Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watnot!watmath!clyde!rutgers!husc6!endor!reiter
From: reiter@endor.UUCP
Newsgroups: comp.arch,comp.lang.c
Subject: Re: String Handling -- Incompetence of run-time libraries
Message-ID: <1530@husc6.UUCP>
Date: Mon, 30-Mar-87 17:49:39 EST
Article-I.D.: husc6.1530
Posted: Mon Mar 30 17:49:39 1987
Date-Received: Wed, 1-Apr-87 01:36:22 EST
References: <15292@amdcad.UUCP> <978@ames.UUCP> <15694@sun.uucp> <18036@ucbvax.BERKELEY.EDU> <1944@hoptoad.uucp>
Sender: news@husc6.UUCP
Reply-To: reiter@harvard.UUCP (Ehud Reiter)
Organization: Aiken Computation Lab Harvard, Cambridge, MA
Lines: 65
Keywords: instruction set architectures, strcpy
Xref: utgpu comp.arch:711 comp.lang.c:1386

In article <1944@hoptoad.uucp> gnu@hoptoad.uucp (John Gilmore) pointed out
that printf could be recoded to run much faster.

This made me curious, and I started playing around with alternate versions of
strcpy.  I did some timings on a microVAX II (cc -O, 4.3BSD) using the library
strcpy, a straightforward version of my own (strcpy2), and a version which
tried to minimize branches out of the instruction sequence (strcpy3).  I ran
strcpy2 and strcpy3 both as procedure calls and as in-line expansions.  The
timings, for 1,000,000 executions of copying "Test string 1", were

routine	  inline?	time (sec)

strcpy	  no		134.7
strcpy2   no		 65.2
strcpy3   no		 51.6
strcpy2   yes		 43.3
strcpy3   yes		 30.8

The most shocking thing about the above is that the library strcpy is half the
speed of a very straightforward C implementation of the routine!  Perhaps
we should spend less time arguing about hardware support for string-handling
routines, and more time making sure that the people who implement the
run-time library are somewhat competent.  It also certainly appears that
in-line expansion of simple library routines is a big win.

I also ran some tests on a SUN-3/180 (SUN UNIX 4.2, cc -O).  The results were:

routine	  inline?	time (sec)

strcpy	  no		 22.0
strcpy2	  no		 21.1
strcpy3	  no		 20.1
strcpy2	  yes		 15.1
strcpy3   yes		 13.7

The SUN library routines seem much more competently coded, but even here,
perhaps a bit of tuning would help (note that my timing was pretty sloppy,
so small differences may not be significant), and in-line expansion would
certainly be a big win.

Routine codes:

strcpy2(ss1,ss2)
char *ss1,*ss2;
{	register char *s1,*s2;

	s1 = ss1;
	s2 = ss2;
	while (*s1++ = *s2++);	}

strcpy3(ss1,ss2)
char *ss1,*ss2;
{	register char *s1,*s2;

	s1 = ss1;
	s2 = ss2;
	while (*s1++ = *s2++)
	{	if (!(*s1++ = *s2++)) break;
		if (!(*s1++ = *s2++)) break;
	< the above statement is repeated 20 times >
			}	}

***************************************************************************
				Ehud Reiter
				reiter@harvard	(ARPA, BITNET, UUCP)