arch/arm26/lib/memcpy.S

   1 /*
   2  *  linux/arch/arm26/lib/memcpy.S
   3  *
   4  *  Copyright (C) 1995-1999 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  *  ASM optimised string functions
  11  */
  12 #include <linux/linkage.h>
  13 #include <asm/assembler.h>
  14
  15                 .text
  16
  17 #define ENTER   \
  18                 mov     ip,sp   ;\
  19                 stmfd   sp!,{r4-r9,fp,ip,lr,pc} ;\
  20                 sub     fp,ip,#4
  21
  22 #define EXIT    \
  23                 LOADREGS(ea, fp, {r4 - r9, fp, sp, pc})
  24
  25 #define EXITEQ  \
  26                 LOADREGS(eqea, fp, {r4 - r9, fp, sp, pc})
  27
  28 /*
  29  * Prototype: void memcpy(void *to,const void *from,unsigned long n);
  30  * ARM3: cant use memcopy here!!!
  31  */
  32 ENTRY(memcpy)
  33 ENTRY(memmove)
  34                 ENTER
  35                 cmp     r1, r0
  36                 bcc     19f
  37                 subs    r2, r2, #4
  38                 blt     6f
  39                 ands    ip, r0, #3
  40                 bne     7f
  41                 ands    ip, r1, #3
  42                 bne     8f
  43
  44 1:              subs    r2, r2, #8
  45                 blt     5f
  46                 subs    r2, r2, #0x14
  47                 blt     3f
  48 2:              ldmia   r1!,{r3 - r9, ip}
  49                 stmia   r0!,{r3 - r9, ip}
  50                 subs    r2, r2, #32
  51                 bge     2b
  52                 cmn     r2, #16
  53                 ldmgeia r1!, {r3 - r6}
  54                 stmgeia r0!, {r3 - r6}
  55                 subge   r2, r2, #0x10
  56 3:              adds    r2, r2, #0x14
  57 4:              ldmgeia r1!, {r3 - r5}
  58                 stmgeia r0!, {r3 - r5}
  59                 subges  r2, r2, #12
  60                 bge     4b
  61 5:              adds    r2, r2, #8
  62                 blt     6f
  63                 subs    r2, r2, #4
  64                 ldrlt   r3, [r1], #4
  65                 ldmgeia r1!, {r4, r5}
  66                 strlt   r3, [r0], #4
  67                 stmgeia r0!, {r4, r5}
  68                 subge   r2, r2, #4
  69
  70 6:              adds    r2, r2, #4
  71                 EXITEQ
  72                 cmp     r2, #2
  73                 ldrb    r3, [r1], #1
  74                 ldrgeb  r4, [r1], #1
  75                 ldrgtb  r5, [r1], #1
  76                 strb    r3, [r0], #1
  77                 strgeb  r4, [r0], #1
  78                 strgtb  r5, [r0], #1
  79                 EXIT
  80
  81 7:              rsb     ip, ip, #4
  82                 cmp     ip, #2
  83                 ldrb    r3, [r1], #1
  84                 ldrgeb  r4, [r1], #1
  85                 ldrgtb  r5, [r1], #1
  86                 strb    r3, [r0], #1
  87                 strgeb  r4, [r0], #1
  88                 strgtb  r5, [r0], #1
  89                 subs    r2, r2, ip
  90                 blt     6b
  91                 ands    ip, r1, #3
  92                 beq     1b
  93
  94 8:              bic     r1, r1, #3
  95                 ldr     r7, [r1], #4
  96                 cmp     ip, #2
  97                 bgt     15f
  98                 beq     11f
  99                 cmp     r2, #12
 100                 blt     10f
 101                 sub     r2, r2, #12
 102 9:              mov     r3, r7, pull #8
 103                 ldmia   r1!, {r4 - r7}
 104                 orr     r3, r3, r4, push #24
 105                 mov     r4, r4, pull #8
 106                 orr     r4, r4, r5, push #24
 107                 mov     r5, r5, pull #8
 108                 orr     r5, r5, r6, push #24
 109                 mov     r6, r6, pull #8
 110                 orr     r6, r6, r7, push #24
 111                 stmia   r0!, {r3 - r6}
 112                 subs    r2, r2, #16
 113                 bge     9b
 114                 adds    r2, r2, #12
 115                 blt     100f
 116 10:             mov     r3, r7, pull #8
 117                 ldr     r7, [r1], #4
 118                 subs    r2, r2, #4
 119                 orr     r3, r3, r7, push #24
 120                 str     r3, [r0], #4
 121                 bge     10b
 122 100:            sub     r1, r1, #3
 123                 b       6b
 124
 125 11:             cmp     r2, #12
 126                 blt     13f             /* */
 127                 sub     r2, r2, #12
 128 12:             mov     r3, r7, pull #16
 129                 ldmia   r1!, {r4 - r7}
 130                 orr     r3, r3, r4, push #16
 131                 mov     r4, r4, pull #16
 132                 orr     r4, r4, r5, push #16
 133                 mov     r5, r5, pull #16
 134                 orr     r5, r5, r6, push #16
 135                 mov     r6, r6, pull #16
 136                 orr     r6, r6, r7, push #16
 137                 stmia   r0!, {r3 - r6}
 138                 subs    r2, r2, #16
 139                 bge     12b
 140                 adds    r2, r2, #12
 141                 blt     14f
 142 13:             mov     r3, r7, pull #16
 143                 ldr     r7, [r1], #4
 144                 subs    r2, r2, #4
 145                 orr     r3, r3, r7, push #16
 146                 str     r3, [r0], #4
 147                 bge     13b
 148 14:             sub     r1, r1, #2
 149                 b       6b
 150
 151 15:             cmp     r2, #12
 152                 blt     17f
 153                 sub     r2, r2, #12
 154 16:             mov     r3, r7, pull #24
 155                 ldmia   r1!, {r4 - r7}
 156                 orr     r3, r3, r4, push #8
 157                 mov     r4, r4, pull #24
 158                 orr     r4, r4, r5, push #8
 159                 mov     r5, r5, pull #24
 160                 orr     r5, r5, r6, push #8
 161                 mov     r6, r6, pull #24
 162                 orr     r6, r6, r7, push #8
 163                 stmia   r0!, {r3 - r6}
 164                 subs    r2, r2, #16
 165                 bge     16b
 166                 adds    r2, r2, #12
 167                 blt     18f
 168 17:             mov     r3, r7, pull #24
 169                 ldr     r7, [r1], #4
 170                 subs    r2, r2, #4
 171                 orr     r3, r3, r7, push #8
 172                 str     r3, [r0], #4
 173                 bge     17b
 174 18:             sub     r1, r1, #1
 175                 b       6b
 176
 177
 178 19:             add     r1, r1, r2
 179                 add     r0, r0, r2
 180                 subs    r2, r2, #4
 181                 blt     24f
 182                 ands    ip, r0, #3
 183                 bne     25f
 184                 ands    ip, r1, #3
 185                 bne     26f
 186
 187 20:             subs    r2, r2, #8
 188                 blt     23f
 189                 subs    r2, r2, #0x14
 190                 blt     22f
 191 21:             ldmdb   r1!, {r3 - r9, ip}
 192                 stmdb   r0!, {r3 - r9, ip}
 193                 subs    r2, r2, #32
 194                 bge     21b
 195 22:             cmn     r2, #16
 196                 ldmgedb r1!, {r3 - r6}
 197                 stmgedb r0!, {r3 - r6}
 198                 subge   r2, r2, #16
 199                 adds    r2, r2, #20
 200                 ldmgedb r1!, {r3 - r5}
 201                 stmgedb r0!, {r3 - r5}
 202                 subge   r2, r2, #12
 203 23:             adds    r2, r2, #8
 204                 blt     24f
 205                 subs    r2, r2, #4
 206                 ldrlt   r3, [r1, #-4]!
 207                 ldmgedb r1!, {r4, r5}
 208                 strlt   r3, [r0, #-4]!
 209                 stmgedb r0!, {r4, r5}
 210                 subge   r2, r2, #4
 211
 212 24:             adds    r2, r2, #4
 213                 EXITEQ
 214                 cmp     r2, #2
 215                 ldrb    r3, [r1, #-1]!
 216                 ldrgeb  r4, [r1, #-1]!
 217                 ldrgtb  r5, [r1, #-1]!
 218                 strb    r3, [r0, #-1]!
 219                 strgeb  r4, [r0, #-1]!
 220                 strgtb  r5, [r0, #-1]!
 221                 EXIT
 222
 223 25:             cmp     ip, #2
 224                 ldrb    r3, [r1, #-1]!
 225                 ldrgeb  r4, [r1, #-1]!
 226                 ldrgtb  r5, [r1, #-1]!
 227                 strb    r3, [r0, #-1]!
 228                 strgeb  r4, [r0, #-1]!
 229                 strgtb  r5, [r0, #-1]!
 230                 subs    r2, r2, ip
 231                 blt     24b
 232                 ands    ip, r1, #3
 233                 beq     20b
 234
 235 26:             bic     r1, r1, #3
 236                 ldr     r3, [r1], #0
 237                 cmp     ip, #2
 238                 blt     34f
 239                 beq     30f
 240                 cmp     r2, #12
 241                 blt     28f
 242                 sub     r2, r2, #12
 243 27:             mov     r7, r3, push #8
 244                 ldmdb   r1!, {r3, r4, r5, r6}
 245                 orr     r7, r7, r6, pull #24
 246                 mov     r6, r6, push #8
 247                 orr     r6, r6, r5, pull #24
 248                 mov     r5, r5, push #8
 249                 orr     r5, r5, r4, pull #24
 250                 mov     r4, r4, push #8
 251                 orr     r4, r4, r3, pull #24
 252                 stmdb   r0!, {r4, r5, r6, r7}
 253                 subs    r2, r2, #16
 254                 bge     27b
 255                 adds    r2, r2, #12
 256                 blt     29f
 257 28:             mov     ip, r3, push #8
 258                 ldr     r3, [r1, #-4]!
 259                 subs    r2, r2, #4
 260                 orr     ip, ip, r3, pull #24
 261                 str     ip, [r0, #-4]!
 262                 bge     28b
 263 29:             add     r1, r1, #3
 264                 b       24b
 265
 266 30:             cmp     r2, #12
 267                 blt     32f
 268                 sub     r2, r2, #12
 269 31:             mov     r7, r3, push #16
 270                 ldmdb   r1!, {r3, r4, r5, r6}
 271                 orr     r7, r7, r6, pull #16
 272                 mov     r6, r6, push #16
 273                 orr     r6, r6, r5, pull #16
 274                 mov     r5, r5, push #16
 275                 orr     r5, r5, r4, pull #16
 276                 mov     r4, r4, push #16
 277                 orr     r4, r4, r3, pull #16
 278                 stmdb   r0!, {r4, r5, r6, r7}
 279                 subs    r2, r2, #16
 280                 bge     31b
 281                 adds    r2, r2, #12
 282                 blt     33f
 283 32:             mov     ip, r3, push #16
 284                 ldr     r3, [r1, #-4]!
 285                 subs    r2, r2, #4
 286                 orr     ip, ip, r3, pull #16
 287                 str     ip, [r0, #-4]!
 288                 bge     32b
 289 33:             add     r1, r1, #2
 290                 b       24b
 291
 292 34:             cmp     r2, #12
 293                 blt     36f
 294                 sub     r2, r2, #12
 295 35:             mov     r7, r3, push #24
 296                 ldmdb   r1!, {r3, r4, r5, r6}
 297                 orr     r7, r7, r6, pull #8
 298                 mov     r6, r6, push #24
 299                 orr     r6, r6, r5, pull #8
 300                 mov     r5, r5, push #24
 301                 orr     r5, r5, r4, pull #8
 302                 mov     r4, r4, push #24
 303                 orr     r4, r4, r3, pull #8
 304                 stmdb   r0!, {r4, r5, r6, r7}
 305                 subs    r2, r2, #16
 306                 bge     35b
 307                 adds    r2, r2, #12
 308                 blt     37f
 309 36:             mov     ip, r3, push #24
 310                 ldr     r3, [r1, #-4]!
 311                 subs    r2, r2, #4
 312                 orr     ip, ip, r3, pull #8
 313                 str     ip, [r0, #-4]!
 314                 bge     36b
 315 37:             add     r1, r1, #1
 316                 b       24b
 317
 318                 .align