Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!purdue!haven!mimsy!chris
From: chris@mimsy.UUCP (Chris Torek)
Newsgroups: comp.lang.c
Subject: loop strength reduction (was Another silly question)
Message-ID: <17692@mimsy.UUCP>
Date: 23 May 89 00:05:39 GMT
References: <17812@cup.portal.com> <607@kl-cs.UUCP> <749@mccc.UUCP> <1677@auspex.auspex.com>
Organization: U of Maryland, Dept. of Computer Science, Coll. Pk., MD 20742
Lines: 50

In article <1677@auspex.auspex.com> guy@auspex.auspex.com (Guy Harris) writes:
>	for (i = 0; i < LEN; i++)
>		a[i] = 0;

>... on most architectures, this requires that the value in "i" be
>multiplied by "sizeof a[0]" before being added to the address
>represented by the address of "a[0]", and do a strength reduction on
>that multiplication; you then find the induction variable not used, and
>eliminate it, and by the time the smoke clears you have the loop in the
>first example generating the same code as the loop in the second
>example.  (I don't know whether there are any compilers that do this or
>not.)

I fed this through gcc (1.35/vax) using -fstrength-reduce and it
produced the following equivalent:

		movl $19,r1
		addl3 $_a,$76,r0
	L4:
		clrl (r0)
		subl2 $4,r0
		decl r1
		jgeq L4
		...

(using `-mgnu' one gets a jsobgeq instead of decl+jgeq).  I am surprised
that, while it inverts the loop counter, it does not use the predecrement
addressing mode---I rather expected

		movl	$19,r1
		moval	_a+80,r0	# or movab; gcc uses movab elsewhere
	L4:	clrl	-(r0)
		jsobgeq	r1,L4

Or, alternatively, post-increment:

		movl	$19,r1
		moval	_a,r0
	L4:	clrl	(r0)+
		jsobgeq	r1,L4

Or (better but less likely) quadword operations:

		movl	$9,r1
		moval	_a,r0
	L4:	clrq	(r0)+
		jsobgeq	r1,L4
-- 
In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 7163)
Domain:	chris@mimsy.umd.edu	Path:	uunet!mimsy!chris