| blob | 005bc93329c3cfab5bee85135b21c9e72b9e2d0f |
1 /*
2 * linux/arch/arm/lib/csumpartialcopy.S
3 *
4 * Copyright (C) 1995-1998 Russell King
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
9 */
10 #include <linux/linkage.h>
11 #include <asm/assembler.h>
13 .text
15 /* Function: __u32 csum_partial_copy_nocheck(const char *src, char *dst, int len, __u32 sum)
16 * Params : r0 = src, r1 = dst, r2 = len, r3 = checksum
17 * Returns : r0 = new checksum
18 */
20 .macro save_regs
21 stmfd sp!, {r4 - r8, fp, ip, lr, pc}
22 .endm
24 .macro load_regs,flags
25 LOADREGS(\flags,fp,{r4 - r8, fp, sp, pc})
26 .endm
28 .macro load1b, reg1
29 ldrb \reg1, [r0], #1
30 .endm
32 .macro load2b, reg1, reg2
33 ldrb \reg1, [r0], #1
34 ldrb \reg2, [r0], #1
35 .endm
37 .macro load1l, reg1
38 ldr \reg1, [r0], #4
39 .endm
41 .macro load2l, reg1, reg2
42 ldr \reg1, [r0], #4
43 ldr \reg2, [r0], #4
44 .endm
46 .macro load4l, reg1, reg2, reg3, reg4
47 ldmia r0!, {\reg1, \reg2, \reg3, \reg4}
48 .endm
50 ENTRY(csum_partial_copy_nocheck)
51 mov ip, sp
52 save_regs
53 sub fp, ip, #4
54 cmp r2, #4
55 blt .too_small
56 tst r1, #2 @ Test destination alignment
57 beq .dst_aligned
58 load2b ip, r8
59 subs r2, r2, #2 @ We do not know if SRC is aligned...
60 orr ip, ip, r8, lsl #8
61 adds r3, r3, ip
62 adcs r3, r3, #0
63 strb ip, [r1], #1
64 mov ip, ip, lsr #8
65 strb ip, [r1], #1 @ Destination now aligned
66 .dst_aligned: tst r0, #3
67 bne .src_not_aligned
68 adds r3, r3, #0
69 bics ip, r2, #15 @ Routine for src & dst aligned
70 beq 2f
71 1: load4l r4, r5, r6, r7
72 stmia r1!, {r4, r5, r6, r7}
73 adcs r3, r3, r4
74 adcs r3, r3, r5
75 adcs r3, r3, r6
76 adcs r3, r3, r7
77 sub ip, ip, #16
78 teq ip, #0
79 bne 1b
80 2: ands ip, r2, #12
81 beq 4f
82 tst ip, #8
83 beq 3f
84 load2l r4, r5
85 stmia r1!, {r4, r5}
86 adcs r3, r3, r4
87 adcs r3, r3, r5
88 tst ip, #4
89 beq 4f
90 3: load1l r4
91 str r4, [r1], #4
92 adcs r3, r3, r4
93 4: ands r2, r2, #3
94 adceq r0, r3, #0
95 load_regs eqea
96 load1l r4
97 tst r2, #2
98 beq .exit
99 adcs r3, r3, r4, lsl #16
100 strb r4, [r1], #1
101 mov r4, r4, lsr #8
102 strb r4, [r1], #1
103 mov r4, r4, lsr #8
104 .exit: tst r2, #1
105 strneb r4, [r1], #1
106 andne r4, r4, #255
107 adcnes r3, r3, r4
108 adcs r0, r3, #0
109 load_regs ea
111 .too_small: teq r2, #0
112 load_regs eqea
113 cmp r2, #2
114 blt .too_small1
115 load2b ip, r8
116 orr ip, ip, r8, lsl #8
117 adds r3, r3, ip
118 strb ip, [r1], #1
119 strb r8, [r1], #1
120 tst r2, #1
121 .too_small1: @ C = 0
122 beq .csum_exit
123 load1b ip
124 strb ip, [r1], #1
125 adcs r3, r3, ip
126 .csum_exit: adc r0, r3, #0
127 load_regs ea
129 .src_not_aligned:
130 cmp r2, #4
131 blt .too_small
132 and ip, r0, #3
133 bic r0, r0, #3
134 load1l r4
135 cmp ip, #2
136 beq .src2_aligned
137 bhi .src3_aligned
138 mov r4, r4, lsr #8
139 adds r3, r3, #0
140 bics ip, r2, #15
141 beq 2f
142 1: load4l r5, r6, r7, r8
143 orr r4, r4, r5, lsl #24
144 mov r5, r5, lsr #8
145 orr r5, r5, r6, lsl #24
146 mov r6, r6, lsr #8
147 orr r6, r6, r7, lsl #24
148 mov r7, r7, lsr #8
149 orr r7, r7, r8, lsl #24
150 stmia r1!, {r4, r5, r6, r7}
151 adcs r3, r3, r4
152 adcs r3, r3, r5
153 adcs r3, r3, r6
154 adcs r3, r3, r7
155 mov r4, r8, lsr #8
156 sub ip, ip, #16
157 teq ip, #0
158 bne 1b
159 2: ands ip, r2, #12
160 beq 4f
161 tst ip, #8
162 beq 3f
163 load2l r5, r6
164 orr r4, r4, r5, lsl #24
165 mov r5, r5, lsr #8
166 orr r5, r5, r6, lsl #24
167 stmia r1!, {r4, r5}
168 adcs r3, r3, r4
169 adcs r3, r3, r5
170 mov r4, r6, lsr #8
171 tst ip, #4
172 beq 4f
173 3: load1l r5
174 orr r4, r4, r5, lsl #24
175 str r4, [r1], #4
176 adcs r3, r3, r4
177 mov r4, r5, lsr #8
178 4: ands r2, r2, #3
179 adceq r0, r3, #0
180 load_regs eqea
181 tst r2, #2
182 beq .exit
183 adcs r3, r3, r4, lsl #16
184 strb r4, [r1], #1
185 mov r4, r4, lsr #8
186 strb r4, [r1], #1
187 mov r4, r4, lsr #8
188 b .exit
190 .src2_aligned: mov r4, r4, lsr #16
191 adds r3, r3, #0
192 bics ip, r2, #15
193 beq 2f
194 1: load4l r5, r6, r7, r8
195 orr r4, r4, r5, lsl #16
196 mov r5, r5, lsr #16
197 orr r5, r5, r6, lsl #16
198 mov r6, r6, lsr #16
199 orr r6, r6, r7, lsl #16
200 mov r7, r7, lsr #16
201 orr r7, r7, r8, lsl #16
202 stmia r1!, {r4, r5, r6, r7}
203 adcs r3, r3, r4
204 adcs r3, r3, r5
205 adcs r3, r3, r6
206 adcs r3, r3, r7
207 mov r4, r8, lsr #16
208 sub ip, ip, #16
209 teq ip, #0
210 bne 1b
211 2: ands ip, r2, #12
212 beq 4f
213 tst ip, #8
214 beq 3f
215 load2l r5, r6
216 orr r4, r4, r5, lsl #16
217 mov r5, r5, lsr #16
218 orr r5, r5, r6, lsl #16
219 stmia r1!, {r4, r5}
220 adcs r3, r3, r4
221 adcs r3, r3, r5
222 mov r4, r6, lsr #16
223 tst ip, #4
224 beq 4f
225 3: load1l r5
226 orr r4, r4, r5, lsl #16
227 str r4, [r1], #4
228 adcs r3, r3, r4
229 mov r4, r5, lsr #16
230 4: ands r2, r2, #3
231 adceq r0, r3, #0
232 load_regs eqea
233 tst r2, #2
234 beq .exit
235 adcs r3, r3, r4, lsl #16
236 strb r4, [r1], #1
237 mov r4, r4, lsr #8
238 strb r4, [r1], #1
239 tst r2, #1
240 adceq r0, r3, #0
241 load_regs eqea
242 load1b r4
243 b .exit
245 .src3_aligned: mov r4, r4, lsr #24
246 adds r3, r3, #0
247 bics ip, r2, #15
248 beq 2f
249 1: load4l r5, r6, r7, r8
250 orr r4, r4, r5, lsl #8
251 mov r5, r5, lsr #24
252 orr r5, r5, r6, lsl #8
253 mov r6, r6, lsr #24
254 orr r6, r6, r7, lsl #8
255 mov r7, r7, lsr #24
256 orr r7, r7, r8, lsl #8
257 stmia r1!, {r4, r5, r6, r7}
258 adcs r3, r3, r4
259 adcs r3, r3, r5
260 adcs r3, r3, r6
261 adcs r3, r3, r7
262 mov r4, r8, lsr #24
263 sub ip, ip, #16
264 teq ip, #0
265 bne 1b
266 2: ands ip, r2, #12
267 beq 4f
268 tst ip, #8
269 beq 3f
270 load2l r5, r6
271 orr r4, r4, r5, lsl #8
272 mov r5, r5, lsr #24
273 orr r5, r5, r6, lsl #8
274 stmia r1!, {r4, r5}
275 adcs r3, r3, r4
276 adcs r3, r3, r5
277 mov r4, r6, lsr #24
278 tst ip, #4
279 beq 4f
280 3: load1l r5
281 orr r4, r4, r5, lsl #8
282 str r4, [r1], #4
283 adcs r3, r3, r4
284 mov r4, r5, lsr #24
285 4: ands r2, r2, #3
286 adceq r0, r3, #0
287 load_regs eqea
288 tst r2, #2
289 beq .exit
290 adcs r3, r3, r4, lsl #16
291 strb r4, [r1], #1
292 load1l r4
293 strb r4, [r1], #1
294 adcs r3, r3, r4, lsl #24
295 mov r4, r4, lsr #8
296 b .exit