Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/18/84; site calgary.UUCP
Path: utzoo!watmath!clyde!cbosgd!ihnp4!alberta!calgary!radford
From: radford@calgary.UUCP (Radford Neal)
Newsgroups: net.lang.c
Subject: Unrolling string copy loop
Message-ID: <341@calgary.UUCP>
Date: Mon, 1-Apr-85 16:33:46 EST
Article-I.D.: calgary.341
Posted: Mon Apr  1 16:33:46 1985
Date-Received: Fri, 5-Apr-85 02:08:11 EST
References: <1049@gloria.UUCP> <3505@alice.UUCP>
Organization: University of Calgary, Calgary, Alberta
Lines: 55

> 	sym.1:
> 		movb	(r2)+,(r1)+
> 		bneq	sym.1

> By the way, Colonel, this loop is not improved by unrolling.

WRONG! I timed the following two routines:

# String copy with ordinary loop.

_sc1:	.word	0
	movl	4(ap),r1
	movl	8(ap),r2

1:	movb	(r1)+,(r2)+
	bneq	1b

	ret

# String copy with unrolled loop.

_sc2:	.word	0
	movl	4(ap),r1
	movl	8(ap),r2

1:	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	beql	2f
	movb	(r1)+,(r2)+
	bneq	1b

2:	ret

The first takes 120 microseconds to copy a thirty character string. The
second takes only 100 microseconds. 

Seems that branches not taken are faster than branches which are taken.

    Radford Neal
    The University of Calgary