A first step to getting better virtual memory behavior would be getting a
VM written in Bigloo, Gambit, or Scheme->C. Then you could monkey around
with the function representations to let you use mmap and others of that
ilk to bring code into the image without this misguided (IMHO) compile an
image to C stuff.
The compile-an-image-to-C stuff was supposed to be a quick hack to
cheaply pick up a lot of mileage. Doing a new VM and retargeting the
compiler is a lot of work. Doing per-module mmap loading would be
*fantastic*, but requires new S48 features, like separate compilation.
Kelsey is working on that, so I am (still) hopeful.