xref: /illumos-gate/usr/src/lib/libc/i386/gen/strncpy.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	"strncpy.s"
27*5d9d9091SRichard Lowe
28*5d9d9091SRichard Lowe/
29*5d9d9091SRichard Lowe/ strncpy(s1, s2, n)
30*5d9d9091SRichard Lowe/
31*5d9d9091SRichard Lowe/ Copies s2 to s1, truncating or null-padding to always copy n bytes
32*5d9d9091SRichard Lowe/ Returns s1
33*5d9d9091SRichard Lowe/
34*5d9d9091SRichard Lowe/ Fast assembly language version of the following C-program strncpy
35*5d9d9091SRichard Lowe/ which represents the `standard' for the C-library.
36*5d9d9091SRichard Lowe/
37*5d9d9091SRichard Lowe/	char *
38*5d9d9091SRichard Lowe/	strncpy(char *s1, const char *s2, size_t n)
39*5d9d9091SRichard Lowe/	{
40*5d9d9091SRichard Lowe/		char	*os1 = s1;
41*5d9d9091SRichard Lowe/
42*5d9d9091SRichard Lowe/		n++;
43*5d9d9091SRichard Lowe/		while ((--n != 0) && ((*s1++ = *s2++) != '\0'))
44*5d9d9091SRichard Lowe/			;
45*5d9d9091SRichard Lowe/		if (n != 0)
46*5d9d9091SRichard Lowe/			while (--n != 0)
47*5d9d9091SRichard Lowe/				*s1++ = '\0';
48*5d9d9091SRichard Lowe/		return (os1);
49*5d9d9091SRichard Lowe/	}
50*5d9d9091SRichard Lowe/
51*5d9d9091SRichard Lowe/ In this assembly language version, the following expression is used
52*5d9d9091SRichard Lowe/ to check if a 32-bit word data contains a null byte or not:
53*5d9d9091SRichard Lowe/	(((A & 0x7f7f7f7f) + 0x7f7f7f7f) | A) & 0x80808080
54*5d9d9091SRichard Lowe/ If the above expression geneates a value other than 0x80808080,
55*5d9d9091SRichard Lowe/ that means the 32-bit word data contains a null byte.
56*5d9d9091SRichard Lowe/
57*5d9d9091SRichard Lowe
58*5d9d9091SRichard Lowe/ Assume relatively long strings and small numbers of nulls at end.
59*5d9d9091SRichard Lowe
60*5d9d9091SRichard Lowe#include "SYS.h"
61*5d9d9091SRichard Lowe
62*5d9d9091SRichard Lowe	ENTRY(strncpy)
63*5d9d9091SRichard Lowe	pushl	%edi			/ save register variables
64*5d9d9091SRichard Lowe	pushl	%esi
65*5d9d9091SRichard Lowe
66*5d9d9091SRichard Lowe	movl	16(%esp), %eax		/ %eax = source string address
67*5d9d9091SRichard Lowe	movl	12(%esp), %edi		/ %edi = destination string address
68*5d9d9091SRichard Lowe	movl	20(%esp), %esi		/ %esi = number of bytes
69*5d9d9091SRichard Lowe
70*5d9d9091SRichard Lowe	testl	$3, %eax		/ if %eax not word aligned
71*5d9d9091SRichard Lowe	jnz	.L1			/ goto .L1
72*5d9d9091SRichard Lowe.L8:
73*5d9d9091SRichard Lowe	cmpl	$4, %esi		/ if number of bytes < 4
74*5d9d9091SRichard Lowe	jb	.L4			/ goto .L4
75*5d9d9091SRichard Lowe	.align	4
76*5d9d9091SRichard Lowe.L2:
77*5d9d9091SRichard Lowe	movl	(%eax), %edx		/ move 1 word from (%eax) to %edx
78*5d9d9091SRichard Lowe	movl	$0x7f7f7f7f, %ecx
79*5d9d9091SRichard Lowe	andl	%edx, %ecx		/ %ecx = %edx & 0x7f7f7f7f
80*5d9d9091SRichard Lowe	addl	$4, %eax		/ next word
81*5d9d9091SRichard Lowe	addl	$0x7f7f7f7f, %ecx	/ %ecx += 0x7f7f7f7f
82*5d9d9091SRichard Lowe	orl	%edx, %ecx		/ %ecx |= %edx
83*5d9d9091SRichard Lowe	andl	$0x80808080, %ecx	/ %ecx &= 0x80808080
84*5d9d9091SRichard Lowe	cmpl	$0x80808080, %ecx	/ if null byte in this word
85*5d9d9091SRichard Lowe	jne	.L3			/ goto .L3
86*5d9d9091SRichard Lowe	movl	%edx, (%edi)		/ copy this word to (%edi)
87*5d9d9091SRichard Lowe	subl	$4, %esi		/ decrement number of bytes by 4
88*5d9d9091SRichard Lowe	addl	$4, %edi		/ next word
89*5d9d9091SRichard Lowe	cmpl	$4, %esi		/ if number of bytes >= 4
90*5d9d9091SRichard Lowe	jae	.L2			/ goto .L2
91*5d9d9091SRichard Lowe	jmp	.L4			/ goto .L4
92*5d9d9091SRichard Lowe.L3:
93*5d9d9091SRichard Lowe	subl	$4, %eax		/ post-incremented
94*5d9d9091SRichard Lowe	.align	4
95*5d9d9091SRichard Lowe.L4:
96*5d9d9091SRichard Lowe	/ (number of bytes < 4) or (a null byte found in the word)
97*5d9d9091SRichard Lowe	cmpl	$0, %esi		/ if number of bytes == 0
98*5d9d9091SRichard Lowe	je	.L7			/ goto .L7 (finished)
99*5d9d9091SRichard Lowe	movb	(%eax), %dl		/ %dl = a byte in (%eax)
100*5d9d9091SRichard Lowe	decl	%esi			/ decrement number of bytes by 1
101*5d9d9091SRichard Lowe	movb	%dl, (%edi)		/ copy %dl to (%edi)
102*5d9d9091SRichard Lowe	incl	%eax			/ next byte
103*5d9d9091SRichard Lowe	incl	%edi			/ next byte
104*5d9d9091SRichard Lowe	cmpb	$0, %dl			/ compare %dl with a null byte
105*5d9d9091SRichard Lowe	je	.L5			/ if %dl is a null, goto .L5
106*5d9d9091SRichard Lowe	jmp	.L4			/ goto .L4
107*5d9d9091SRichard Lowe	.align	4
108*5d9d9091SRichard Lowe.L1:
109*5d9d9091SRichard Lowe	/ %eax not aligned
110*5d9d9091SRichard Lowe	cmpl	$0, %esi		/ if number of bytes == 0
111*5d9d9091SRichard Lowe	je	.L7			/ goto .L7 (finished)
112*5d9d9091SRichard Lowe	movb	(%eax), %dl		/ %dl = a byte in (%eax)
113*5d9d9091SRichard Lowe	decl	%esi			/ decrement number of bytes by 1
114*5d9d9091SRichard Lowe	movb	%dl, (%edi)		/ copy %dl to (%edi)
115*5d9d9091SRichard Lowe	incl	%edi			/ next byte
116*5d9d9091SRichard Lowe	incl	%eax			/ next byte
117*5d9d9091SRichard Lowe	cmpb	$0, %dl			/ compare %dl with a null byte
118*5d9d9091SRichard Lowe	je	.L5			/ if %dl is a null, goto .L5
119*5d9d9091SRichard Lowe	testl	$3, %eax		/ if %eax word aligned
120*5d9d9091SRichard Lowe	jz	.L8			/ goto .L8
121*5d9d9091SRichard Lowe	jmp	.L1			/ goto .L1 (not word aligned)
122*5d9d9091SRichard Lowe	.align	4
123*5d9d9091SRichard Lowe.L5:
124*5d9d9091SRichard Lowe	movl	%esi, %ecx		/ %ecx = length to copy null bytes
125*5d9d9091SRichard Lowe	xorl	%eax, %eax		/ clear %eax
126*5d9d9091SRichard Lowe	shrl	$2, %ecx		/ %ecx = words to copy null bytes
127*5d9d9091SRichard Lowe	rep ; sstol			/ rep;sstol is optimal
128*5d9d9091SRichard Lowe	andl	$3, %esi		/ %esi = leftover bytes
129*5d9d9091SRichard Lowe.L6:
130*5d9d9091SRichard Lowe	cmpl	$0, %esi		/ if number of bytes == 0
131*5d9d9091SRichard Lowe	jz	.L7			/ goto .L7 (finished)
132*5d9d9091SRichard Lowe	movb	$0, (%edi)		/ move a null byte to (%edi)
133*5d9d9091SRichard Lowe	decl	%esi			/ decrement number of bytes by 1
134*5d9d9091SRichard Lowe	incl	%edi			/ next byte
135*5d9d9091SRichard Lowe	jmp	.L6			/ goto .L6
136*5d9d9091SRichard Lowe	.align	4
137*5d9d9091SRichard Lowe.L7:
138*5d9d9091SRichard Lowe	movl	12(%esp), %eax		/ return the destination address
139*5d9d9091SRichard Lowe	popl	%esi			/ restore register variables
140*5d9d9091SRichard Lowe	popl	%edi
141*5d9d9091SRichard Lowe	ret
142*5d9d9091SRichard Lowe	SET_SIZE(strncpy)
143