arch/sh/mm/copy_page.S

   1 /* $Id: copy_page.S,v 1.8 2003/08/25 17:03:10 lethal Exp $
   2  *
   3  * copy_page, __copy_user_page, __copy_user implementation of SuperH
   4  *
   5  * Copyright (C) 2001  Niibe Yutaka & Kaz Kojima
   6  * Copyright (C) 2002  Toshinobu Sugioka
   7  *
   8  */
   9 #include <linux/linkage.h>
  10
  11 /*
  12  * copy_page_slow
  13  * @to: P1 address
  14  * @from: P1 address
  15  *
  16  * void copy_page_slow(void *to, void *from)
  17  */
  18
  19 /*
  20  * r0, r1, r2, r3, r4, r5, r6, r7 --- scratch
  21  * r8 --- from + 4096
  22  * r9 --- not used
  23  * r10 --- to
  24  * r11 --- from
  25  */
  26 ENTRY(copy_page_slow)
  27         mov.l   r8,@-r15
  28         mov.l   r10,@-r15
  29         mov.l   r11,@-r15
  30         mov     r4,r10
  31         mov     r5,r11
  32         mov     r5,r8
  33         mov.w   .L4096,r0
  34         add     r0,r8
  35         !
  36 1:      mov.l   @r11+,r0
  37         mov.l   @r11+,r1
  38         mov.l   @r11+,r2
  39         mov.l   @r11+,r3
  40         mov.l   @r11+,r4
  41         mov.l   @r11+,r5
  42         mov.l   @r11+,r6
  43         mov.l   @r11+,r7
  44 #if defined(CONFIG_CPU_SH3)
  45         mov.l   r0,@r10
  46 #elif defined(CONFIG_CPU_SH4)
  47         movca.l r0,@r10
  48         mov     r10,r0
  49 #endif
  50         add     #32,r10
  51         mov.l   r7,@-r10
  52         mov.l   r6,@-r10
  53         mov.l   r5,@-r10
  54         mov.l   r4,@-r10
  55         mov.l   r3,@-r10
  56         mov.l   r2,@-r10
  57         mov.l   r1,@-r10
  58 #if defined(CONFIG_CPU_SH4)
  59         ocbwb   @r0
  60 #endif
  61         cmp/eq  r11,r8
  62         bf/s    1b
  63          add    #28,r10
  64         !
  65         mov.l   @r15+,r11
  66         mov.l   @r15+,r10
  67         mov.l   @r15+,r8
  68         rts
  69          nop
  70
  71 #if defined(CONFIG_CPU_SH4)
  72 /*
  73  * __copy_user_page
  74  * @to: P1 address (with same color)
  75  * @from: P1 address
  76  * @orig_to: P1 address
  77  *
  78  * void __copy_user_page(void *to, void *from, void *orig_to)
  79  */
  80
  81 /*
  82  * r0, r1, r2, r3, r4, r5, r6, r7 --- scratch
  83  * r8 --- from + 4096
  84  * r9 --- orig_to
  85  * r10 --- to
  86  * r11 --- from
  87  */
  88 ENTRY(__copy_user_page)
  89         mov.l   r8,@-r15
  90         mov.l   r9,@-r15
  91         mov.l   r10,@-r15
  92         mov.l   r11,@-r15
  93         mov     r4,r10
  94         mov     r5,r11
  95         mov     r6,r9
  96         mov     r5,r8
  97         mov.w   .L4096,r0
  98         add     r0,r8
  99         !
 100 1:      ocbi    @r9
 101         add     #32,r9
 102         mov.l   @r11+,r0
 103         mov.l   @r11+,r1
 104         mov.l   @r11+,r2
 105         mov.l   @r11+,r3
 106         mov.l   @r11+,r4
 107         mov.l   @r11+,r5
 108         mov.l   @r11+,r6
 109         mov.l   @r11+,r7
 110         movca.l r0,@r10
 111         mov     r10,r0
 112         add     #32,r10
 113         mov.l   r7,@-r10
 114         mov.l   r6,@-r10
 115         mov.l   r5,@-r10
 116         mov.l   r4,@-r10
 117         mov.l   r3,@-r10
 118         mov.l   r2,@-r10
 119         mov.l   r1,@-r10
 120         ocbwb   @r0
 121         cmp/eq  r11,r8
 122         bf/s    1b
 123          add    #28,r10
 124         !
 125         mov.l   @r15+,r11
 126         mov.l   @r15+,r10
 127         mov.l   @r15+,r9
 128         mov.l   @r15+,r8
 129         rts
 130          nop
 131 #endif
 132 .L4096: .word   4096
 133 /*
 134  * __kernel_size_t __copy_user(void *to, const void *from, __kernel_size_t n);
 135  * Return the number of bytes NOT copied
 136  */
 137 #define EX(...)                 \
 138         9999: __VA_ARGS__ ;             \
 139         .section __ex_table, "a";       \
 140         .long 9999b, 6000f      ;       \
 141         .previous
 142 ENTRY(__copy_user)
 143         tst     r6,r6           ! Check explicitly for zero
 144         bf      1f
 145         rts
 146          mov    #0,r0           ! normal return
 147 1:
 148         mov.l   r10,@-r15
 149         mov.l   r9,@-r15
 150         mov.l   r8,@-r15
 151         mov     r4,r3
 152         add     r6,r3           ! last destination address
 153         mov     #12,r0          ! Check if small number of bytes
 154         cmp/gt  r0,r6
 155         bt      2f
 156         bra     .L_cleanup_loop
 157          nop
 158 2:
 159         neg     r5,r0           ! Calculate bytes needed to align source
 160         add     #4,r0
 161         and     #3,r0
 162         tst     r0,r0
 163         bt      .L_jump
 164         mov     r0,r1
 165
 166 .L_loop1:
 167         ! Copy bytes to align source
 168 EX(     mov.b   @r5+,r0         )
 169         dt      r1
 170 EX(     mov.b   r0,@r4          )
 171         add     #-1,r6
 172         bf/s    .L_loop1
 173          add    #1,r4
 174
 175 .L_jump:
 176         mov     r6,r2           ! Calculate number of longwords to copy
 177         shlr2   r2
 178         tst     r2,r2
 179         bt      .L_cleanup
 180
 181         mov     r4,r0           ! Jump to appropriate routine
 182         and     #3,r0
 183         mov     r0,r1
 184         shll2   r1
 185         mova    .L_jump_tbl,r0
 186         mov.l   @(r0,r1),r1
 187         jmp     @r1
 188          nop
 189
 190         .align 2
 191 .L_jump_tbl:
 192         .long   .L_dest00
 193         .long   .L_dest01
 194         .long   .L_dest10
 195         .long   .L_dest11
 196
 197 ! Destination = 00
 198
 199 .L_dest00:
 200         mov     r2,r7
 201         shlr2   r7
 202         shlr    r7
 203         tst     r7,r7
 204         mov     #7,r0
 205         bt/s    1f
 206          and    r0,r2
 207         .align 2
 208 2:
 209 EX(     mov.l   @r5+,r0         )
 210 EX(     mov.l   @r5+,r8         )
 211 EX(     mov.l   @r5+,r9         )
 212 EX(     mov.l   @r5+,r10        )
 213 EX(     mov.l   r0,@r4          )
 214 EX(     mov.l   r8,@(4,r4)      )
 215 EX(     mov.l   r9,@(8,r4)      )
 216 EX(     mov.l   r10,@(12,r4)    )
 217 EX(     mov.l   @r5+,r0         )
 218 EX(     mov.l   @r5+,r8         )
 219 EX(     mov.l   @r5+,r9         )
 220 EX(     mov.l   @r5+,r10        )
 221         dt      r7
 222 EX(     mov.l   r0,@(16,r4)     )
 223 EX(     mov.l   r8,@(20,r4)     )
 224 EX(     mov.l   r9,@(24,r4)     )
 225 EX(     mov.l   r10,@(28,r4)    )
 226         bf/s    2b
 227          add    #32,r4
 228         tst     r2,r2
 229         bt      .L_cleanup
 230 1:
 231 EX(     mov.l   @r5+,r0         )
 232         dt      r2
 233 EX(     mov.l   r0,@r4          )
 234         bf/s    1b
 235          add    #4,r4
 236
 237         bra     .L_cleanup
 238          nop
 239
 240 ! Destination = 10
 241
 242 .L_dest10:
 243         mov     r2,r7
 244         shlr2   r7
 245         shlr    r7
 246         tst     r7,r7
 247         mov     #7,r0
 248         bt/s    1f
 249          and    r0,r2
 250 2:
 251         dt      r7
 252 #ifdef __LITTLE_ENDIAN__
 253 EX(     mov.l   @r5+,r0         )
 254 EX(     mov.l   @r5+,r1         )
 255 EX(     mov.l   @r5+,r8         )
 256 EX(     mov.l   @r5+,r9         )
 257 EX(     mov.l   @r5+,r10        )
 258 EX(     mov.w   r0,@r4          )
 259         add     #2,r4
 260         xtrct   r1,r0
 261         xtrct   r8,r1
 262         xtrct   r9,r8
 263         xtrct   r10,r9
 264
 265 EX(     mov.l   r0,@r4          )
 266 EX(     mov.l   r1,@(4,r4)      )
 267 EX(     mov.l   r8,@(8,r4)      )
 268 EX(     mov.l   r9,@(12,r4)     )
 269
 270 EX(     mov.l   @r5+,r1         )
 271 EX(     mov.l   @r5+,r8         )
 272 EX(     mov.l   @r5+,r0         )
 273         xtrct   r1,r10
 274         xtrct   r8,r1
 275         xtrct   r0,r8
 276         shlr16  r0
 277 EX(     mov.l   r10,@(16,r4)    )
 278 EX(     mov.l   r1,@(20,r4)     )
 279 EX(     mov.l   r8,@(24,r4)     )
 280 EX(     mov.w   r0,@(28,r4)     )
 281         bf/s    2b
 282          add    #30,r4
 283 #else
 284 EX(     mov.l   @(28,r5),r0     )
 285 EX(     mov.l   @(24,r5),r8     )
 286 EX(     mov.l   @(20,r5),r9     )
 287 EX(     mov.l   @(16,r5),r10    )
 288 EX(     mov.w   r0,@(30,r4)     )
 289         add     #-2,r4
 290         xtrct   r8,r0
 291         xtrct   r9,r8
 292         xtrct   r10,r9
 293 EX(     mov.l   r0,@(28,r4)     )
 294 EX(     mov.l   r8,@(24,r4)     )
 295 EX(     mov.l   r9,@(20,r4)     )
 296
 297 EX(     mov.l   @(12,r5),r0     )
 298 EX(     mov.l   @(8,r5),r8      )
 299         xtrct   r0,r10
 300 EX(     mov.l   @(4,r5),r9      )
 301         mov.l   r10,@(16,r4)
 302 EX(     mov.l   @r5,r10         )
 303         xtrct   r8,r0
 304         xtrct   r9,r8
 305         xtrct   r10,r9
 306 EX(     mov.l   r0,@(12,r4)     )
 307 EX(     mov.l   r8,@(8,r4)      )
 308         swap.w  r10,r0
 309 EX(     mov.l   r9,@(4,r4)      )
 310 EX(     mov.w   r0,@(2,r4)      )
 311
 312         add     #32,r5
 313         bf/s    2b
 314          add    #34,r4
 315 #endif
 316         tst     r2,r2
 317         bt      .L_cleanup
 318
 319 1:      ! Read longword, write two words per iteration
 320 EX(     mov.l   @r5+,r0         )
 321         dt      r2
 322 #ifdef __LITTLE_ENDIAN__
 323 EX(     mov.w   r0,@r4          )
 324         shlr16  r0
 325 EX(     mov.w   r0,@(2,r4)      )
 326 #else
 327 EX(     mov.w   r0,@(2,r4)      )
 328         shlr16  r0
 329 EX(     mov.w   r0,@r4          )
 330 #endif
 331         bf/s    1b
 332          add    #4,r4
 333
 334         bra     .L_cleanup
 335          nop
 336
 337 ! Destination = 01 or 11
 338
 339 .L_dest01:
 340 .L_dest11:
 341         ! Read longword, write byte, word, byte per iteration
 342 EX(     mov.l   @r5+,r0         )
 343         dt      r2
 344 #ifdef __LITTLE_ENDIAN__
 345 EX(     mov.b   r0,@r4          )
 346         shlr8   r0
 347         add     #1,r4
 348 EX(     mov.w   r0,@r4          )
 349         shlr16  r0
 350 EX(     mov.b   r0,@(2,r4)      )
 351         bf/s    .L_dest01
 352          add    #3,r4
 353 #else
 354 EX(     mov.b   r0,@(3,r4)      )
 355         shlr8   r0
 356         swap.w  r0,r7
 357 EX(     mov.b   r7,@r4          )
 358         add     #1,r4
 359 EX(     mov.w   r0,@r4          )
 360         bf/s    .L_dest01
 361          add    #3,r4
 362 #endif
 363
 364 ! Cleanup last few bytes
 365 .L_cleanup:
 366         mov     r6,r0
 367         and     #3,r0
 368         tst     r0,r0
 369         bt      .L_exit
 370         mov     r0,r6
 371
 372 .L_cleanup_loop:
 373 EX(     mov.b   @r5+,r0         )
 374         dt      r6
 375 EX(     mov.b   r0,@r4          )
 376         bf/s    .L_cleanup_loop
 377          add    #1,r4
 378
 379 .L_exit:
 380         mov     #0,r0           ! normal return
 381 5000:
 382
 383 # Exception handler:
 384 .section .fixup, "ax"
 385 6000:
 386         mov.l   8000f,r1
 387         mov     r3,r0
 388         jmp     @r1
 389          sub    r4,r0
 390         .align  2
 391 8000:   .long   5000b
 392
 393 .previous
 394         mov.l   @r15+,r8
 395         mov.l   @r15+,r9
 396         rts
 397          mov.l  @r15+,r10