I know, I know -- why bother optimizing bytecode?  (Compelling reasons
have long been debated -- for me, it's a need to do dynamic- / meta-
programming.)

Anyway, I've written a small patch for the OCaml bytecode compiler and
runtime that significantly improves the performance of a few
floating-point intensive benchmarks (from 30 to 40%) by reducing
boxing and unboxing of intermediate floating point results.

This is still preliminary work, and honestly I'm a little suspicious
that it works so well.  Testing has also been light -- it compiles and
runs my 'benchmark' script when patched against ocaml 3.08.1 and
3.08.4 (as well as MetaOcaml, though the patch needs a little
coaxing).

A short write-up can be found at http://dem.inim.us/ocamlfp/README-FP.html

The patch is found at http://dem.inim.us/ocamlfp/ocamlfp-0.1.diff.gz

I'd greatly appreciate feedback, criticism, testing on other platforms.

Thanks in advance,
Jeff