From mboxrd@z Thu Jan  1 00:00:00 1970
Received: (from weis@localhost) by pauillac.inria.fr (8.7.6/8.7.3) id TAA16569 for caml-red; Thu, 14 Dec 2000 19:30:15 +0100 (MET)
Received: from concorde.inria.fr (concorde.inria.fr [192.93.2.39]) by pauillac.inria.fr (8.7.6/8.7.3) with ESMTP id JAA26332 for <caml-list@pauillac.inria.fr>; Thu, 14 Dec 2000 09:03:46 +0100 (MET)
Received: from dynabook ([133.11.12.224])
	by concorde.inria.fr (8.11.1/8.10.0) with ESMTP id eBE83iL02851
	for <caml-list@inria.fr>; Thu, 14 Dec 2000 09:03:45 +0100 (MET)
Received: from dynabook
	([127.0.0.1] helo=localhost ident=sumii)
	by dynabook with esmtp (Exim 3.12 #1 (Debian))
	id 146TLZ-0000BC-00; Thu, 14 Dec 2000 17:02:37 +0900
To: checker@d6.com
From: eijiro_sumii@anet.ne.jp
Cc: caml-list@inria.fr, gerd@gerd-stolpmann.de, sumii@venus.is.s.u-tokyo.ac.jp
Subject: Re: substring match like "strstr"
In-Reply-To: <4.3.2.7.2.20001213223322.00cb3c90@shell16.ba.best.com>
References: <4.3.2.7.2.20001211103237.00c12100@shell16.ba.best.com>
	<20001214123635H.sumii@yl.is.s.u-tokyo.ac.jp>
	<4.3.2.7.2.20001213223322.00cb3c90@shell16.ba.best.com>
X-Mailer: Mew version 1.94.2 on Emacs 20.7 / Mule 4.1 (AOI)
Mime-Version: 1.0
Content-Type: Text/Plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-Id: <20001214170237M.sumii@yl.is.s.u-tokyo.ac.jp>
Date: Thu, 14 Dec 2000 17:02:37 +0900
X-Dispatcher: imput version 991025(IM133)
Sender: weis@pauillac.inria.fr

> Is that C the simple C one I posted, or the libc one on sparc?

It's the one in libc on Solaris, so may be tuned by hand in assembly.

> Your new strstr is actually slower on x86 than your first one, but I
> haven't looked into why yet.

That's interesting.  I confirmed it myself on my x86 machine with my
program:

	strstr_imp2	88.530
	strstr_fun	124.210
	strstr_fun2	127.120
	glibc strstr	61.310

Seeing such strong machine dependency, it seems yet more reasonable to
use the "default" strstr provided by the system (unless the _pattern_
is long enough, in which case KMP or BM might perform better), even
though there is some overhead --- about 5% in my profiling on SPARC
--- of interfacing a C function to OCaml.  This is a natural (though
boring) conclusion, I think.

Eijiro