1 | /* SPDX-License-Identifier: GPL-2.0 */ |
2 | /* $Id: memcpy.S,v 1.3 2001/07/27 11:50:52 gniibe Exp $ |
3 | * |
4 | * "memcpy" implementation of SuperH |
5 | * |
6 | * Copyright (C) 1999 Niibe Yutaka |
7 | * |
8 | */ |
9 | |
10 | /* |
11 | * void *memcpy(void *dst, const void *src, size_t n); |
12 | * No overlap between the memory of DST and of SRC are assumed. |
13 | */ |
14 | |
15 | #include <linux/linkage.h> |
16 | ENTRY(memcpy) |
17 | tst r6,r6 |
18 | bt/s 9f ! if n=0, do nothing |
19 | mov r4,r0 |
20 | sub r4,r5 ! From here, r5 has the distance to r0 |
21 | add r6,r0 ! From here, r0 points the end of copying point |
22 | mov #12,r1 |
23 | cmp/gt r6,r1 |
24 | bt/s 7f ! if it's too small, copy a byte at once |
25 | add #-1,r5 |
26 | add #1,r5 |
27 | ! From here, r6 is free |
28 | ! |
29 | ! r4 --> [ ... ] DST [ ... ] SRC |
30 | ! [ ... ] [ ... ] |
31 | ! : : |
32 | ! r0 --> [ ... ] r0+r5 --> [ ... ] |
33 | ! |
34 | ! |
35 | mov r5,r1 |
36 | mov #3,r2 |
37 | and r2,r1 |
38 | shll2 r1 |
39 | mov r0,r3 ! Save the value on R0 to R3 |
40 | mova jmptable,r0 |
41 | add r1,r0 |
42 | mov.l @r0,r1 |
43 | jmp @r1 |
44 | mov r3,r0 ! and back to R0 |
45 | .balign 4 |
46 | jmptable: |
47 | .long case0 |
48 | .long case1 |
49 | .long case2 |
50 | .long case3 |
51 | |
52 | ! copy a byte at once |
53 | 7: mov r4,r2 |
54 | add #1,r2 |
55 | 8: |
56 | cmp/hi r2,r0 |
57 | mov.b @(r0,r5),r1 |
58 | bt/s 8b ! while (r0>r2) |
59 | mov.b r1,@-r0 |
60 | 9: |
61 | rts |
62 | nop |
63 | |
64 | case0: |
65 | ! |
66 | ! GHIJ KLMN OPQR --> GHIJ KLMN OPQR |
67 | ! |
68 | ! First, align to long word boundary |
69 | mov r0,r3 |
70 | and r2,r3 |
71 | tst r3,r3 |
72 | bt/s 2f |
73 | add #-4,r5 |
74 | add #3,r5 |
75 | 1: dt r3 |
76 | mov.b @(r0,r5),r1 |
77 | bf/s 1b |
78 | mov.b r1,@-r0 |
79 | ! |
80 | add #-3,r5 |
81 | 2: ! Second, copy a long word at once |
82 | mov r4,r2 |
83 | add #7,r2 |
84 | 3: mov.l @(r0,r5),r1 |
85 | cmp/hi r2,r0 |
86 | bt/s 3b |
87 | mov.l r1,@-r0 |
88 | ! |
89 | ! Third, copy a byte at once, if necessary |
90 | cmp/eq r4,r0 |
91 | bt/s 9b |
92 | add #3,r5 |
93 | bra 8b |
94 | add #-6,r2 |
95 | |
96 | case1: |
97 | ! |
98 | ! GHIJ KLMN OPQR --> ...G HIJK LMNO PQR. |
99 | ! |
100 | ! First, align to long word boundary |
101 | mov r0,r3 |
102 | and r2,r3 |
103 | tst r3,r3 |
104 | bt/s 2f |
105 | add #-1,r5 |
106 | 1: dt r3 |
107 | mov.b @(r0,r5),r1 |
108 | bf/s 1b |
109 | mov.b r1,@-r0 |
110 | ! |
111 | 2: ! Second, read a long word and write a long word at once |
112 | mov.l @(r0,r5),r1 |
113 | add #-4,r5 |
114 | mov r4,r2 |
115 | add #7,r2 |
116 | ! |
117 | #ifdef __LITTLE_ENDIAN__ |
118 | 3: mov r1,r3 ! RQPO |
119 | shll16 r3 |
120 | shll8 r3 ! Oxxx |
121 | mov.l @(r0,r5),r1 ! NMLK |
122 | mov r1,r6 |
123 | shlr8 r6 ! xNML |
124 | or r6,r3 ! ONML |
125 | cmp/hi r2,r0 |
126 | bt/s 3b |
127 | mov.l r3,@-r0 |
128 | #else |
129 | 3: mov r1,r3 ! OPQR |
130 | shlr16 r3 |
131 | shlr8 r3 ! xxxO |
132 | mov.l @(r0,r5),r1 ! KLMN |
133 | mov r1,r6 |
134 | shll8 r6 ! LMNx |
135 | or r6,r3 ! LMNO |
136 | cmp/hi r2,r0 |
137 | bt/s 3b |
138 | mov.l r3,@-r0 |
139 | #endif |
140 | ! |
141 | ! Third, copy a byte at once, if necessary |
142 | cmp/eq r4,r0 |
143 | bt/s 9b |
144 | add #4,r5 |
145 | bra 8b |
146 | add #-6,r2 |
147 | |
148 | case2: |
149 | ! |
150 | ! GHIJ KLMN OPQR --> ..GH IJKL MNOP QR.. |
151 | ! |
152 | ! First, align to word boundary |
153 | tst #1,r0 |
154 | bt/s 2f |
155 | add #-1,r5 |
156 | mov.b @(r0,r5),r1 |
157 | mov.b r1,@-r0 |
158 | ! |
159 | 2: ! Second, read a word and write a word at once |
160 | add #-1,r5 |
161 | mov r4,r2 |
162 | add #3,r2 |
163 | ! |
164 | 3: mov.w @(r0,r5),r1 |
165 | cmp/hi r2,r0 |
166 | bt/s 3b |
167 | mov.w r1,@-r0 |
168 | ! |
169 | ! Third, copy a byte at once, if necessary |
170 | cmp/eq r4,r0 |
171 | bt/s 9b |
172 | add #1,r5 |
173 | mov.b @(r0,r5),r1 |
174 | rts |
175 | mov.b r1,@-r0 |
176 | |
177 | case3: |
178 | ! |
179 | ! GHIJ KLMN OPQR --> .GHI JKLM NOPQ R... |
180 | ! |
181 | ! First, align to long word boundary |
182 | mov r0,r3 |
183 | and r2,r3 |
184 | tst r3,r3 |
185 | bt/s 2f |
186 | add #-1,r5 |
187 | 1: dt r3 |
188 | mov.b @(r0,r5),r1 |
189 | bf/s 1b |
190 | mov.b r1,@-r0 |
191 | ! |
192 | 2: ! Second, read a long word and write a long word at once |
193 | add #-2,r5 |
194 | mov.l @(r0,r5),r1 |
195 | add #-4,r5 |
196 | mov r4,r2 |
197 | add #7,r2 |
198 | ! |
199 | #ifdef __LITTLE_ENDIAN__ |
200 | 3: mov r1,r3 ! RQPO |
201 | shll8 r3 ! QPOx |
202 | mov.l @(r0,r5),r1 ! NMLK |
203 | mov r1,r6 |
204 | shlr16 r6 |
205 | shlr8 r6 ! xxxN |
206 | or r6,r3 ! QPON |
207 | cmp/hi r2,r0 |
208 | bt/s 3b |
209 | mov.l r3,@-r0 |
210 | #else |
211 | 3: mov r1,r3 ! OPQR |
212 | shlr8 r3 ! xOPQ |
213 | mov.l @(r0,r5),r1 ! KLMN |
214 | mov r1,r6 |
215 | shll16 r6 |
216 | shll8 r6 ! Nxxx |
217 | or r6,r3 ! NOPQ |
218 | cmp/hi r2,r0 |
219 | bt/s 3b |
220 | mov.l r3,@-r0 |
221 | #endif |
222 | ! |
223 | ! Third, copy a byte at once, if necessary |
224 | cmp/eq r4,r0 |
225 | bt/s 9b |
226 | add #6,r5 |
227 | bra 8b |
228 | add #-6,r2 |
229 | |