xref: /onnv-gate/usr/src/lib/libc/sparc/gen/strncpy.s (revision 7298:b69e27387f74)
10Sstevel@tonic-gate/*
20Sstevel@tonic-gate * CDDL HEADER START
30Sstevel@tonic-gate *
40Sstevel@tonic-gate * The contents of this file are subject to the terms of the
56812Sraf * Common Development and Distribution License (the "License").
66812Sraf * You may not use this file except in compliance with the License.
70Sstevel@tonic-gate *
80Sstevel@tonic-gate * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
90Sstevel@tonic-gate * or http://www.opensolaris.org/os/licensing.
100Sstevel@tonic-gate * See the License for the specific language governing permissions
110Sstevel@tonic-gate * and limitations under the License.
120Sstevel@tonic-gate *
130Sstevel@tonic-gate * When distributing Covered Code, include this CDDL HEADER in each
140Sstevel@tonic-gate * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
150Sstevel@tonic-gate * If applicable, add the following below this CDDL HEADER, with the
160Sstevel@tonic-gate * fields enclosed by brackets "[]" replaced with your own identifying
170Sstevel@tonic-gate * information: Portions Copyright [yyyy] [name of copyright owner]
180Sstevel@tonic-gate *
190Sstevel@tonic-gate * CDDL HEADER END
200Sstevel@tonic-gate */
216812Sraf
220Sstevel@tonic-gate/*
236812Sraf * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
240Sstevel@tonic-gate * Use is subject to license terms.
250Sstevel@tonic-gate */
260Sstevel@tonic-gate
27*7298SMark.J.Nelson@Sun.COM	.file	"strncpy.s"
280Sstevel@tonic-gate
290Sstevel@tonic-gate/*
300Sstevel@tonic-gate * strncpy(s1, s2)
310Sstevel@tonic-gate *
320Sstevel@tonic-gate * Copy string s2 to s1, truncating or null-padding to always copy n bytes
330Sstevel@tonic-gate * return s1.
340Sstevel@tonic-gate *
350Sstevel@tonic-gate * Fast assembler language version of the following C-program for strncpy
360Sstevel@tonic-gate * which represents the `standard' for the C-library.
370Sstevel@tonic-gate *
380Sstevel@tonic-gate *	char *
390Sstevel@tonic-gate *	strncpy(char *s1, const char *s2, size_t n)
400Sstevel@tonic-gate *	{
410Sstevel@tonic-gate *		char *os1 = s1;
420Sstevel@tonic-gate *
430Sstevel@tonic-gate *		n++;
440Sstevel@tonic-gate *		while ((--n != 0) &&  ((*s1++ = *s2++) != '\0'))
450Sstevel@tonic-gate *			;
460Sstevel@tonic-gate *		if (n != 0)
470Sstevel@tonic-gate *			while (--n != 0)
480Sstevel@tonic-gate *				*s1++ = '\0';
490Sstevel@tonic-gate *		return (os1);
500Sstevel@tonic-gate *	}
510Sstevel@tonic-gate */
520Sstevel@tonic-gate
530Sstevel@tonic-gate#include <sys/asm_linkage.h>
540Sstevel@tonic-gate
550Sstevel@tonic-gate	! strncpy works similarly to strcpy, except that n bytes of s2
560Sstevel@tonic-gate	! are copied to s1. If a null character is reached in s2 yet more
570Sstevel@tonic-gate	! bytes remain to be copied, strncpy will copy null bytes into
580Sstevel@tonic-gate	! the destination string.
590Sstevel@tonic-gate	!
600Sstevel@tonic-gate	! This implementation works by first aligning the src ptr and
610Sstevel@tonic-gate	! performing small copies until it is aligned.  Then, the string
620Sstevel@tonic-gate	! is copied based upon destination alignment.  (byte, half-word,
630Sstevel@tonic-gate	! word, etc.)
640Sstevel@tonic-gate
650Sstevel@tonic-gate	ENTRY(strncpy)
660Sstevel@tonic-gate
670Sstevel@tonic-gate	.align 32
680Sstevel@tonic-gate	subcc	%g0, %o2, %o4		! n = -n
690Sstevel@tonic-gate	bz	.doneshort		! if n == 0, done
700Sstevel@tonic-gate	cmp	%o2, 7			! n < 7 ?
710Sstevel@tonic-gate	add	%o1, %o2, %o3		! src = src + n
720Sstevel@tonic-gate	blu	.shortcpy		! n < 7, use byte-wise copy
730Sstevel@tonic-gate	add	%o0, %o2, %o2		! dst = dst + n
740Sstevel@tonic-gate	andcc	%o1, 3, %o5		! src word aligned ?
750Sstevel@tonic-gate	bz	.wordaligned		! yup
760Sstevel@tonic-gate	save	%sp, -0x40, %sp		! create new register window
770Sstevel@tonic-gate	sub	%i5, 4, %i5		! bytes until src aligned
780Sstevel@tonic-gate	nop				! align loop on 16-byte boundary
790Sstevel@tonic-gate	nop				! align loop on 16-byte boundary
800Sstevel@tonic-gate
810Sstevel@tonic-gate.alignsrc:
820Sstevel@tonic-gate	ldub	[%i3 + %i4], %i1	! src[]
830Sstevel@tonic-gate	stb	%i1, [%i2 + %i4]	! dst[] = src[]
840Sstevel@tonic-gate	inccc	%i4			! src++, dst++, n--
850Sstevel@tonic-gate	bz	.done			! n == 0, done
860Sstevel@tonic-gate	tst     %i1			! end of src reached (null byte) ?
870Sstevel@tonic-gate	bz,a	.bytepad		! yes, at least one byte to pad here
880Sstevel@tonic-gate	add 	%i2, %i4, %l0		! need single dest pointer for fill
890Sstevel@tonic-gate	inccc	%i5			! src aligned now?
900Sstevel@tonic-gate	bnz	.alignsrc		! no, copy another byte
910Sstevel@tonic-gate	.empty
920Sstevel@tonic-gate
930Sstevel@tonic-gate.wordaligned:
940Sstevel@tonic-gate	add	%i2, %i4, %l0		! dst
950Sstevel@tonic-gate	sethi	%hi(0x01010101), %l1	! Alan Mycroft's magic1
960Sstevel@tonic-gate	sub	%i2, 4, %i2		! adjust for dest pre-incr in cpy loops
970Sstevel@tonic-gate	or	%l1, %lo(0x01010101),%l1!  finish loading magic1
980Sstevel@tonic-gate	andcc	%l0, 3, %g1		! destination word aligned ?
990Sstevel@tonic-gate	bnz	.dstnotaligned		! nope
1000Sstevel@tonic-gate	sll	%l1, 7, %i5		! create Alan Mycroft's magic2
1010Sstevel@tonic-gate
1020Sstevel@tonic-gate.storeword:
1030Sstevel@tonic-gate	lduw	[%i3 + %i4], %i1	! src dword
1040Sstevel@tonic-gate	addcc	%i4, 4, %i4		! n += 4, src += 4, dst += 4
1050Sstevel@tonic-gate	bcs	.lastword		! if counter wraps, last word
1060Sstevel@tonic-gate	andn	%i5, %i1, %g1		! ~dword & 0x80808080
1070Sstevel@tonic-gate	sub	%i1, %l1, %l0		! dword - 0x01010101
1080Sstevel@tonic-gate	andcc	%l0, %g1, %g0		! ((dword - 0x01010101) & ~dword & 0x80808080)
1090Sstevel@tonic-gate	bz,a	.storeword		! no zero byte if magic expression == 0
1100Sstevel@tonic-gate	stw	%i1, [%i2 + %i4]	! store word to dst (address pre-incremented)
1110Sstevel@tonic-gate
1120Sstevel@tonic-gate	! n has not expired, but src is at the end. we need to push out the
1130Sstevel@tonic-gate	! remaining src bytes and then start padding with null bytes
1140Sstevel@tonic-gate
1150Sstevel@tonic-gate.zerobyte:
1160Sstevel@tonic-gate	add	%i2, %i4, %l0		! pointer to dest string
1170Sstevel@tonic-gate	srl	%i1, 24, %g1		! first byte
1180Sstevel@tonic-gate	stb	%g1, [%l0]		! store it
1190Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
1200Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
1210Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null bytes
1220Sstevel@tonic-gate	srl	%i1, 16, %g1		! second byte
1230Sstevel@tonic-gate	stb	%g1, [%l0 + 1]		! store it
1240Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
1250Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
1260Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
1270Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null bytes
1280Sstevel@tonic-gate	srl	%i1, 8, %g1		! third byte
1290Sstevel@tonic-gate	stb	%g1, [%l0 + 2]		! store it
1300Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
1310Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
1320Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
1330Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null bytes
1340Sstevel@tonic-gate	stb	%i1, [%l0 + 3]		! store fourth byte
1350Sstevel@tonic-gate	addcc	%i4, 8, %g0		! number of pad bytes < 8 ?
1360Sstevel@tonic-gate	bcs	.bytepad		! yes, do simple byte wise fill
1370Sstevel@tonic-gate	add	%l0, 4, %l0		! dst += 4
1380Sstevel@tonic-gate	andcc	%l0, 3, %l1		! dst offset relative to word boundary
1390Sstevel@tonic-gate	bz	.fillaligned		! dst already word aligned
1400Sstevel@tonic-gate
1410Sstevel@tonic-gate	! here there is a least one more byte to zero out: otherwise we would
1420Sstevel@tonic-gate	! have exited through label .lastword
1430Sstevel@tonic-gate
1440Sstevel@tonic-gate	sub	%l1, 4, %l1		! bytes to align dst to word boundary
1450Sstevel@tonic-gate.makealigned:
1460Sstevel@tonic-gate	stb	%g0, [%l0]		! dst[] = 0
1470Sstevel@tonic-gate	addcc	%i4, 1, %i4		! n--
1480Sstevel@tonic-gate	bz	.done			! n == 0, we are done
1490Sstevel@tonic-gate	addcc	%l1, 1, %l1		! any more byte needed to align
1500Sstevel@tonic-gate	bnz	.makealigned		! yup, pad another byte
1510Sstevel@tonic-gate	add	%l0, 1, %l0		! dst++
1520Sstevel@tonic-gate	nop				! pad to align copy loop below
1530Sstevel@tonic-gate
1540Sstevel@tonic-gate	! here we know that there at least another 4 bytes to pad, since
1550Sstevel@tonic-gate	! we don't get here unless there were >= 8 bytes to pad to begin
1560Sstevel@tonic-gate	! with, and we have padded at most 3 bytes suring dst aligning
1570Sstevel@tonic-gate
1580Sstevel@tonic-gate.fillaligned:
1590Sstevel@tonic-gate	add	%i4, 3, %i2		! round up to next word boundary
1600Sstevel@tonic-gate	and	%i2, -4, %l1		! pointer to next word boundary
1610Sstevel@tonic-gate	and	%i2, 4, %i2		! word count odd ? 4 : 0
1620Sstevel@tonic-gate	stw	%g0, [%l0]		! store first word
1630Sstevel@tonic-gate	addcc	%l1, %i2, %l1		! dword count == 1 ?
1640Sstevel@tonic-gate	add	%i4, %i2, %i4		! if word count odd, n -= 4
1650Sstevel@tonic-gate	bz	.bytepad		! if word count == 1, pad bytes left
1660Sstevel@tonic-gate	add	%l0, %i2, %l0		! bump dst if word count odd
1670Sstevel@tonic-gate
1680Sstevel@tonic-gate.fillword:
1690Sstevel@tonic-gate	addcc	%l1, 8, %l1		! count -= 8
1700Sstevel@tonic-gate	stw	%g0, [%l0]		! dst[n] = 0
1710Sstevel@tonic-gate	stw	%g0, [%l0 + 4]		! dst[n+4] = 0
1720Sstevel@tonic-gate	add	%l0, 8, %l0		! dst += 8
1730Sstevel@tonic-gate	bcc	.fillword		! fill words until count == 0
1740Sstevel@tonic-gate	addcc	%i4, 8, %i4		! n -= 8
1750Sstevel@tonic-gate	bz	.done			! if n == 0, we are done
1760Sstevel@tonic-gate	.empty
1770Sstevel@tonic-gate
1780Sstevel@tonic-gate.bytepad:
1790Sstevel@tonic-gate	and	%i4, 1, %i2		! byte count odd ? 1 : 0
1800Sstevel@tonic-gate	stb	%g0, [%l0]		! store first byte
1810Sstevel@tonic-gate	addcc	%i4, %i2, %i4		! byte count == 1 ?
1820Sstevel@tonic-gate	bz	.done			! yup, we are done
1830Sstevel@tonic-gate	add	%l0, %i2, %l0		! bump pointer if odd
1840Sstevel@tonic-gate
1850Sstevel@tonic-gate.fillbyte:
1860Sstevel@tonic-gate	addcc	%i4, 2, %i4		! n -= 2
1870Sstevel@tonic-gate	stb	%g0, [%l0]		! dst[n] = 0
1880Sstevel@tonic-gate	stb	%g0, [%l0 + 1]		! dst[n+1] = 0
1890Sstevel@tonic-gate	bnz	.fillbyte		! fill until n == 0
1900Sstevel@tonic-gate	add	%l0, 2, %l0		! dst += 2
1910Sstevel@tonic-gate
1920Sstevel@tonic-gate.done:
1930Sstevel@tonic-gate	ret				! done
1940Sstevel@tonic-gate	restore	%i0, %g0, %o0		! restore reg window, return dst
1950Sstevel@tonic-gate
1960Sstevel@tonic-gate	! this is the last word. It may contain null bytes. store bytes
1970Sstevel@tonic-gate	! until n == 0. if null byte encountered, continue
1980Sstevel@tonic-gate
1990Sstevel@tonic-gate.lastword:
2000Sstevel@tonic-gate	sub	%i4, 4, %i4		! undo counter pre-increment
2010Sstevel@tonic-gate	add	%i2, 4, %i2		! adjust dst for counter un-bumping
2020Sstevel@tonic-gate
2030Sstevel@tonic-gate	srl	%i1, 24, %g1		! first byte
2040Sstevel@tonic-gate	stb	%g1, [%i2 + %i4]	! store it
2050Sstevel@tonic-gate	inccc	%i4			! n--
2060Sstevel@tonic-gate	bz	.done			! if n == 0, we're done
2070Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
2080Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
2090Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null
2100Sstevel@tonic-gate	srl	%i1, 16, %g1		! second byte
2110Sstevel@tonic-gate	stb	%g1, [%i2 + %i4]	! store it
2120Sstevel@tonic-gate	inccc	%i4			! n--
2130Sstevel@tonic-gate	bz	.done			! if n == 0, we're done
2140Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
2150Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
2160Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
2170Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null
2180Sstevel@tonic-gate	srl	%i1, 8, %g1		! third byte
2190Sstevel@tonic-gate	stb	%g1, [%i2 + %i4]	! store it
2200Sstevel@tonic-gate	inccc	%i4			! n--
2210Sstevel@tonic-gate	bz	.done			! if n == 0, we're done
2220Sstevel@tonic-gate	and	%g1, 0xff, %g1		! isolate byte
2230Sstevel@tonic-gate	sub	%g1, 1, %g1		! byte == 0 ? -1 : byte - 1
2240Sstevel@tonic-gate	sra	%g1, 31, %g1		! byte == 0 ? -1 : 0
2250Sstevel@tonic-gate	andn	%i1, %g1, %i1		! if byte == 0, start padding with null
2260Sstevel@tonic-gate	ba	.done			! here n must be zero, we are done
2270Sstevel@tonic-gate	stb	%i1, [%i2 + %i4]	! store fourth byte
2280Sstevel@tonic-gate
2290Sstevel@tonic-gate.dstnotaligned:
2300Sstevel@tonic-gate	cmp	%g1, 2			! dst half word aligned?
2310Sstevel@tonic-gate	be	.storehalfword2		! yup, store half word at a time
2320Sstevel@tonic-gate	.empty
2330Sstevel@tonic-gate.storebyte:
2340Sstevel@tonic-gate	lduw	[%i3 + %i4], %i1	! x = src[]
2350Sstevel@tonic-gate	addcc	%i4, 4, %i4		! src += 4, dst += 4, n -= 4
2360Sstevel@tonic-gate	bcs	.lastword		! if counter wraps, last word
2370Sstevel@tonic-gate	andn	%i5, %i1, %g1		! ~x & 0x80808080
2380Sstevel@tonic-gate	sub	%i1, %l1, %l0		! x - 0x01010101
2390Sstevel@tonic-gate	andcc	%l0, %g1, %g0		! ((x - 0x01010101) & ~x & 0x80808080)
2400Sstevel@tonic-gate	bnz	.zerobyte		! end of src found, may need to pad
2410Sstevel@tonic-gate	add	%i2, %i4, %l0		! dst (in pointer form)
2420Sstevel@tonic-gate	srl	%i1, 24, %g1		! %g1<7:0> = 1st byte; half-word aligned now
2430Sstevel@tonic-gate	stb	%g1, [%l0]		! store first byte
2440Sstevel@tonic-gate	srl	%i1, 8, %g1		! %g1<15:0> = bytes 2, 3
2450Sstevel@tonic-gate	sth	%g1, [%l0 + 1]		! store bytes 2, 3
2460Sstevel@tonic-gate	ba	.storebyte		! next word
2470Sstevel@tonic-gate	stb	%i1, [%l0 + 3]		! store fourth byte
2480Sstevel@tonic-gate	nop
2490Sstevel@tonic-gate	nop
2500Sstevel@tonic-gate
2510Sstevel@tonic-gate.storehalfword:
2520Sstevel@tonic-gate	lduw	[%i3 + %i4], %i1	! x = src[]
2530Sstevel@tonic-gate.storehalfword2:
2540Sstevel@tonic-gate	addcc	%i4, 4, %i4		! src += 4, dst += 4, n -= 4
2550Sstevel@tonic-gate	bcs	.lastword		! if counter wraps, last word
2560Sstevel@tonic-gate	andn	%i5, %i1, %g1		! ~x & 0x80808080
2570Sstevel@tonic-gate	sub	%i1, %l1, %l0		! x - 0x01010101
2580Sstevel@tonic-gate	andcc	%l0, %g1, %g0		! ((x -0x01010101) & ~x & 0x8080808080)
2590Sstevel@tonic-gate	bnz	.zerobyte		! x has zero byte, handle end cases
2600Sstevel@tonic-gate	add	%i2, %i4, %l0		! dst (in pointer form)
2610Sstevel@tonic-gate	srl	%i1, 16, %g1		! %g1<15:0> = bytes 1, 2
2620Sstevel@tonic-gate	sth	%g1, [%l0]		! store bytes 1, 2
2630Sstevel@tonic-gate	ba	.storehalfword		! next dword
2640Sstevel@tonic-gate	sth	%i1, [%l0 + 2]		! store bytes 3, 4
2650Sstevel@tonic-gate
2660Sstevel@tonic-gate.shortcpy:
2670Sstevel@tonic-gate	ldub	[%o3 + %o4], %o5	! src[]
2680Sstevel@tonic-gate	stb	%o5, [%o2 + %o4]	! dst[] = src[]
2690Sstevel@tonic-gate	inccc	%o4			! src++, dst++, n--
2700Sstevel@tonic-gate	bz	.doneshort		! if n == 0, done
2710Sstevel@tonic-gate	tst	%o5			! src[] == 0 ?
2720Sstevel@tonic-gate	bnz,a	.shortcpy		! nope, next byte
2730Sstevel@tonic-gate	nop				! empty delay slot
2740Sstevel@tonic-gate
2750Sstevel@tonic-gate.padbyte:
2760Sstevel@tonic-gate	stb	%g0, [%o2 + %o4]	! dst[] = 0
2770Sstevel@tonic-gate.padbyte2:
2780Sstevel@tonic-gate	addcc	%o4, 1, %o4		! dst++, n--
2790Sstevel@tonic-gate	bnz,a	.padbyte2		! if n != 0, next byte
2800Sstevel@tonic-gate	stb	%g0, [%o2 + %o4]	! dst[] = 0
2810Sstevel@tonic-gate	nop				! align label below to 16-byte boundary
2820Sstevel@tonic-gate
2830Sstevel@tonic-gate.doneshort:
2840Sstevel@tonic-gate	retl				! return from leaf
2850Sstevel@tonic-gate	nop				! empty delay slot
2860Sstevel@tonic-gate	SET_SIZE(strncpy)
287