[MLton] Re: cygwoes

[Stephen Weeks]

> I did a little more investigating of the Virtual{Alloc,Free}
> problems over the weekend, and I'm mostly convinced that (at least
> through Cygwin), they are not working as advertised by the API.

Great (sort of).  That seems to be our only problem on Cygwin.  I just
changed MLton to use mmap on Cygwin, cross-built MLton for Cygwin, and
ran all the regressions on directly on Cygwin with the cross-built
MLton.  All regressions passed, except for a case-sensitivity issue in
the file name reported by MLton.Exn.history, which shows up in the
exnHistory and exnHistory3 regressions.  Other than that MLton
performed fine, and in particular had no problem invoking gcc.

I did see the same issue that we saw before with mmap not giving as
much memory as one can get with VirtualAlloc.  On my machine with 1G
RAM, mmap only gave ~500M.  I didn't actually try the VirtualAlloc
test, so I guess it could be some Windows setting preventing Cygwin
from getting enough memory (suggestions for things to try
appreciated), but I'll guess that the issue is due to Cygwin.

Assuming the problem with VirtualAlloc looks unsolvable, I propose
that we move forward with the release with MLton-on-Cygwin using
mmap.