From: Raymond Toy
Subject: Re: Which implementation is best for floating points?
Date:
Message-ID: <4nafozfqm5.fsf@rtp.ericsson.se>
Raymond Toy <ยทยทยท@rtp.ericsson.se> writes:
>
> In my own tests, I find it easier to get fast floating-point code with
> CMUCL than Allegro. Sometimes the same carefully crafted code is 2 or
> more times faster than Allegro (but sometimes the same speed). This
> holds for both the Linux and Sparc version. I think others get the
> same. This may not be true for non-floating point code, or if you are
> careless about declaring stuff. Allegro GC is seems better, though.
>
Just one other comment. Even though I only had the Unix demo CD,
Franz was VERY helpful in answering my questions on how to get their
compiler to produce better code. The final result with really not too
many tweaks was that Allegro was still a bit slower. However, this
was caused by the fact that Allegro didn't do type propagation for
sqrt and log functions so that it had to use generic functions. In
CMUCL with the later releases, this propagation can be done so that
sqrt and log are either inlined FP instructions (x86) or calls to the
C functions (sparc). The person I talked to said that he is actively
working on making these things better.
Franz sounds like a good place to get a commercially supported Lisp
that is well-supported.
Ray