xref: /linux/arch/csky/abiv2/memcmp.S (revision 3eb66e91a25497065c5322b1268cbc3953642227)
1*c5af58b7SGuo Ren/* SPDX-License-Identifier: GPL-2.0 */
2*c5af58b7SGuo Ren// Copyright (C) 2018 Hangzhou C-SKY Microsystems co.,ltd.
3*c5af58b7SGuo Ren
4*c5af58b7SGuo Ren#include <linux/linkage.h>
5*c5af58b7SGuo Ren#include "sysdep.h"
6*c5af58b7SGuo Ren
7*c5af58b7SGuo RenENTRY(memcmp)
8*c5af58b7SGuo Ren	/* Test if len less than 4 bytes.  */
9*c5af58b7SGuo Ren	mov	r3, r0
10*c5af58b7SGuo Ren	movi	r0, 0
11*c5af58b7SGuo Ren	mov	r12, r4
12*c5af58b7SGuo Ren	cmplti	r2, 4
13*c5af58b7SGuo Ren	bt	.L_compare_by_byte
14*c5af58b7SGuo Ren
15*c5af58b7SGuo Ren	andi	r13, r0, 3
16*c5af58b7SGuo Ren	movi	r19, 4
17*c5af58b7SGuo Ren
18*c5af58b7SGuo Ren	/* Test if s1 is not 4 bytes aligned.  */
19*c5af58b7SGuo Ren	bnez	r13, .L_s1_not_aligned
20*c5af58b7SGuo Ren
21*c5af58b7SGuo Ren	LABLE_ALIGN
22*c5af58b7SGuo Ren.L_s1_aligned:
23*c5af58b7SGuo Ren	/* If dest is aligned, then copy.  */
24*c5af58b7SGuo Ren	zext	r18, r2, 31, 4
25*c5af58b7SGuo Ren	/* Test if len less than 16 bytes.  */
26*c5af58b7SGuo Ren	bez	r18, .L_compare_by_word
27*c5af58b7SGuo Ren
28*c5af58b7SGuo Ren.L_compare_by_4word:
29*c5af58b7SGuo Ren	/* If aligned, load word each time.  */
30*c5af58b7SGuo Ren	ldw	r20, (r3, 0)
31*c5af58b7SGuo Ren	ldw	r21, (r1, 0)
32*c5af58b7SGuo Ren	/* If s1[i] != s2[i], goto .L_byte_check.  */
33*c5af58b7SGuo Ren	cmpne	r20, r21
34*c5af58b7SGuo Ren	bt	.L_byte_check
35*c5af58b7SGuo Ren
36*c5af58b7SGuo Ren	ldw	r20, (r3, 4)
37*c5af58b7SGuo Ren	ldw	r21, (r1, 4)
38*c5af58b7SGuo Ren	cmpne	r20, r21
39*c5af58b7SGuo Ren	bt	.L_byte_check
40*c5af58b7SGuo Ren
41*c5af58b7SGuo Ren	ldw	r20, (r3, 8)
42*c5af58b7SGuo Ren	ldw	r21, (r1, 8)
43*c5af58b7SGuo Ren	cmpne	r20, r21
44*c5af58b7SGuo Ren	bt	.L_byte_check
45*c5af58b7SGuo Ren
46*c5af58b7SGuo Ren	ldw	r20, (r3, 12)
47*c5af58b7SGuo Ren	ldw	r21, (r1, 12)
48*c5af58b7SGuo Ren	cmpne	r20, r21
49*c5af58b7SGuo Ren	bt	.L_byte_check
50*c5af58b7SGuo Ren
51*c5af58b7SGuo Ren	PRE_BNEZAD (r18)
52*c5af58b7SGuo Ren	addi	a3, 16
53*c5af58b7SGuo Ren	addi	a1, 16
54*c5af58b7SGuo Ren
55*c5af58b7SGuo Ren	BNEZAD (r18, .L_compare_by_4word)
56*c5af58b7SGuo Ren
57*c5af58b7SGuo Ren.L_compare_by_word:
58*c5af58b7SGuo Ren	zext	r18, r2, 3, 2
59*c5af58b7SGuo Ren	bez	r18, .L_compare_by_byte
60*c5af58b7SGuo Ren.L_compare_by_word_loop:
61*c5af58b7SGuo Ren	ldw	r20, (r3, 0)
62*c5af58b7SGuo Ren	ldw	r21, (r1, 0)
63*c5af58b7SGuo Ren	addi	r3, 4
64*c5af58b7SGuo Ren	PRE_BNEZAD (r18)
65*c5af58b7SGuo Ren	cmpne	r20, r21
66*c5af58b7SGuo Ren	addi    r1, 4
67*c5af58b7SGuo Ren	bt	.L_byte_check
68*c5af58b7SGuo Ren	BNEZAD (r18, .L_compare_by_word_loop)
69*c5af58b7SGuo Ren
70*c5af58b7SGuo Ren.L_compare_by_byte:
71*c5af58b7SGuo Ren        zext    r18, r2, 1, 0
72*c5af58b7SGuo Ren        bez     r18, .L_return
73*c5af58b7SGuo Ren.L_compare_by_byte_loop:
74*c5af58b7SGuo Ren        ldb     r0, (r3, 0)
75*c5af58b7SGuo Ren        ldb     r4, (r1, 0)
76*c5af58b7SGuo Ren        addi    r3, 1
77*c5af58b7SGuo Ren        subu    r0, r4
78*c5af58b7SGuo Ren        PRE_BNEZAD (r18)
79*c5af58b7SGuo Ren        addi    r1, 1
80*c5af58b7SGuo Ren        bnez    r0, .L_return
81*c5af58b7SGuo Ren        BNEZAD (r18, .L_compare_by_byte_loop)
82*c5af58b7SGuo Ren
83*c5af58b7SGuo Ren.L_return:
84*c5af58b7SGuo Ren        mov     r4, r12
85*c5af58b7SGuo Ren        rts
86*c5af58b7SGuo Ren
87*c5af58b7SGuo Ren# ifdef __CSKYBE__
88*c5af58b7SGuo Ren/* d[i] != s[i] in word, so we check byte 0.  */
89*c5af58b7SGuo Ren.L_byte_check:
90*c5af58b7SGuo Ren        xtrb0   r0, r20
91*c5af58b7SGuo Ren        xtrb0   r2, r21
92*c5af58b7SGuo Ren        subu    r0, r2
93*c5af58b7SGuo Ren        bnez    r0, .L_return
94*c5af58b7SGuo Ren
95*c5af58b7SGuo Ren        /* check byte 1 */
96*c5af58b7SGuo Ren        xtrb1   r0, r20
97*c5af58b7SGuo Ren        xtrb1   r2, r21
98*c5af58b7SGuo Ren        subu    r0, r2
99*c5af58b7SGuo Ren        bnez    r0, .L_return
100*c5af58b7SGuo Ren
101*c5af58b7SGuo Ren        /* check byte 2 */
102*c5af58b7SGuo Ren        xtrb2   r0, r20
103*c5af58b7SGuo Ren        xtrb2   r2, r21
104*c5af58b7SGuo Ren        subu    r0, r2
105*c5af58b7SGuo Ren        bnez    r0, .L_return
106*c5af58b7SGuo Ren
107*c5af58b7SGuo Ren        /* check byte 3 */
108*c5af58b7SGuo Ren        xtrb3   r0, r20
109*c5af58b7SGuo Ren        xtrb3   r2, r21
110*c5af58b7SGuo Ren        subu    r0, r2
111*c5af58b7SGuo Ren# else
112*c5af58b7SGuo Ren/* s1[i] != s2[i] in word, so we check byte 3.  */
113*c5af58b7SGuo Ren.L_byte_check:
114*c5af58b7SGuo Ren	xtrb3	r0, r20
115*c5af58b7SGuo Ren	xtrb3	r2, r21
116*c5af58b7SGuo Ren        subu    r0, r2
117*c5af58b7SGuo Ren        bnez    r0, .L_return
118*c5af58b7SGuo Ren
119*c5af58b7SGuo Ren	/* check byte 2 */
120*c5af58b7SGuo Ren	xtrb2	r0, r20
121*c5af58b7SGuo Ren	xtrb2	r2, r21
122*c5af58b7SGuo Ren        subu    r0, r2
123*c5af58b7SGuo Ren        bnez    r0, .L_return
124*c5af58b7SGuo Ren
125*c5af58b7SGuo Ren	/* check byte 1 */
126*c5af58b7SGuo Ren	xtrb1	r0, r20
127*c5af58b7SGuo Ren	xtrb1	r2, r21
128*c5af58b7SGuo Ren	subu	r0, r2
129*c5af58b7SGuo Ren	bnez    r0, .L_return
130*c5af58b7SGuo Ren
131*c5af58b7SGuo Ren	/* check byte 0 */
132*c5af58b7SGuo Ren	xtrb0	r0, r20
133*c5af58b7SGuo Ren	xtrb0	r2, r21
134*c5af58b7SGuo Ren	subu	r0, r2
135*c5af58b7SGuo Ren	br	.L_return
136*c5af58b7SGuo Ren# endif /* !__CSKYBE__ */
137*c5af58b7SGuo Ren
138*c5af58b7SGuo Ren/* Compare when s1 is not aligned.  */
139*c5af58b7SGuo Ren.L_s1_not_aligned:
140*c5af58b7SGuo Ren	sub	r13, r19, r13
141*c5af58b7SGuo Ren	sub	r2, r13
142*c5af58b7SGuo Ren.L_s1_not_aligned_loop:
143*c5af58b7SGuo Ren	ldb	r0, (r3, 0)
144*c5af58b7SGuo Ren	ldb	r4, (r1, 0)
145*c5af58b7SGuo Ren	addi	r3, 1
146*c5af58b7SGuo Ren	subu	r0, r4
147*c5af58b7SGuo Ren	PRE_BNEZAD (r13)
148*c5af58b7SGuo Ren	addi	r1, 1
149*c5af58b7SGuo Ren	bnez	r0, .L_return
150*c5af58b7SGuo Ren	BNEZAD (r13, .L_s1_not_aligned_loop)
151*c5af58b7SGuo Ren	br	.L_s1_aligned
152*c5af58b7SGuo RenENDPROC(memcmp)
153