Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watnot!watmath!clyde!cbatt!decuac!cvl!mimsy!chris
From: chris@mimsy.UUCP
Newsgroups: comp.arch,comp.lang.c
Subject: Re: String Handling -- Incompetence of run-time libraries
Message-ID: <6071@mimsy.UUCP>
Date: Wed, 1-Apr-87 13:45:05 EST
Article-I.D.: mimsy.6071
Posted: Wed Apr  1 13:45:05 1987
Date-Received: Sat, 4-Apr-87 07:32:31 EST
References: <15292@amdcad.UUCP> <978@ames.UUCP> <15694@sun.uucp> <9996@sri-spam.istc.sri.com>
Organization: U of Maryland, Dept. of Computer Science, Coll. Pk., MD 20742
Lines: 81
Keywords: instruction set architectures, strcpy
Xref: utgpu comp.arch:745 comp.lang.c:1421

[To comp.arch readers who just want to read about architectures:
skip this.  It is just correcting various points in other articles,
none of which had much to do with architecture in the first place.]

In some article, someone adds these arrows and notes:
>>$strcpy2(ss1,ss2)
>>$char *ss1,*ss2;<---------------------------- put "register" on this line
>>${	register char *s1,*s2;<-------------- as many compilers will use this
>>$					      as local variables overriding
>>$	s1 = ss1;			      the arguments.
	[etc]

In article <9996@sri-spam.istc.sri.com> robert@sri-spam.istc.sri.com
(Robert Allen) writes:
>To further speed this up you could write it as:
>	strcpy2(ss1,ss2)
>	register char *ss1, *ss2;
>	{
>	    while (*s1++ = *s2++)
>	    ;
>	}
>thus eliminating two assignments and two local variables.

Now these people all seem to be talking about Vaxen; I feel I must
point out the various errors here.  It is clear that neither the
original someone nor Robert Allen actually compiled these to
assembly.  (Use `cc -O -S' to do this.)

First, the notes attached to the arrows are wrong.  Compilers cannot
override the arguments when the names differ.  The loop copies from
s2 to s1; the parameters were ss2 and ss1.  Second, the code produced
for the second example (once corrected to `while (*ss1++ = *ss2++)')
is *identical* to the code for the first!  (Vax running 32V, 3BSD,
4BSD, and, no doubt, Sys3, Sys5, V8, and V9.)  The two assignments
the second version attempts to eliminate must still occur, for
`ss1' and `ss2' are passed on the stack, and must be copied into
the two registers---the same two registers allocated in the first
version.

The proper way to speed strcpy() on a MicroVAX-II is no doubt to
use the following assembly code:

	_strcpy:.globl	_strcpy
		.word	0		# save no registers
		movq	4(ap),r1	# get s1 and s2 into r1 and r2
		movl	r1,r0		# save s1
	1:	movb	(r2)+,(r1)+	# *s1++ = *s2++
		bneq	1b		# loop until a zero is moved
		ret			# return original s1 in r0

Note that this is remarkably similar to the compiler's output
for the original code, modified to have the proper return value:

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

		while ((*s1++ = *s2++) != 0)
			/* void */;
		return (ss1);		/* must return the original value */
	}

		.globl	_strcpy
	_strcpy:
		.word	0xc00		# save r11 and r10
		movl	4(ap),r11	# here is s1
		movl	8(ap),r10	# and s2
	L16:	movb	(r10)+,(r11)+	# *s1++ = *s2++
		jneq	L16		# (this assembles to a `bneq')
		movl	4(ap),r0	# return the original s1
		ret

All one can improve on the locc-poor MicroVAX-II is the register
usage and the parameter grabbing.  (c2, at least from 32V to 4.3BSD,
will never turn two `movl's into a `movq'.  Ah well.)
-- 
In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 7690)
UUCP:	seismo!mimsy!chris	ARPA/CSNet:	chris@mimsy.umd.edu