xref: /OK3568_Linux_fs/kernel/arch/riscv/lib/memcpy.S (revision 4882a59341e53eb6f0b4789bf948001014eff981)
1*4882a593Smuzhiyun/* SPDX-License-Identifier: GPL-2.0-only */
2*4882a593Smuzhiyun/*
3*4882a593Smuzhiyun * Copyright (C) 2013 Regents of the University of California
4*4882a593Smuzhiyun */
5*4882a593Smuzhiyun
6*4882a593Smuzhiyun#include <linux/linkage.h>
7*4882a593Smuzhiyun#include <asm/asm.h>
8*4882a593Smuzhiyun
9*4882a593Smuzhiyun/* void *memcpy(void *, const void *, size_t) */
10*4882a593SmuzhiyunENTRY(__memcpy)
11*4882a593SmuzhiyunWEAK(memcpy)
12*4882a593Smuzhiyun	move t6, a0  /* Preserve return value */
13*4882a593Smuzhiyun
14*4882a593Smuzhiyun	/* Defer to byte-oriented copy for small sizes */
15*4882a593Smuzhiyun	sltiu a3, a2, 128
16*4882a593Smuzhiyun	bnez a3, 4f
17*4882a593Smuzhiyun	/* Use word-oriented copy only if low-order bits match */
18*4882a593Smuzhiyun	andi a3, t6, SZREG-1
19*4882a593Smuzhiyun	andi a4, a1, SZREG-1
20*4882a593Smuzhiyun	bne a3, a4, 4f
21*4882a593Smuzhiyun
22*4882a593Smuzhiyun	beqz a3, 2f  /* Skip if already aligned */
23*4882a593Smuzhiyun	/*
24*4882a593Smuzhiyun	 * Round to nearest double word-aligned address
25*4882a593Smuzhiyun	 * greater than or equal to start address
26*4882a593Smuzhiyun	 */
27*4882a593Smuzhiyun	andi a3, a1, ~(SZREG-1)
28*4882a593Smuzhiyun	addi a3, a3, SZREG
29*4882a593Smuzhiyun	/* Handle initial misalignment */
30*4882a593Smuzhiyun	sub a4, a3, a1
31*4882a593Smuzhiyun1:
32*4882a593Smuzhiyun	lb a5, 0(a1)
33*4882a593Smuzhiyun	addi a1, a1, 1
34*4882a593Smuzhiyun	sb a5, 0(t6)
35*4882a593Smuzhiyun	addi t6, t6, 1
36*4882a593Smuzhiyun	bltu a1, a3, 1b
37*4882a593Smuzhiyun	sub a2, a2, a4  /* Update count */
38*4882a593Smuzhiyun
39*4882a593Smuzhiyun2:
40*4882a593Smuzhiyun	andi a4, a2, ~((16*SZREG)-1)
41*4882a593Smuzhiyun	beqz a4, 4f
42*4882a593Smuzhiyun	add a3, a1, a4
43*4882a593Smuzhiyun3:
44*4882a593Smuzhiyun	REG_L a4,       0(a1)
45*4882a593Smuzhiyun	REG_L a5,   SZREG(a1)
46*4882a593Smuzhiyun	REG_L a6, 2*SZREG(a1)
47*4882a593Smuzhiyun	REG_L a7, 3*SZREG(a1)
48*4882a593Smuzhiyun	REG_L t0, 4*SZREG(a1)
49*4882a593Smuzhiyun	REG_L t1, 5*SZREG(a1)
50*4882a593Smuzhiyun	REG_L t2, 6*SZREG(a1)
51*4882a593Smuzhiyun	REG_L t3, 7*SZREG(a1)
52*4882a593Smuzhiyun	REG_L t4, 8*SZREG(a1)
53*4882a593Smuzhiyun	REG_L t5, 9*SZREG(a1)
54*4882a593Smuzhiyun	REG_S a4,       0(t6)
55*4882a593Smuzhiyun	REG_S a5,   SZREG(t6)
56*4882a593Smuzhiyun	REG_S a6, 2*SZREG(t6)
57*4882a593Smuzhiyun	REG_S a7, 3*SZREG(t6)
58*4882a593Smuzhiyun	REG_S t0, 4*SZREG(t6)
59*4882a593Smuzhiyun	REG_S t1, 5*SZREG(t6)
60*4882a593Smuzhiyun	REG_S t2, 6*SZREG(t6)
61*4882a593Smuzhiyun	REG_S t3, 7*SZREG(t6)
62*4882a593Smuzhiyun	REG_S t4, 8*SZREG(t6)
63*4882a593Smuzhiyun	REG_S t5, 9*SZREG(t6)
64*4882a593Smuzhiyun	REG_L a4, 10*SZREG(a1)
65*4882a593Smuzhiyun	REG_L a5, 11*SZREG(a1)
66*4882a593Smuzhiyun	REG_L a6, 12*SZREG(a1)
67*4882a593Smuzhiyun	REG_L a7, 13*SZREG(a1)
68*4882a593Smuzhiyun	REG_L t0, 14*SZREG(a1)
69*4882a593Smuzhiyun	REG_L t1, 15*SZREG(a1)
70*4882a593Smuzhiyun	addi a1, a1, 16*SZREG
71*4882a593Smuzhiyun	REG_S a4, 10*SZREG(t6)
72*4882a593Smuzhiyun	REG_S a5, 11*SZREG(t6)
73*4882a593Smuzhiyun	REG_S a6, 12*SZREG(t6)
74*4882a593Smuzhiyun	REG_S a7, 13*SZREG(t6)
75*4882a593Smuzhiyun	REG_S t0, 14*SZREG(t6)
76*4882a593Smuzhiyun	REG_S t1, 15*SZREG(t6)
77*4882a593Smuzhiyun	addi t6, t6, 16*SZREG
78*4882a593Smuzhiyun	bltu a1, a3, 3b
79*4882a593Smuzhiyun	andi a2, a2, (16*SZREG)-1  /* Update count */
80*4882a593Smuzhiyun
81*4882a593Smuzhiyun4:
82*4882a593Smuzhiyun	/* Handle trailing misalignment */
83*4882a593Smuzhiyun	beqz a2, 6f
84*4882a593Smuzhiyun	add a3, a1, a2
85*4882a593Smuzhiyun
86*4882a593Smuzhiyun	/* Use word-oriented copy if co-aligned to word boundary */
87*4882a593Smuzhiyun	or a5, a1, t6
88*4882a593Smuzhiyun	or a5, a5, a3
89*4882a593Smuzhiyun	andi a5, a5, 3
90*4882a593Smuzhiyun	bnez a5, 5f
91*4882a593Smuzhiyun7:
92*4882a593Smuzhiyun	lw a4, 0(a1)
93*4882a593Smuzhiyun	addi a1, a1, 4
94*4882a593Smuzhiyun	sw a4, 0(t6)
95*4882a593Smuzhiyun	addi t6, t6, 4
96*4882a593Smuzhiyun	bltu a1, a3, 7b
97*4882a593Smuzhiyun
98*4882a593Smuzhiyun	ret
99*4882a593Smuzhiyun
100*4882a593Smuzhiyun5:
101*4882a593Smuzhiyun	lb a4, 0(a1)
102*4882a593Smuzhiyun	addi a1, a1, 1
103*4882a593Smuzhiyun	sb a4, 0(t6)
104*4882a593Smuzhiyun	addi t6, t6, 1
105*4882a593Smuzhiyun	bltu a1, a3, 5b
106*4882a593Smuzhiyun6:
107*4882a593Smuzhiyun	ret
108*4882a593SmuzhiyunEND(__memcpy)
109