xref: /illumos-gate/usr/src/lib/libc/i386/gen/strncat.S (revision 55fea89dcaa64928bed4327112404dcb3e07b79f)
1*5d9d9091SRichard Lowe/*
2*5d9d9091SRichard Lowe * CDDL HEADER START
3*5d9d9091SRichard Lowe *
4*5d9d9091SRichard Lowe * The contents of this file are subject to the terms of the
5*5d9d9091SRichard Lowe * Common Development and Distribution License (the "License").
6*5d9d9091SRichard Lowe * You may not use this file except in compliance with the License.
7*5d9d9091SRichard Lowe *
8*5d9d9091SRichard Lowe * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9*5d9d9091SRichard Lowe * or http://www.opensolaris.org/os/licensing.
10*5d9d9091SRichard Lowe * See the License for the specific language governing permissions
11*5d9d9091SRichard Lowe * and limitations under the License.
12*5d9d9091SRichard Lowe *
13*5d9d9091SRichard Lowe * When distributing Covered Code, include this CDDL HEADER in each
14*5d9d9091SRichard Lowe * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15*5d9d9091SRichard Lowe * If applicable, add the following below this CDDL HEADER, with the
16*5d9d9091SRichard Lowe * fields enclosed by brackets "[]" replaced with your own identifying
17*5d9d9091SRichard Lowe * information: Portions Copyright [yyyy] [name of copyright owner]
18*5d9d9091SRichard Lowe *
19*5d9d9091SRichard Lowe * CDDL HEADER END
20*5d9d9091SRichard Lowe */
21*5d9d9091SRichard Lowe/*
22*5d9d9091SRichard Lowe * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
23*5d9d9091SRichard Lowe * Use is subject to license terms.
24*5d9d9091SRichard Lowe */
25*5d9d9091SRichard Lowe
26*5d9d9091SRichard Lowe	.file	"strncat.s"
27*5d9d9091SRichard Lowe
28*5d9d9091SRichard Lowe/
29*5d9d9091SRichard Lowe/ strncat(s1, s2, n)
30*5d9d9091SRichard Lowe/
31*5d9d9091SRichard Lowe/ Concatenates s2 on the end of s1.  s1's space must be large enough.
32*5d9d9091SRichard Lowe/ At most n characters are moved.
33*5d9d9091SRichard Lowe/ Returns s1.
34*5d9d9091SRichard Lowe/
35*5d9d9091SRichard Lowe/ Fast assembly language version of the following C-program strncat
36*5d9d9091SRichard Lowe/ which represents the `standard' for the C-library.
37*5d9d9091SRichard Lowe/
38*5d9d9091SRichard Lowe/	char *
39*5d9d9091SRichard Lowe/	strncat(char *s1, const char *s2, size_t n)
40*5d9d9091SRichard Lowe/	{
41*5d9d9091SRichard Lowe/		char	*os1 = s1;
42*5d9d9091SRichard Lowe/
43*5d9d9091SRichard Lowe/		n++;
44*5d9d9091SRichard Lowe/		while (*s1++)
45*5d9d9091SRichard Lowe/			;
46*5d9d9091SRichard Lowe/		--s1;
47*5d9d9091SRichard Lowe/		while (*s1++ = *s2++)
48*5d9d9091SRichard Lowe/			if (--n == 0) {
49*5d9d9091SRichard Lowe/				s1[-1] = '\0';
50*5d9d9091SRichard Lowe/				break;
51*5d9d9091SRichard Lowe/			}
52*5d9d9091SRichard Lowe/		return (os1);
53*5d9d9091SRichard Lowe/	}
54*5d9d9091SRichard Lowe/
55*5d9d9091SRichard Lowe/ In this assembly language version, the following expression is used
56*5d9d9091SRichard Lowe/ to check if a 32-bit word data contains a null byte or not:
57*5d9d9091SRichard Lowe/	(((A & 0x7f7f7f7f) + 0x7f7f7f7f) | A) & 0x80808080
58*5d9d9091SRichard Lowe/ If the above expression geneates a value other than 0x80808080,
59*5d9d9091SRichard Lowe/ that means the 32-bit word data contains a null byte.
60*5d9d9091SRichard Lowe/
61*5d9d9091SRichard Lowe
62*5d9d9091SRichard Lowe#include "SYS.h"
63*5d9d9091SRichard Lowe
64*5d9d9091SRichard Lowe	ENTRY(strncat)
65*5d9d9091SRichard Lowe	pushl	%edi			/ save register variables
66*5d9d9091SRichard Lowe	pushl	%esi
67*5d9d9091SRichard Lowe	movl	12(%esp), %edi		/ %edi = destination string address
68*5d9d9091SRichard Lowe	testl	$3, %edi		/ if %edi not word aligned
69*5d9d9091SRichard Lowe	jnz	.L1			/ goto .L1
70*5d9d9091SRichard Lowe	.align	4
71*5d9d9091SRichard Lowe.L2:
72*5d9d9091SRichard Lowe	movl	(%edi), %edx		/ move 1 word from (%edi) to %edx
73*5d9d9091SRichard Lowe	movl	$0x7f7f7f7f, %ecx
74*5d9d9091SRichard Lowe	andl	%edx, %ecx		/ %ecx = %edx & 0x7f7f7f7f
75*5d9d9091SRichard Lowe	addl	$4, %edi		/ next word
76*5d9d9091SRichard Lowe	addl	$0x7f7f7f7f, %ecx	/ %ecx += 0x7f7f7f7f
77*5d9d9091SRichard Lowe	orl	%edx, %ecx		/ %ecx |= %edx
78*5d9d9091SRichard Lowe	andl	$0x80808080, %ecx	/ %ecx &= 0x80808080
79*5d9d9091SRichard Lowe	cmpl	$0x80808080, %ecx	/ if no null byte in this word
80*5d9d9091SRichard Lowe	je	.L2			/ goto .L2
81*5d9d9091SRichard Lowe	subl	$4, %edi		/ post-incremented
82*5d9d9091SRichard Lowe.L1:
83*5d9d9091SRichard Lowe	cmpb	$0, (%edi)		/ if a byte in (%edi) is null
84*5d9d9091SRichard Lowe	je	.L3			/ goto .L3
85*5d9d9091SRichard Lowe	incl	%edi			/ next byte
86*5d9d9091SRichard Lowe	testl	$3, %edi		/ if %edi not word aligned
87*5d9d9091SRichard Lowe	jnz	.L1			/ goto .L1
88*5d9d9091SRichard Lowe	jmp	.L2			/ goto .L2 (%edi word aligned)
89*5d9d9091SRichard Lowe	.align	4
90*5d9d9091SRichard Lowe.L3:
91*5d9d9091SRichard Lowe	/ %edi points to a null byte in destination string
92*5d9d9091SRichard Lowe	movl	16(%esp), %eax		/ %eax = source string address
93*5d9d9091SRichard Lowe	movl	20(%esp), %esi		/ %esi = number of bytes
94*5d9d9091SRichard Lowe
95*5d9d9091SRichard Lowe	testl	$3, %eax		/ if %eax not word aligned
96*5d9d9091SRichard Lowe	jnz	.L4			/ goto .L4
97*5d9d9091SRichard Lowe	cmpl	$4, %esi		/ if number of bytes < 4
98*5d9d9091SRichard Lowe	jb	.L7			/ goto .L7
99*5d9d9091SRichard Lowe	.align	4
100*5d9d9091SRichard Lowe.L5:
101*5d9d9091SRichard Lowe	movl	(%eax), %edx		/ move 1 word from (%eax) to %edx
102*5d9d9091SRichard Lowe	movl	$0x7f7f7f7f, %ecx
103*5d9d9091SRichard Lowe	andl	%edx, %ecx		/ %ecx = %edx & 0x7f7f7f7f
104*5d9d9091SRichard Lowe	addl	$4, %eax		/ next word
105*5d9d9091SRichard Lowe	addl	$0x7f7f7f7f, %ecx	/ %ecx += 0x7f7f7f7f
106*5d9d9091SRichard Lowe	orl	%edx, %ecx		/ %ecx |= %edx
107*5d9d9091SRichard Lowe	andl	$0x80808080, %ecx	/ %ecx &= 0x80808080
108*5d9d9091SRichard Lowe	cmpl	$0x80808080, %ecx	/ if null byte in this word
109*5d9d9091SRichard Lowe	jne	.L6			/ goto .L6
110*5d9d9091SRichard Lowe	movl	%edx, (%edi)		/ copy this word to (%edi)
111*5d9d9091SRichard Lowe	subl	$4, %esi		/ decrement number of bytes by 4
112*5d9d9091SRichard Lowe	addl	$4, %edi		/ next word
113*5d9d9091SRichard Lowe	cmpl	$4, %esi		/ if number of bytes >= 4
114*5d9d9091SRichard Lowe	jae	.L5			/ goto .L5
115*5d9d9091SRichard Lowe	jmp	.L7			/ goto .L7
116*5d9d9091SRichard Lowe.L6:
117*5d9d9091SRichard Lowe	subl	$4, %eax		/ post-incremented
118*5d9d9091SRichard Lowe	.align	4
119*5d9d9091SRichard Lowe.L7:
120*5d9d9091SRichard Lowe	/ number of bytes < 4  or  a null byte found in the word
121*5d9d9091SRichard Lowe	cmpl	$0, %esi		/ if number of bytes == 0
122*5d9d9091SRichard Lowe	jz	.L8			/ goto .L8 (finished)
123*5d9d9091SRichard Lowe	movb	(%eax), %dl		/ %dl = a byte in (%eax)
124*5d9d9091SRichard Lowe	decl	%esi			/ decrement number of bytes by 1
125*5d9d9091SRichard Lowe	movb	%dl, (%edi)		/ copy %dl to (%edi)
126*5d9d9091SRichard Lowe	incl	%eax			/ next byte
127*5d9d9091SRichard Lowe	incl	%edi			/ next byte
128*5d9d9091SRichard Lowe	cmpb	$0, %dl			/ compare %dl with a null byte
129*5d9d9091SRichard Lowe	je	.L9			/ if %dl is a null, goto .L9
130*5d9d9091SRichard Lowe	jmp	.L7			/ goto .L7
131*5d9d9091SRichard Lowe	.align	4
132*5d9d9091SRichard Lowe
133*5d9d9091SRichard Lowe.L4:
134*5d9d9091SRichard Lowe	/ %eax not aligned
135*5d9d9091SRichard Lowe	cmpl	$0, %esi		/ if number of bytes == 0
136*5d9d9091SRichard Lowe	jz	.L8			/ goto .L8 (finished)
137*5d9d9091SRichard Lowe	movb	(%eax), %dl		/ %dl = a byte in (%eax)
138*5d9d9091SRichard Lowe	decl	%esi			/ decrement number of bytes by 1
139*5d9d9091SRichard Lowe	movb	%dl, (%edi)		/ copy %dl to (%edi)
140*5d9d9091SRichard Lowe	incl	%edi			/ next byte
141*5d9d9091SRichard Lowe	incl	%eax			/ next byte
142*5d9d9091SRichard Lowe	cmpb	$0, %dl			/ compare %dl with a null byte
143*5d9d9091SRichard Lowe	je	.L9			/ if %dl is a null, goto .L9
144*5d9d9091SRichard Lowe	jmp	.L4			/ goto .L4
145*5d9d9091SRichard Lowe	.align	4
146*5d9d9091SRichard Lowe.L8:
147*5d9d9091SRichard Lowe	movb	$0, (%edi)		/ null termination
148*5d9d9091SRichard Lowe.L9:
149*5d9d9091SRichard Lowe	movl	12(%esp), %eax		/ return the destination address
150*5d9d9091SRichard Lowe	popl	%esi			/ restore register variables
151*5d9d9091SRichard Lowe	popl	%edi
152*5d9d9091SRichard Lowe	ret
153*5d9d9091SRichard Lowe	SET_SIZE(strncat)
154