FFT benchmark

[Stephen Weeks]

> I  finally  sat  down  and  looked  at  the hot loop you sent me from the FFT
> benchmark.  One inefficiency is the repeated references  to  the  same  stack
> (MLton stack) location. 
...
> I would think that doing this kind of common subexpression elimination  (just
> of  the  result  of  loads) would be a good idea any way.  Especially for the
> extra-indirection stuff (all the loads above).
...


Thanks for the observations.  Both of those should probably go in.

> I tried to look at the generated C code, but nothing I  ran  on  any  of  the
> versions  of the FFT benchmark produced similar C code.  What did you compile
> and with what version of MLton?  Could you send me the source you used?

It's at http://www.sourcelight.com/MLton/benchmarks/fft.sml.  You need to put a
call "val _ = Main.doit()" at the end to make it do something.  I used the
latest internal version of MLton, but 20000712 should work as well.