xref: /onnv-gate/usr/src/lib/libc/sparc/gen/strncpy.s (revision 0:68f95e015346)
1*0Sstevel@tonic-gate/*
2*0Sstevel@tonic-gate * CDDL HEADER START
3*0Sstevel@tonic-gate *
4*0Sstevel@tonic-gate * The contents of this file are subject to the terms of the
5*0Sstevel@tonic-gate * Common Development and Distribution License, Version 1.0 only
6*0Sstevel@tonic-gate * (the "License").  You may not use this file except in compliance
7*0Sstevel@tonic-gate * with the License.
8*0Sstevel@tonic-gate *
9*0Sstevel@tonic-gate * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10*0Sstevel@tonic-gate * or http://www.opensolaris.org/os/licensing.
11*0Sstevel@tonic-gate * See the License for the specific language governing permissions
12*0Sstevel@tonic-gate * and limitations under the License.
13*0Sstevel@tonic-gate *
14*0Sstevel@tonic-gate * When distributing Covered Code, include this CDDL HEADER in each
15*0Sstevel@tonic-gate * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16*0Sstevel@tonic-gate * If applicable, add the following below this CDDL HEADER, with the
17*0Sstevel@tonic-gate * fields enclosed by brackets "[]" replaced with your own identifying
18*0Sstevel@tonic-gate * information: Portions Copyright [yyyy] [name of copyright owner]
19*0Sstevel@tonic-gate *
20*0Sstevel@tonic-gate * CDDL HEADER END
21*0Sstevel@tonic-gate */
22*0Sstevel@tonic-gate/*
23*0Sstevel@tonic-gate * Copyright 2003 Sun Microsystems, Inc.  All rights reserved.
24*0Sstevel@tonic-gate * Use is subject to license terms.
25*0Sstevel@tonic-gate */
26*0Sstevel@tonic-gate
27*0Sstevel@tonic-gate
28*0Sstevel@tonic-gate.ident	"%Z%%M%	%I%	%E% SMI"
29*0Sstevel@tonic-gate
30*0Sstevel@tonic-gate	.file	"%M%"
31*0Sstevel@tonic-gate
32*0Sstevel@tonic-gate/*
33*0Sstevel@tonic-gate * strncpy(s1, s2)
34*0Sstevel@tonic-gate *
35*0Sstevel@tonic-gate * Copy string s2 to s1, truncating or null-padding to always copy n bytes
36*0Sstevel@tonic-gate * return s1.
37*0Sstevel@tonic-gate *
38*0Sstevel@tonic-gate * Fast assembler language version of the following C-program for strncpy
39*0Sstevel@tonic-gate * which represents the `standard' for the C-library.
40*0Sstevel@tonic-gate *
41*0Sstevel@tonic-gate *	char *
42*0Sstevel@tonic-gate *	strncpy(char *s1, const char *s2, size_t n)
43*0Sstevel@tonic-gate *	{
44*0Sstevel@tonic-gate *		char *os1 = s1;
45*0Sstevel@tonic-gate *
46*0Sstevel@tonic-gate *		n++;
47*0Sstevel@tonic-gate *		while ((--n != 0) &&  ((*s1++ = *s2++) != '\0'))
48*0Sstevel@tonic-gate *			;
49*0Sstevel@tonic-gate *		if (n != 0)
50*0Sstevel@tonic-gate *			while (--n != 0)
51*0Sstevel@tonic-gate *				*s1++ = '\0';
52*0Sstevel@tonic-gate *		return (os1);
53*0Sstevel@tonic-gate *	}
54*0Sstevel@tonic-gate */
55*0Sstevel@tonic-gate
56*0Sstevel@tonic-gate#include <sys/asm_linkage.h>
57*0Sstevel@tonic-gate#include "synonyms.h"
58*0Sstevel@tonic-gate
59*0Sstevel@tonic-gate	! strncpy works similarly to strcpy, except that n bytes of s2
60*0Sstevel@tonic-gate	! are copied to s1. If a null character is reached in s2 yet more
61*0Sstevel@tonic-gate	! bytes remain to be copied, strncpy will copy null bytes into
62*0Sstevel@tonic-gate	! the destination string.
63*0Sstevel@tonic-gate	!
64*0Sstevel@tonic-gate	! This implementation works by first aligning the src ptr and
65*0Sstevel@tonic-gate	! performing small copies until it is aligned.  Then, the string
66*0Sstevel@tonic-gate	! is copied based upon destination alignment.  (byte, half-word,
67*0Sstevel@tonic-gate	! word, etc.)
68*0Sstevel@tonic-gate
69*0Sstevel@tonic-gate	ENTRY(strncpy)
70*0Sstevel@tonic-gate
71*0Sstevel@tonic-gate	.align 32
72*0Sstevel@tonic-gate	subcc	%g0, %o2, %o4		! n = -n
73*0Sstevel@tonic-gate	bz	.doneshort		! if n == 0, done
74*0Sstevel@tonic-gate	cmp	%o2, 7			! n < 7 ?
75*0Sstevel@tonic-gate	add	%o1, %o2, %o3		! src = src + n
76*0Sstevel@tonic-gate	blu	.shortcpy		! n < 7, use byte-wise copy
77*0Sstevel@tonic-gate	add	%o0, %o2, %o2		! dst = dst + n
78*0Sstevel@tonic-gate	andcc	%o1, 3, %o5		! src word aligned ?
79*0Sstevel@tonic-gate	bz	.wordaligned		! yup
80*0Sstevel@tonic-gate	save	%sp, -0x40, %sp		! create new register window
81*0Sstevel@tonic-gate	sub	%i5, 4, %i5		! bytes until src aligned
82*0Sstevel@tonic-gate	nop				! align loop on 16-byte boundary
83*0Sstevel@tonic-gate	nop				! align loop on 16-byte boundary
84*0Sstevel@tonic-gate
85*0Sstevel@tonic-gate.alignsrc:
86*0Sstevel@tonic-gate	ldub	[%i3 + %i4], %i1	! src[]
87*0Sstevel@tonic-gate	stb	%i1, [%i2 + %i4]	! dst[] = src[]
88*0Sstevel@tonic-gate	inccc	%i4			! src++, dst++, n--
89*0Sstevel@tonic-gate	bz	.done			! n == 0, done
90*0Sstevel@tonic-gate	tst     %i1			! end of src reached (null byte) ?
91*0Sstevel@tonic-gate	bz,a	.bytepad		! yes, at least one byte to pad here
92*0Sstevel@tonic-gate	add 	%i2, %i4, %l0		! need single dest pointer for fill
93*0Sstevel@tonic-gate	inccc	%i5			! src aligned now?
94*0Sstevel@tonic-gate	bnz	.alignsrc		! no, copy another byte
95*0Sstevel@tonic-gate	.empty
96*0Sstevel@tonic-gate
97*0Sstevel@tonic-gate.wordaligned:
98*0Sstevel@tonic-gate	add	%i2, %i4, %l0		! dst
99*0Sstevel@tonic-gate	sethi	%hi(0x01010101), %l1	! Alan Mycroft's magic1
100*0Sstevel@tonic-gate	sub	%i2, 4, %i2		! adjust for dest pre-incr in cpy loops
101*0Sstevel@tonic-gate	or	%l1, %lo(0x01010101),%l1!  finish loading magic1
102*0Sstevel@tonic-gate	andcc	%l0, 3, %g1		! destination word aligned ?
103*0Sstevel@tonic-gate	bnz	.dstnotaligned		! nope
104*0Sstevel@tonic-gate	sll	%l1, 7, %i5		! create Alan Mycroft's magic2
105*0Sstevel@tonic-gate
106*0Sstevel@tonic-gate.storeword:
107*0Sstevel@tonic-gate	lduw	[%i3 + %i4], %i1	! src dword
108*0Sstevel@tonic-gate	addcc	%i4, 4, %i4		! n += 4, src += 4, dst += 4
109*0Sstevel@tonic-gate	bcs	.lastword		! if counter wraps, last word
110*0Sstevel@tonic-gate	andn	%i5, %i1, %g1		! ~dword & 0x80808080
111*0Sstevel@tonic-gate	sub	%i1, %l1, %l0		! dword - 0x01010101
112*0Sstevel@tonic-gate	andcc	%l0, %g1, %g0		! ((dword - 0x01010101) & ~dword & 0x80808080)
113*0Sstevel@tonic-gate	bz,a	.storeword		! no zero byte if magic expression == 0
114*0Sstevel@tonic-gate	stw	%i1, [%i2 + %i4]	! store word to dst (address pre-incremented)
115*0Sstevel@tonic-gate
116*0Sstevel@tonic-gate	! n has not expired, but src is at the end. we need to push out the
117*0Sstevel@tonic-gate	! remaining src bytes and then start padding with null bytes
118*0Sstevel@tonic-gate
119*0Sstevel@tonic-gate.zerobyte:
120*0Sstevel@tonic-gate	add	%i2, %i4, %l0		! pointer to dest string
121*0Sstevel@tonic-gate	srl	%i1, 24, %g1		! first byte
122*0Sstevel@tonic-gate	stb	%g1, [%l0]		! store it
123*0Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
124*0Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
125*0Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null bytes
126*0Sstevel@tonic-gate	srl	%i1, 16, %g1		! second byte
127*0Sstevel@tonic-gate	stb	%g1, [%l0 + 1]		! store it
128*0Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
129*0Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
130*0Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
131*0Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null bytes
132*0Sstevel@tonic-gate	srl	%i1, 8, %g1		! third byte
133*0Sstevel@tonic-gate	stb	%g1, [%l0 + 2]		! store it
134*0Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
135*0Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
136*0Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
137*0Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null bytes
138*0Sstevel@tonic-gate	stb	%i1, [%l0 + 3]		! store fourth byte
139*0Sstevel@tonic-gate	addcc	%i4, 8, %g0		! number of pad bytes < 8 ?
140*0Sstevel@tonic-gate	bcs	.bytepad		! yes, do simple byte wise fill
141*0Sstevel@tonic-gate	add	%l0, 4, %l0		! dst += 4
142*0Sstevel@tonic-gate	andcc	%l0, 3, %l1		! dst offset relative to word boundary
143*0Sstevel@tonic-gate	bz	.fillaligned		! dst already word aligned
144*0Sstevel@tonic-gate
145*0Sstevel@tonic-gate	! here there is a least one more byte to zero out: otherwise we would
146*0Sstevel@tonic-gate	! have exited through label .lastword
147*0Sstevel@tonic-gate
148*0Sstevel@tonic-gate	sub	%l1, 4, %l1		! bytes to align dst to word boundary
149*0Sstevel@tonic-gate.makealigned:
150*0Sstevel@tonic-gate	stb	%g0, [%l0]		! dst[] = 0
151*0Sstevel@tonic-gate	addcc	%i4, 1, %i4		! n--
152*0Sstevel@tonic-gate	bz	.done			! n == 0, we are done
153*0Sstevel@tonic-gate	addcc	%l1, 1, %l1		! any more byte needed to align
154*0Sstevel@tonic-gate	bnz	.makealigned		! yup, pad another byte
155*0Sstevel@tonic-gate	add	%l0, 1, %l0		! dst++
156*0Sstevel@tonic-gate	nop				! pad to align copy loop below
157*0Sstevel@tonic-gate
158*0Sstevel@tonic-gate	! here we know that there at least another 4 bytes to pad, since
159*0Sstevel@tonic-gate	! we don't get here unless there were >= 8 bytes to pad to begin
160*0Sstevel@tonic-gate	! with, and we have padded at most 3 bytes suring dst aligning
161*0Sstevel@tonic-gate
162*0Sstevel@tonic-gate.fillaligned:
163*0Sstevel@tonic-gate	add	%i4, 3, %i2		! round up to next word boundary
164*0Sstevel@tonic-gate	and	%i2, -4, %l1		! pointer to next word boundary
165*0Sstevel@tonic-gate	and	%i2, 4, %i2		! word count odd ? 4 : 0
166*0Sstevel@tonic-gate	stw	%g0, [%l0]		! store first word
167*0Sstevel@tonic-gate	addcc	%l1, %i2, %l1		! dword count == 1 ?
168*0Sstevel@tonic-gate	add	%i4, %i2, %i4		! if word count odd, n -= 4
169*0Sstevel@tonic-gate	bz	.bytepad		! if word count == 1, pad bytes left
170*0Sstevel@tonic-gate	add	%l0, %i2, %l0		! bump dst if word count odd
171*0Sstevel@tonic-gate
172*0Sstevel@tonic-gate.fillword:
173*0Sstevel@tonic-gate	addcc	%l1, 8, %l1		! count -= 8
174*0Sstevel@tonic-gate	stw	%g0, [%l0]		! dst[n] = 0
175*0Sstevel@tonic-gate	stw	%g0, [%l0 + 4]		! dst[n+4] = 0
176*0Sstevel@tonic-gate	add	%l0, 8, %l0		! dst += 8
177*0Sstevel@tonic-gate	bcc	.fillword		! fill words until count == 0
178*0Sstevel@tonic-gate	addcc	%i4, 8, %i4		! n -= 8
179*0Sstevel@tonic-gate	bz	.done			! if n == 0, we are done
180*0Sstevel@tonic-gate	.empty
181*0Sstevel@tonic-gate
182*0Sstevel@tonic-gate.bytepad:
183*0Sstevel@tonic-gate	and	%i4, 1, %i2		! byte count odd ? 1 : 0
184*0Sstevel@tonic-gate	stb	%g0, [%l0]		! store first byte
185*0Sstevel@tonic-gate	addcc	%i4, %i2, %i4		! byte count == 1 ?
186*0Sstevel@tonic-gate	bz	.done			! yup, we are done
187*0Sstevel@tonic-gate	add	%l0, %i2, %l0		! bump pointer if odd
188*0Sstevel@tonic-gate
189*0Sstevel@tonic-gate.fillbyte:
190*0Sstevel@tonic-gate	addcc	%i4, 2, %i4		! n -= 2
191*0Sstevel@tonic-gate	stb	%g0, [%l0]		! dst[n] = 0
192*0Sstevel@tonic-gate	stb	%g0, [%l0 + 1]		! dst[n+1] = 0
193*0Sstevel@tonic-gate	bnz	.fillbyte		! fill until n == 0
194*0Sstevel@tonic-gate	add	%l0, 2, %l0		! dst += 2
195*0Sstevel@tonic-gate
196*0Sstevel@tonic-gate.done:
197*0Sstevel@tonic-gate	ret				! done
198*0Sstevel@tonic-gate	restore	%i0, %g0, %o0		! restore reg window, return dst
199*0Sstevel@tonic-gate
200*0Sstevel@tonic-gate	! this is the last word. It may contain null bytes. store bytes
201*0Sstevel@tonic-gate	! until n == 0. if null byte encountered, continue
202*0Sstevel@tonic-gate
203*0Sstevel@tonic-gate.lastword:
204*0Sstevel@tonic-gate	sub	%i4, 4, %i4		! undo counter pre-increment
205*0Sstevel@tonic-gate	add	%i2, 4, %i2		! adjust dst for counter un-bumping
206*0Sstevel@tonic-gate
207*0Sstevel@tonic-gate	srl	%i1, 24, %g1		! first byte
208*0Sstevel@tonic-gate	stb	%g1, [%i2 + %i4]	! store it
209*0Sstevel@tonic-gate	inccc	%i4			! n--
210*0Sstevel@tonic-gate	bz	.done			! if n == 0, we're done
211*0Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
212*0Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
213*0Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null
214*0Sstevel@tonic-gate	srl	%i1, 16, %g1		! second byte
215*0Sstevel@tonic-gate	stb	%g1, [%i2 + %i4]	! store it
216*0Sstevel@tonic-gate	inccc	%i4			! n--
217*0Sstevel@tonic-gate	bz	.done			! if n == 0, we're done
218*0Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
219*0Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
220*0Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
221*0Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null
222*0Sstevel@tonic-gate	srl	%i1, 8, %g1		! third byte
223*0Sstevel@tonic-gate	stb	%g1, [%i2 + %i4]	! store it
224*0Sstevel@tonic-gate	inccc	%i4			! n--
225*0Sstevel@tonic-gate	bz	.done			! if n == 0, we're done
226*0Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
227*0Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
228*0Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
229*0Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null
230*0Sstevel@tonic-gate	ba	.done			! here n must be zero, we are done
231*0Sstevel@tonic-gate	stb	%i1, [%i2 + %i4]	! store fourth byte
232*0Sstevel@tonic-gate
233*0Sstevel@tonic-gate.dstnotaligned:
234*0Sstevel@tonic-gate	cmp	%g1, 2			! dst half word aligned?
235*0Sstevel@tonic-gate	be	.storehalfword2		! yup, store half word at a time
236*0Sstevel@tonic-gate	.empty
237*0Sstevel@tonic-gate.storebyte:
238*0Sstevel@tonic-gate	lduw	[%i3 + %i4], %i1	! x = src[]
239*0Sstevel@tonic-gate	addcc	%i4, 4, %i4		! src += 4, dst += 4, n -= 4
240*0Sstevel@tonic-gate	bcs	.lastword		! if counter wraps, last word
241*0Sstevel@tonic-gate	andn	%i5, %i1, %g1		! ~x & 0x80808080
242*0Sstevel@tonic-gate	sub	%i1, %l1, %l0		! x - 0x01010101
243*0Sstevel@tonic-gate	andcc	%l0, %g1, %g0		! ((x - 0x01010101) & ~x & 0x80808080)
244*0Sstevel@tonic-gate	bnz	.zerobyte		! end of src found, may need to pad
245*0Sstevel@tonic-gate	add	%i2, %i4, %l0		! dst (in pointer form)
246*0Sstevel@tonic-gate	srl	%i1, 24, %g1		! %g1<7:0> = 1st byte; half-word aligned now
247*0Sstevel@tonic-gate	stb	%g1, [%l0]		! store first byte
248*0Sstevel@tonic-gate	srl	%i1, 8, %g1		! %g1<15:0> = bytes 2, 3
249*0Sstevel@tonic-gate	sth	%g1, [%l0 + 1]		! store bytes 2, 3
250*0Sstevel@tonic-gate	ba	.storebyte		! next word
251*0Sstevel@tonic-gate	stb	%i1, [%l0 + 3]		! store fourth byte
252*0Sstevel@tonic-gate	nop
253*0Sstevel@tonic-gate	nop
254*0Sstevel@tonic-gate
255*0Sstevel@tonic-gate.storehalfword:
256*0Sstevel@tonic-gate	lduw	[%i3 + %i4], %i1	! x = src[]
257*0Sstevel@tonic-gate.storehalfword2:
258*0Sstevel@tonic-gate	addcc	%i4, 4, %i4		! src += 4, dst += 4, n -= 4
259*0Sstevel@tonic-gate	bcs	.lastword		! if counter wraps, last word
260*0Sstevel@tonic-gate	andn	%i5, %i1, %g1		! ~x & 0x80808080
261*0Sstevel@tonic-gate	sub	%i1, %l1, %l0		! x - 0x01010101
262*0Sstevel@tonic-gate	andcc	%l0, %g1, %g0		! ((x -0x01010101) & ~x & 0x8080808080)
263*0Sstevel@tonic-gate	bnz	.zerobyte		! x has zero byte, handle end cases
264*0Sstevel@tonic-gate	add	%i2, %i4, %l0		! dst (in pointer form)
265*0Sstevel@tonic-gate	srl	%i1, 16, %g1		! %g1<15:0> = bytes 1, 2
266*0Sstevel@tonic-gate	sth	%g1, [%l0]		! store bytes 1, 2
267*0Sstevel@tonic-gate	ba	.storehalfword		! next dword
268*0Sstevel@tonic-gate	sth	%i1, [%l0 + 2]		! store bytes 3, 4
269*0Sstevel@tonic-gate
270*0Sstevel@tonic-gate.shortcpy:
271*0Sstevel@tonic-gate	ldub	[%o3 + %o4], %o5	! src[]
272*0Sstevel@tonic-gate	stb	%o5, [%o2 + %o4]	! dst[] = src[]
273*0Sstevel@tonic-gate	inccc	%o4			! src++, dst++, n--
274*0Sstevel@tonic-gate	bz	.doneshort		! if n == 0, done
275*0Sstevel@tonic-gate	tst	%o5			! src[] == 0 ?
276*0Sstevel@tonic-gate	bnz,a	.shortcpy		! nope, next byte
277*0Sstevel@tonic-gate	nop				! empty delay slot
278*0Sstevel@tonic-gate
279*0Sstevel@tonic-gate.padbyte:
280*0Sstevel@tonic-gate	stb	%g0, [%o2 + %o4]	! dst[] = 0
281*0Sstevel@tonic-gate.padbyte2:
282*0Sstevel@tonic-gate	addcc	%o4, 1, %o4		! dst++, n--
283*0Sstevel@tonic-gate	bnz,a	.padbyte2		! if n != 0, next byte
284*0Sstevel@tonic-gate	stb	%g0, [%o2 + %o4]	! dst[] = 0
285*0Sstevel@tonic-gate	nop				! align label below to 16-byte boundary
286*0Sstevel@tonic-gate
287*0Sstevel@tonic-gate.doneshort:
288*0Sstevel@tonic-gate	retl				! return from leaf
289*0Sstevel@tonic-gate	nop				! empty delay slot
290*0Sstevel@tonic-gate	SET_SIZE(strncpy)
291