Path: utzoo!censor!becker!geac!jtsv16!uunet!samsung!think!leander.think.com!barmar
From: barmar@leander.think.com (Barry Margolin)
Newsgroups: comp.lang.lisp
Subject: Re: Can there be an efficient SWITCH form in CL?
Message-ID: <31749@news.Think.COM>
Date: 21 Nov 89 02:39:44 GMT
References: <32711@ucbvax.BERKELEY.EDU>
Sender: news@Think.COM
Distribution: usa
Organization: Thinking Machines Corporation, Cambridge MA
Lines: 26

In article <32711@ucbvax.BERKELEY.EDU> norvig@cogsci.berkeley.edu (Peter Norvig) writes:
>				      Or should we just insist that
>compiler writers produce better code for (CASE n ...), and for
>(aref (vector ...) n)?

After racking our brains a couple of years ago over this one, we complained
to Symbolics.  As a result, the compiler in Genera 7.2 optimizes CASE
macros when all the keys are non-negative fixnums and the space of keys
isn't too sparse (if memory serves, "too sparse" is when fewer than half of
the keys between 0 and the largest key are used).  Earlier this year I sent
them the minor modifications necessary to remove the reference to 0.

By the way, depending on the memory architecture of the machine, converting
a series of tests into a jump through a dispatch table might *not* be an
optimization.  Some systems simultaneously prefetch both branches of a
conditional jump, but this can't be done easily with a dispatch.  You can
often get better overall performance by putting the most common cases
first.  This is why such optimization should be left to the compiler
(although it's still a hard problem, since different models of Symbolics
36xx's have different prefetch and caching characteristics, yet there is
only one compiler for the whole family).

Barry Margolin, Thinking Machines Corp.

barmar@think.com
{uunet,harvard}!think!barmar