack slots to stack
> slots.

O.k.  I think the reg to reg is where shuffle code could be cut down.  The
codegen tries to figure some of this stuff out to reduce shuffle, but the
more explict the better.