arch/sh/mm/copy_page.S

   1 /*
   2  * copy_page, __copy_user_page, __copy_user implementation of SuperH
   3  *
   4  * Copyright (C) 2001  Niibe Yutaka & Kaz Kojima
   5  * Copyright (C) 2002  Toshinobu Sugioka
   6  * Copyright (C) 2006  Paul Mundt
   7  */
   8 #include <linux/linkage.h>
   9 #include <asm/page.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 + PAGE_SIZE
  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.l   .Lpsz,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 + PAGE_SIZE
  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.l   .Lpsz,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         .align 2
 133 .Lpsz:  .long   PAGE_SIZE
 134 /*
 135  * __kernel_size_t __copy_user(void *to, const void *from, __kernel_size_t n);
 136  * Return the number of bytes NOT copied
 137  */
 138 #define EX(...)                 \
 139         9999: __VA_ARGS__ ;             \
 140         .section __ex_table, "a";       \
 141         .long 9999b, 6000f      ;       \
 142         .previous
 143 ENTRY(__copy_user)
 144         tst     r6,r6           ! Check explicitly for zero
 145         bf      1f
 146         rts
 147          mov    #0,r0           ! normal return
 148 1:
 149         mov.l   r10,@-r15
 150         mov.l   r9,@-r15
 151         mov.l   r8,@-r15
 152         mov     r4,r3
 153         add     r6,r3           ! last destination address
 154         mov     #12,r0          ! Check if small number of bytes
 155         cmp/gt  r0,r6
 156         bt      2f
 157         bra     .L_cleanup_loop
 158          nop
 159 2:
 160         neg     r5,r0           ! Calculate bytes needed to align source
 161         add     #4,r0
 162         and     #3,r0
 163         tst     r0,r0
 164         bt      .L_jump
 165         mov     r0,r1
 166
 167 .L_loop1:
 168         ! Copy bytes to align source
 169 EX(     mov.b   @r5+,r0         )
 170         dt      r1
 171 EX(     mov.b   r0,@r4          )
 172         add     #-1,r6
 173         bf/s    .L_loop1
 174          add    #1,r4
 175
 176 .L_jump:
 177         mov     r6,r2           ! Calculate number of longwords to copy
 178         shlr2   r2
 179         tst     r2,r2
 180         bt      .L_cleanup
 181
 182         mov     r4,r0           ! Jump to appropriate routine
 183         and     #3,r0
 184         mov     r0,r1
 185         shll2   r1
 186         mova    .L_jump_tbl,r0
 187         mov.l   @(r0,r1),r1
 188         jmp     @r1
 189          nop
 190
 191         .align 2
 192 .L_jump_tbl:
 193         .long   .L_dest00
 194         .long   .L_dest01
 195         .long   .L_dest10
 196         .long   .L_dest11
 197
 198 ! Destination = 00
 199
 200 .L_dest00:
 201         mov     r2,r7
 202         shlr2   r7
 203         shlr    r7
 204         tst     r7,r7
 205         mov     #7,r0
 206         bt/s    1f
 207          and    r0,r2
 208         .align 2
 209 2:
 210 EX(     mov.l   @r5+,r0         )
 211 EX(     mov.l   @r5+,r8         )
 212 EX(     mov.l   @r5+,r9         )
 213 EX(     mov.l   @r5+,r10        )
 214 EX(     mov.l   r0,@r4          )
 215 EX(     mov.l   r8,@(4,r4)      )
 216 EX(     mov.l   r9,@(8,r4)      )
 217 EX(     mov.l   r10,@(12,r4)    )
 218 EX(     mov.l   @r5+,r0         )
 219 EX(     mov.l   @r5+,r8         )
 220 EX(     mov.l   @r5+,r9         )
 221 EX(     mov.l   @r5+,r10        )
 222         dt      r7
 223 EX(     mov.l   r0,@(16,r4)     )
 224 EX(     mov.l   r8,@(20,r4)     )
 225 EX(     mov.l   r9,@(24,r4)     )
 226 EX(     mov.l   r10,@(28,r4)    )
 227         bf/s    2b
 228          add    #32,r4
 229         tst     r2,r2
 230         bt      .L_cleanup
 231 1:
 232 EX(     mov.l   @r5+,r0         )
 233         dt      r2
 234 EX(     mov.l   r0,@r4          )
 235         bf/s    1b
 236          add    #4,r4
 237
 238         bra     .L_cleanup
 239          nop
 240
 241 ! Destination = 10
 242
 243 .L_dest10:
 244         mov     r2,r7
 245         shlr2   r7
 246         shlr    r7
 247         tst     r7,r7
 248         mov     #7,r0
 249         bt/s    1f
 250          and    r0,r2
 251 2:
 252         dt      r7
 253 #ifdef __LITTLE_ENDIAN__
 254 EX(     mov.l   @r5+,r0         )
 255 EX(     mov.l   @r5+,r1         )
 256 EX(     mov.l   @r5+,r8         )
 257 EX(     mov.l   @r5+,r9         )
 258 EX(     mov.l   @r5+,r10        )
 259 EX(     mov.w   r0,@r4          )
 260         add     #2,r4
 261         xtrct   r1,r0
 262         xtrct   r8,r1
 263         xtrct   r9,r8
 264         xtrct   r10,r9
 265
 266 EX(     mov.l   r0,@r4          )
 267 EX(     mov.l   r1,@(4,r4)      )
 268 EX(     mov.l   r8,@(8,r4)      )
 269 EX(     mov.l   r9,@(12,r4)     )
 270
 271 EX(     mov.l   @r5+,r1         )
 272 EX(     mov.l   @r5+,r8         )
 273 EX(     mov.l   @r5+,r0         )
 274         xtrct   r1,r10
 275         xtrct   r8,r1
 276         xtrct   r0,r8
 277         shlr16  r0
 278 EX(     mov.l   r10,@(16,r4)    )
 279 EX(     mov.l   r1,@(20,r4)     )
 280 EX(     mov.l   r8,@(24,r4)     )
 281 EX(     mov.w   r0,@(28,r4)     )
 282         bf/s    2b
 283          add    #30,r4
 284 #else
 285 EX(     mov.l   @(28,r5),r0     )
 286 EX(     mov.l   @(24,r5),r8     )
 287 EX(     mov.l   @(20,r5),r9     )
 288 EX(     mov.l   @(16,r5),r10    )
 289 EX(     mov.w   r0,@(30,r4)     )
 290         add     #-2,r4
 291         xtrct   r8,r0
 292         xtrct   r9,r8
 293         xtrct   r10,r9
 294 EX(     mov.l   r0,@(28,r4)     )
 295 EX(     mov.l   r8,@(24,r4)     )
 296 EX(     mov.l   r9,@(20,r4)     )
 297
 298 EX(     mov.l   @(12,r5),r0     )
 299 EX(     mov.l   @(8,r5),r8      )
 300         xtrct   r0,r10
 301 EX(     mov.l   @(4,r5),r9      )
 302         mov.l   r10,@(16,r4)
 303 EX(     mov.l   @r5,r10         )
 304         xtrct   r8,r0
 305         xtrct   r9,r8
 306         xtrct   r10,r9
 307 EX(     mov.l   r0,@(12,r4)     )
 308 EX(     mov.l   r8,@(8,r4)      )
 309         swap.w  r10,r0
 310 EX(     mov.l   r9,@(4,r4)      )
 311 EX(     mov.w   r0,@(2,r4)      )
 312
 313         add     #32,r5
 314         bf/s    2b
 315          add    #34,r4
 316 #endif
 317         tst     r2,r2
 318         bt      .L_cleanup
 319
 320 1:      ! Read longword, write two words per iteration
 321 EX(     mov.l   @r5+,r0         )
 322         dt      r2
 323 #ifdef __LITTLE_ENDIAN__
 324 EX(     mov.w   r0,@r4          )
 325         shlr16  r0
 326 EX(     mov.w   r0,@(2,r4)      )
 327 #else
 328 EX(     mov.w   r0,@(2,r4)      )
 329         shlr16  r0
 330 EX(     mov.w   r0,@r4          )
 331 #endif
 332         bf/s    1b
 333          add    #4,r4
 334
 335         bra     .L_cleanup
 336          nop
 337
 338 ! Destination = 01 or 11
 339
 340 .L_dest01:
 341 .L_dest11:
 342         ! Read longword, write byte, word, byte per iteration
 343 EX(     mov.l   @r5+,r0         )
 344         dt      r2
 345 #ifdef __LITTLE_ENDIAN__
 346 EX(     mov.b   r0,@r4          )
 347         shlr8   r0
 348         add     #1,r4
 349 EX(     mov.w   r0,@r4          )
 350         shlr16  r0
 351 EX(     mov.b   r0,@(2,r4)      )
 352         bf/s    .L_dest01
 353          add    #3,r4
 354 #else
 355 EX(     mov.b   r0,@(3,r4)      )
 356         shlr8   r0
 357         swap.w  r0,r7
 358 EX(     mov.b   r7,@r4          )
 359         add     #1,r4
 360 EX(     mov.w   r0,@r4          )
 361         bf/s    .L_dest01
 362          add    #3,r4
 363 #endif
 364
 365 ! Cleanup last few bytes
 366 .L_cleanup:
 367         mov     r6,r0
 368         and     #3,r0
 369         tst     r0,r0
 370         bt      .L_exit
 371         mov     r0,r6
 372
 373 .L_cleanup_loop:
 374 EX(     mov.b   @r5+,r0         )
 375         dt      r6
 376 EX(     mov.b   r0,@r4          )
 377         bf/s    .L_cleanup_loop
 378          add    #1,r4
 379
 380 .L_exit:
 381         mov     #0,r0           ! normal return
 382 5000:
 383
 384 # Exception handler:
 385 .section .fixup, "ax"
 386 6000:
 387         mov.l   8000f,r1
 388         mov     r3,r0
 389         jmp     @r1
 390          sub    r4,r0
 391         .align  2
 392 8000:   .long   5000b
 393
 394 .previous
 395         mov.l   @r15+,r8
 396         mov.l   @r15+,r9
 397         rts
 398          mov.l  @r15+,r10