A good starting point would be to go through the BLAS-
these are used quite heavily in numerical work.  I would
also look a bit at the vector features in FORTRAN 90-
shorthands for array sectioning, etc.

I also had an interest a few years ago in doing this...
no time.  But it should be straightforward, and I think
this approach is a good way to go.  The fundamental
thing that was missing was the notion of homogenous
arrays of known, primitive types such as float, double,
and integer ( to corresponed with comparable abilities
in FORTRAN ).  The Vector/Matrix primitives would be
super to have.

I wouldn't expect it to be as fast as FORTRAN, but it
ought to be pretty good- I think they still say that
C is slower, I've heard about 30% in some cases.  So,
there is perhaps some leeway to do better than C, or
C++ for that matter.

So, for what it's worth, I strongly encourage you to
do this- I wish I could offer more, but time is very
short and I really ought to get those Modules working
for everyone.

les