benchmarking

[Matthew Fluet]

> So  only the new SML/NJ is actual competition, and it is beating us in logic,
> nucleic, barnes-hut, and ray.  Do we have any notion of what is  going  wrong
> on these?

nucleic, barnes-hut, and ray are all floating-point intesive.  I'm
currently extending the backend to not flush the floating-point stack at
the end of every basic block (yes, it is that bad); i.e., we'll carry live
float values the same way we are carrying live integral values.  Maybe
that will help.  As best I can make out grepping through the SML/NJ
source, the fast-fp option is off by default, so they may be even
better than it appears.

I don't know about logic.