[plt-scheme] Statistics (V301.5 Speed Up)

From: Will M. Farr (farr at MIT.EDU)
Date: Tue Feb 14 13:31:12 EST 2006

On 14 Feb 2006, at 11:20 AM, Alexander Friedman wrote:

>> - generic arithmetic preventing arithmetic optimizations (not that
>> anybody really knows how to fix this)
> Yes. I don't see any obvious way to fix this either.

As someone who uses Scheme to do pretty heavy-duty numerical 
simulations, could I put in a request for some fl+ fl- fl* fl/, etc 
operators?  (These are already in the SRFI with proposed R6RS 
arithmetic.)  I realize that this solution to the generic arithmetic 
problem is not perfect, but if the typechecks and boxing/unboxing 
operations associated with these operators were compiled in the right 
way, a pretty simple peephole optimizer would be able to eliminate most 
of them---for example, the ocaml bytecode interpreter recently gained a 
double-float result register (unboxed), and numerical performance 
improved quite a bit (see http://dem.inim.us/ocamlfp/README-FP.html ).  
Naively, I would expect even greater improvements if there's a proper 
native-code compiler sitting behind the floating-point specific 
operators.  You're not going to catch Bigloo or another compiler that 
does full SUA, etc, but you would make me very happy :).


Posted on the users mailing list.