local refs

[Stephen Weeks]

> (devil's advocate:) if it is o.k. to turn any handler at a non-tail call
> into a "push right before a nontail call and popp right after the return
> or handle", then why bother maintaining HandlerPush/Pop at all.

That is certainly one strategy we will explore.  I just don't see any
alternative for implementing the SSA algorithm.  If we're really
worried, we can measure how many of the bogus pushes and pops we
introduce.

> I'm not commenting on the complication of the transformation, I'm talking
> about how significantly the program deviates from the original.

Understood (I wanted to make sure). 

> Compile times still disappointing; run times unchanged.

I don't see anything noticeably bad compile-time wise, and the
runtimes look on the whole OK.  Code sizes are even a percent or two
better.  It's already at the point where I would vote for leaving it
on.

Obviously we need to add one benchmark to our tests that uses an int
ref for looping.