benchmarks

[Henry Cejtin]

I  agree that all the time in running the smith-normal-form program is in the
gmp IntInf libraries, but here it was the compile time that got much quicker.
Do  you have any idea on why x86-codegen was more than twice as fast?  To see
the effect, just run the smith-normal-form thing on small matricies.  Then it
never goes to the library code (because all the numbers are small).

The  must-not-alias  thing is probably only relevant for mutable objects, but
re `We know that heap objects of different types must not alias', this  isn't
quite  true  in the case of arrays that get cast into vectors or IntInf.int's
that come from (or go to) arrays or vectors.

For my FFT (which is currently 1/2 the speed of the GNU Scientific library) I
would say that the extra level of indirection in arrays of complex's  is  the
major cost.