CWS paper

[Matthew Fluet]

> Here are the numbers using the native backend.  They show that a lot of the
> weirdness was caused by using C.  Now, both is always as good as call with the
> exception of tensor.  And both is always as good as either with the exception of 
> knuth-bendix and tensor.

I suspect that part of the issue with tensor is just the native backend
struggling with floating point.  You'll notice that the native backend is
slower than the C backend on call, cont, and both.  I suspect the speedup
in none arises from the absent trampolining costs.

>                    none  call  cont  both
> tensor              5.5   5.2   5.0   5.3 (native)
> tensor              6.0   4.0   4.2   4.2 (C)