xref: /illumos-gate/usr/src/lib/libc/sparc/gen/memcmp.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/*
23*5d9d9091SRichard Lowe * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24*5d9d9091SRichard Lowe * Use is subject to license terms.
25*5d9d9091SRichard Lowe */
26*5d9d9091SRichard Lowe
27*5d9d9091SRichard Lowe	.file	"memcmp.s"
28*5d9d9091SRichard Lowe
29*5d9d9091SRichard Lowe/*
30*5d9d9091SRichard Lowe * memcmp(s1, s2, len)
31*5d9d9091SRichard Lowe *
32*5d9d9091SRichard Lowe * Compare n bytes:  s1>s2: >0  s1==s2: 0  s1<s2: <0
33*5d9d9091SRichard Lowe *
34*5d9d9091SRichard Lowe * Fast assembler language version of the following C-program for memcmp
35*5d9d9091SRichard Lowe * which represents the `standard' for the C-library.
36*5d9d9091SRichard Lowe *
37*5d9d9091SRichard Lowe *	int
38*5d9d9091SRichard Lowe *	memcmp(const void *s1, const void *s2, size_t n)
39*5d9d9091SRichard Lowe *	{
40*5d9d9091SRichard Lowe *		if (s1 != s2 && n != 0) {
41*5d9d9091SRichard Lowe *			const char *ps1 = s1;
42*5d9d9091SRichard Lowe *			const char *ps2 = s2;
43*5d9d9091SRichard Lowe *			do {
44*5d9d9091SRichard Lowe *				if (*ps1++ != *ps2++)
45*5d9d9091SRichard Lowe *					return (ps1[-1] - ps2[-1]);
46*5d9d9091SRichard Lowe *			} while (--n != 0);
47*5d9d9091SRichard Lowe *		}
48*5d9d9091SRichard Lowe *		return (NULL);
49*5d9d9091SRichard Lowe *	}
50*5d9d9091SRichard Lowe */
51*5d9d9091SRichard Lowe
52*5d9d9091SRichard Lowe#include <sys/asm_linkage.h>
53*5d9d9091SRichard Lowe
54*5d9d9091SRichard Lowe	ANSI_PRAGMA_WEAK(memcmp,function)
55*5d9d9091SRichard Lowe
56*5d9d9091SRichard Lowe	ENTRY(memcmp)
57*5d9d9091SRichard Lowe	st	%g2, [%sp + 68]		! g2 must be restored before retl
58*5d9d9091SRichard Lowe	cmp	%o0, %o1		! s1 == s2?
59*5d9d9091SRichard Lowe	be	.cmpeq
60*5d9d9091SRichard Lowe	cmp	%o2, 17
61*5d9d9091SRichard Lowe	bleu,a	.cmpbyt			! for small counts go do bytes
62*5d9d9091SRichard Lowe	sub	%o1, %o0, %o1
63*5d9d9091SRichard Lowe
64*5d9d9091SRichard Lowe	andcc	%o0, 3, %o3		! is s1 aligned?
65*5d9d9091SRichard Lowe	bz,a	.iss2			! if so go check s2
66*5d9d9091SRichard Lowe	andcc	%o1, 3, %o4		! is s2 aligned?
67*5d9d9091SRichard Lowe	cmp	%o3, 2
68*5d9d9091SRichard Lowe	be	.algn2
69*5d9d9091SRichard Lowe	cmp	%o3, 3
70*5d9d9091SRichard Lowe
71*5d9d9091SRichard Lowe.algn1:	ldub	[%o0], %o4		! cmp one byte
72*5d9d9091SRichard Lowe	inc	%o0
73*5d9d9091SRichard Lowe	ldub	[%o1], %o5
74*5d9d9091SRichard Lowe	inc	%o1
75*5d9d9091SRichard Lowe	dec	%o2
76*5d9d9091SRichard Lowe	be	.algn3
77*5d9d9091SRichard Lowe	cmp	%o4, %o5
78*5d9d9091SRichard Lowe	be	.algn2
79*5d9d9091SRichard Lowe	nop
80*5d9d9091SRichard Lowe	b,a	.noteq
81*5d9d9091SRichard Lowe
82*5d9d9091SRichard Lowe.algn2:	lduh	[%o0], %o4
83*5d9d9091SRichard Lowe	inc	2, %o0
84*5d9d9091SRichard Lowe	ldub	[%o1], %o5
85*5d9d9091SRichard Lowe	inc	1, %o1
86*5d9d9091SRichard Lowe	srl	%o4, 8, %o3
87*5d9d9091SRichard Lowe	cmp	%o3, %o5
88*5d9d9091SRichard Lowe	be,a	1f
89*5d9d9091SRichard Lowe	ldub	[%o1], %o5		! delay slot, get next byte from s2
90*5d9d9091SRichard Lowe	b	.noteq
91*5d9d9091SRichard Lowe	mov	%o3, %o4		! delay slot, move *s1 to %o4
92*5d9d9091SRichard Lowe1:	inc	%o1
93*5d9d9091SRichard Lowe	dec	2, %o2
94*5d9d9091SRichard Lowe	and	%o4, 0xff, %o4
95*5d9d9091SRichard Lowe	cmp	%o4, %o5
96*5d9d9091SRichard Lowe.algn3:	be,a	.iss2
97*5d9d9091SRichard Lowe	andcc	%o1, 3, %o4		! delay slot, is s2 aligned?
98*5d9d9091SRichard Lowe	b,a	.noteq
99*5d9d9091SRichard Lowe
100*5d9d9091SRichard Lowe.cmpbyt:b	.bytcmp
101*5d9d9091SRichard Lowe	deccc	%o2
102*5d9d9091SRichard Lowe1:	ldub	[%o0 + %o1], %o5	! byte compare loop
103*5d9d9091SRichard Lowe	inc	%o0
104*5d9d9091SRichard Lowe	cmp	%o4, %o5
105*5d9d9091SRichard Lowe	be,a	.bytcmp
106*5d9d9091SRichard Lowe	deccc	%o2			! delay slot, compare count (len)
107*5d9d9091SRichard Lowe	b,a	.noteq
108*5d9d9091SRichard Lowe.bytcmp:bgeu,a	1b
109*5d9d9091SRichard Lowe	ldub	[%o0], %o4
110*5d9d9091SRichard Lowe.cmpeq:	ld	[%sp + 68], %g2
111*5d9d9091SRichard Lowe	retl				! strings compare equal
112*5d9d9091SRichard Lowe	clr	%o0
113*5d9d9091SRichard Lowe
114*5d9d9091SRichard Lowe.noteq_word:				! words aren't equal. find unequal byte
115*5d9d9091SRichard Lowe	srl	%o4, 24, %o1		! first byte
116*5d9d9091SRichard Lowe	srl	%o5, 24, %o2
117*5d9d9091SRichard Lowe	cmp	%o1, %o2
118*5d9d9091SRichard Lowe	bne	1f
119*5d9d9091SRichard Lowe	sll	%o4, 8, %o4
120*5d9d9091SRichard Lowe	sll	%o5, 8, %o5
121*5d9d9091SRichard Lowe	srl	%o4, 24, %o1
122*5d9d9091SRichard Lowe	srl	%o5, 24, %o2
123*5d9d9091SRichard Lowe	cmp	%o1, %o2
124*5d9d9091SRichard Lowe	bne	1f
125*5d9d9091SRichard Lowe	sll	%o4, 8, %o4
126*5d9d9091SRichard Lowe	sll	%o5, 8, %o5
127*5d9d9091SRichard Lowe	srl	%o4, 24, %o1
128*5d9d9091SRichard Lowe	srl	%o5, 24, %o2
129*5d9d9091SRichard Lowe	cmp	%o1, %o2
130*5d9d9091SRichard Lowe	bne	1f
131*5d9d9091SRichard Lowe	sll	%o4, 8, %o4
132*5d9d9091SRichard Lowe	sll	%o5, 8, %o5
133*5d9d9091SRichard Lowe	srl	%o4, 24, %o1
134*5d9d9091SRichard Lowe	srl	%o5, 24, %o2
135*5d9d9091SRichard Lowe1:
136*5d9d9091SRichard Lowe	ld	[%sp + 68], %g2
137*5d9d9091SRichard Lowe	retl
138*5d9d9091SRichard Lowe	sub	%o1, %o2, %o0		! delay slot
139*5d9d9091SRichard Lowe
140*5d9d9091SRichard Lowe.noteq:
141*5d9d9091SRichard Lowe	ld	[%sp + 68], %g2
142*5d9d9091SRichard Lowe	retl				! strings aren't equal
143*5d9d9091SRichard Lowe	sub	%o4, %o5, %o0		! delay slot, return(*s1 - *s2)
144*5d9d9091SRichard Lowe
145*5d9d9091SRichard Lowe.iss2:	andn	%o2, 3, %o3		! count of aligned bytes
146*5d9d9091SRichard Lowe	and	%o2, 3, %o2		! remaining bytes
147*5d9d9091SRichard Lowe	bz	.w4cmp			! if s2 word aligned, compare words
148*5d9d9091SRichard Lowe	cmp	%o4, 2
149*5d9d9091SRichard Lowe	be	.w2cmp			! s2 half aligned
150*5d9d9091SRichard Lowe	cmp	%o4, 1
151*5d9d9091SRichard Lowe
152*5d9d9091SRichard Lowe.w3cmp:
153*5d9d9091SRichard Lowe	dec	4, %o3			! avoid reading beyond the last byte
154*5d9d9091SRichard Lowe	inc	4, %o2
155*5d9d9091SRichard Lowe	ldub	[%o1], %g1		! read a byte to align for word reads
156*5d9d9091SRichard Lowe	inc	1, %o1
157*5d9d9091SRichard Lowe	be	.w1cmp			! aligned to 1 or 3 bytes
158*5d9d9091SRichard Lowe	sll	%g1, 24, %o5
159*5d9d9091SRichard Lowe
160*5d9d9091SRichard Lowe	sub	%o1, %o0, %o1
161*5d9d9091SRichard Lowe2:	ld	[%o0 + %o1], %g1
162*5d9d9091SRichard Lowe	ld	[%o0], %o4
163*5d9d9091SRichard Lowe	inc	4, %o0
164*5d9d9091SRichard Lowe	srl	%g1, 8, %g2		! merge with the other half
165*5d9d9091SRichard Lowe	or	%g2, %o5, %o5
166*5d9d9091SRichard Lowe	cmp	%o4, %o5
167*5d9d9091SRichard Lowe	bne	.noteq_word
168*5d9d9091SRichard Lowe	deccc	4, %o3
169*5d9d9091SRichard Lowe	bnz	2b
170*5d9d9091SRichard Lowe	sll	%g1, 24, %o5
171*5d9d9091SRichard Lowe	sub	%o1, 1, %o1		! used 3 bytes of the last word read
172*5d9d9091SRichard Lowe	b	.bytcmp
173*5d9d9091SRichard Lowe	deccc	%o2
174*5d9d9091SRichard Lowe
175*5d9d9091SRichard Lowe.w1cmp:
176*5d9d9091SRichard Lowe	dec	4, %o3			! avoid reading beyond the last byte
177*5d9d9091SRichard Lowe	inc	4, %o2
178*5d9d9091SRichard Lowe	lduh	[%o1], %g1		! read 3 bytes to word align
179*5d9d9091SRichard Lowe	inc	2, %o1
180*5d9d9091SRichard Lowe	sll	%g1, 8, %g2
181*5d9d9091SRichard Lowe	or	%o5, %g2, %o5
182*5d9d9091SRichard Lowe
183*5d9d9091SRichard Lowe	sub	%o1, %o0, %o1
184*5d9d9091SRichard Lowe3:	ld	[%o0 + %o1], %g1
185*5d9d9091SRichard Lowe	ld	[%o0], %o4
186*5d9d9091SRichard Lowe	inc	4, %o0
187*5d9d9091SRichard Lowe	srl	%g1, 24, %g2		! merge with the other half
188*5d9d9091SRichard Lowe	or	%g2, %o5, %o5
189*5d9d9091SRichard Lowe	cmp	%o4, %o5
190*5d9d9091SRichard Lowe	bne	.noteq_word
191*5d9d9091SRichard Lowe	deccc	4, %o3
192*5d9d9091SRichard Lowe	bnz	3b
193*5d9d9091SRichard Lowe	sll	%g1, 8, %o5
194*5d9d9091SRichard Lowe	sub	%o1, 3, %o1		! used 1 byte of the last word read
195*5d9d9091SRichard Lowe	b	.bytcmp
196*5d9d9091SRichard Lowe	deccc	%o2
197*5d9d9091SRichard Lowe
198*5d9d9091SRichard Lowe.w2cmp:
199*5d9d9091SRichard Lowe	dec	4, %o3			! avoid reading beyond the last byte
200*5d9d9091SRichard Lowe	inc	4, %o2
201*5d9d9091SRichard Lowe	lduh	[%o1], %g1		! read a halfword to align s2
202*5d9d9091SRichard Lowe	inc	2, %o1
203*5d9d9091SRichard Lowe	sll	%g1, 16, %o5
204*5d9d9091SRichard Lowe
205*5d9d9091SRichard Lowe	sub	%o1, %o0, %o1
206*5d9d9091SRichard Lowe4:	ld	[%o0 + %o1], %g1	! read a word from s2
207*5d9d9091SRichard Lowe	ld	[%o0], %o4		! read a word from s1
208*5d9d9091SRichard Lowe	inc	4, %o0
209*5d9d9091SRichard Lowe	srl	%g1, 16, %g2		! merge with the other half
210*5d9d9091SRichard Lowe	or	%g2, %o5, %o5
211*5d9d9091SRichard Lowe	cmp	%o4, %o5
212*5d9d9091SRichard Lowe	bne	.noteq_word
213*5d9d9091SRichard Lowe	deccc	4, %o3
214*5d9d9091SRichard Lowe	bnz	4b
215*5d9d9091SRichard Lowe	sll	%g1, 16, %o5
216*5d9d9091SRichard Lowe	sub	%o1, 2, %o1		! only used half of the last read word
217*5d9d9091SRichard Lowe	b	.bytcmp
218*5d9d9091SRichard Lowe	deccc	%o2
219*5d9d9091SRichard Lowe
220*5d9d9091SRichard Lowe.w4cmp:
221*5d9d9091SRichard Lowe	sub	%o1, %o0, %o1
222*5d9d9091SRichard Lowe	ld	[%o0 + %o1], %o5
223*5d9d9091SRichard Lowe5:	ld	[%o0], %o4
224*5d9d9091SRichard Lowe	inc	4, %o0
225*5d9d9091SRichard Lowe	cmp	%o4, %o5
226*5d9d9091SRichard Lowe	bne	.noteq_word
227*5d9d9091SRichard Lowe	deccc	4, %o3
228*5d9d9091SRichard Lowe	bnz,a	5b
229*5d9d9091SRichard Lowe	ld	[%o0 + %o1], %o5
230*5d9d9091SRichard Lowe	b	.bytcmp			! compare remaining bytes, if any
231*5d9d9091SRichard Lowe	deccc	%o2
232*5d9d9091SRichard Lowe
233*5d9d9091SRichard Lowe	SET_SIZE(memcmp)
234