xref: /linux/arch/csky/abiv2/memcpy.S (revision 4e0ae876f77bc01a7e77724dea57b4b82bd53244)
1/* SPDX-License-Identifier: GPL-2.0 */
2// Copyright (C) 2018 Hangzhou C-SKY Microsystems co.,ltd.
3
4#include <linux/linkage.h>
5#include "sysdep.h"
6
7ENTRY(__memcpy)
8ENTRY(memcpy)
9	/* Test if len less than 4 bytes.  */
10	mov	r12, r0
11	cmplti	r2, 4
12	bt	.L_copy_by_byte
13
14	andi	r13, r0, 3
15	movi	r19, 4
16	/* Test if dest is not 4 bytes aligned.  */
17	bnez	r13, .L_dest_not_aligned
18
19/* Hardware can handle unaligned access directly.  */
20.L_dest_aligned:
21	/* If dest is aligned, then copy.  */
22	zext	r18, r2, 31, 4
23
24	/* Test if len less than 16 bytes.  */
25	bez	r18, .L_len_less_16bytes
26	movi	r19, 0
27
28	LABLE_ALIGN
29.L_len_larger_16bytes:
30#if defined(__CK860__)
31	ldw	r3, (r1, 0)
32	stw	r3, (r0, 0)
33	ldw	r3, (r1, 4)
34	stw	r3, (r0, 4)
35	ldw	r3, (r1, 8)
36	stw	r3, (r0, 8)
37	ldw	r3, (r1, 12)
38	addi	r1, 16
39	stw	r3, (r0, 12)
40	addi	r0, 16
41#else
42	ldw	r20, (r1, 0)
43	ldw	r21, (r1, 4)
44	ldw	r22, (r1, 8)
45	ldw	r23, (r1, 12)
46	stw	r20, (r0, 0)
47	stw	r21, (r0, 4)
48	stw	r22, (r0, 8)
49	stw	r23, (r0, 12)
50	PRE_BNEZAD (r18)
51	addi	r1, 16
52	addi	r0, 16
53#endif
54	BNEZAD (r18, .L_len_larger_16bytes)
55
56.L_len_less_16bytes:
57	zext	r18, r2, 3, 2
58	bez	r18, .L_copy_by_byte
59.L_len_less_16bytes_loop:
60	ldw	r3, (r1, 0)
61	PRE_BNEZAD (r18)
62	addi	r1, 4
63	stw	r3, (r0, 0)
64	addi	r0, 4
65	BNEZAD (r18, .L_len_less_16bytes_loop)
66
67/* Test if len less than 4 bytes.  */
68.L_copy_by_byte:
69	zext	r18, r2, 1, 0
70	bez	r18, .L_return
71.L_copy_by_byte_loop:
72	ldb	r3, (r1, 0)
73	PRE_BNEZAD (r18)
74	addi	r1, 1
75	stb	r3, (r0, 0)
76	addi	r0, 1
77	BNEZAD (r18, .L_copy_by_byte_loop)
78
79.L_return:
80	mov	r0, r12
81	rts
82
83/*
84 * If dest is not aligned, just copying some bytes makes the
85 * dest align.
86 */
87.L_dest_not_aligned:
88	sub	r13, r19, r13
89	sub	r2, r13
90
91/* Makes the dest align.  */
92.L_dest_not_aligned_loop:
93	ldb	r3, (r1, 0)
94	PRE_BNEZAD (r13)
95	addi	r1, 1
96	stb	r3, (r0, 0)
97	addi	r0, 1
98	BNEZAD (r13, .L_dest_not_aligned_loop)
99	cmplti	r2, 4
100	bt	.L_copy_by_byte
101
102	/* Check whether the src is aligned.  */
103	jbr	.L_dest_aligned
104ENDPROC(__memcpy)
105