> I forgot one thing: This comment is only correct for 32 bit code. 64 bit
> code has a uniform array layout for all element types.

Yes, but a generic access to a float array still involves an extra
boxing operation.  So, your advice is sound for 64-bit plaforms as well.

I had a very quick look at the original poster's code: it seems
reasonable performance-wise.  There are a few extra hacks that could
be done to increase performance at some expense in code clarity, for
instance turning

            a.(i).(j) <- f i j a.(i).(j);
in
            (let b = a.(i) in b.(j) <- f i j b.(j));

but I wouldn't recommend doing this unless profiling exhibits a big
"hot spot" in this function.

- Xavier Leroy