When I did the earlier Fibonacci benchmarking, I found it useful to take the code generated by ac, and run it inside mzscheme. (Note that :a will drop you from Arc to Scheme.) That is, running the exact same code that Arc does, just from the Scheme REPL. (This should take the same time as running in Arc, or else something is very wrong.) Then I could tweak the code a function at a time evolving from the Arc-generated code to the Scheme code, and see where the time was going.
Overall, it makes no sense that Arc would be faster. But it would be very interesting to find out why you're getting those results. Could there be some subtle algorithm difference, maybe lack of tail recursion, in the Scheme version?
Ok, I tried that. Running the fib code in a mzscheme REPL is slower than in an Arc REPL. But, if once in Arc REPL you type :a (and get to the mzscheme REPL with Arc functions loaded), the same Scheme code is then a little faster than the Arc counterpart. That means there is something in ac.scm that speeds up mzscheme computations. Maybe the fact that it's embedded in a module ? I really don't know, but at least this is not an aberration anymore.
