benchmarks

[Matthew Fluet]

> > Strangely, there are some decent speedups in ratio-regions and
> > wc-input1, a 2X speed up in matrix-multiply, and an amazing 9X
> > speedup in md5!
> 
> Truly amazing.  Hopefully we can figure it out and attribute it to
> something other than caching effects.  We should keep in mind that
> another possibility is bugs :-(.

matrix-multiply could be caching; could also be loop alignment, which I
know is less than perfect in the x86-codegen.  On the todo is using loop
header info to do better alignment for the right labels.

md5 is really strange.  I turned on a couple of print statements and the
output is the same under the release and the experimental.