pseudo-register addressing

[Matthew Fluet]

> Like I said before: there could be a cache advantage.  If you use a fixed
> block for pseudo-registers then they will stay in the L1 cache.  If you put
> them on the stack and you spend time doing a lot of stack pushes and pops
> (from non-tail calls and returns) then the places you store them in in the
> various stack frames might not.  I have no idea how big a disadvantage this
> could be.

At some point in the not too distant future, I want to play around with 
"Linux x86 Performance-Monitoring Counters Driver" [1] which should let us
measure things like cache misses, and really see if this is an issue.

[1]
http://www.csd.uu.se/~mikpe/linux/perfctr/