gcc/config/sh/lib1funcs.asm

   1 /* Copyright (C) 1994, 1995, 1997, 1998, 1999, 2000, 2001
   2    Free Software Foundation, Inc.
   3
   4 This file is free software; you can redistribute it and/or modify it
   5 under the terms of the GNU General Public License as published by the
   6 Free Software Foundation; either version 2, or (at your option) any
   7 later version.
   8
   9 In addition to the permissions in the GNU General Public License, the
  10 Free Software Foundation gives you unlimited permission to link the
  11 compiled version of this file into combinations with other programs,
  12 and to distribute those combinations without any restriction coming
  13 from the use of this file.  (The General Public License restrictions
  14 do apply in other respects; for example, they cover modification of
  15 the file, and distribution when not linked into a combine
  16 executable.)
  17
  18 This file is distributed in the hope that it will be useful, but
  19 WITHOUT ANY WARRANTY; without even the implied warranty of
  20 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  21 General Public License for more details.
  22
  23 You should have received a copy of the GNU General Public License
  24 along with this program; see the file COPYING.  If not, write to
  25 the Free Software Foundation, 59 Temple Place - Suite 330,
  26 Boston, MA 02111-1307, USA.  */
  27
  28 !! libgcc routines for the Hitachi SH cpu.
  29 !! Contributed by Steve Chamberlain.
  30 !! sac@cygnus.com
  31
  32 !! ashiftrt_r4_x, ___ashrsi3, ___ashlsi3, ___lshrsi3 routines
  33 !! recoded in assembly by Toshiyasu Morita
  34 !! tm@netcom.com
  35
  36 /* SH2 optimizations for ___ashrsi3, ___ashlsi3, ___lshrsi3 and
  37    ELF local label prefixes by J"orn Rennecke
  38    amylaar@cygnus.com  */
  39
  40 #ifdef __ELF__
  41 #define LOCAL(X) .L_##X
  42 #else
  43 #define LOCAL(X) L_##X
  44 #endif
  45
  46 #ifdef __linux__
  47 #define GLOBAL(X) __##X
  48 #endif
  49
  50 #ifndef GLOBAL
  51 #define GLOBAL(X) ___##X
  52 #endif
  53
  54 #ifdef L_ashiftrt
  55         .global GLOBAL(ashiftrt_r4_0)
  56         .global GLOBAL(ashiftrt_r4_1)
  57         .global GLOBAL(ashiftrt_r4_2)
  58         .global GLOBAL(ashiftrt_r4_3)
  59         .global GLOBAL(ashiftrt_r4_4)
  60         .global GLOBAL(ashiftrt_r4_5)
  61         .global GLOBAL(ashiftrt_r4_6)
  62         .global GLOBAL(ashiftrt_r4_7)
  63         .global GLOBAL(ashiftrt_r4_8)
  64         .global GLOBAL(ashiftrt_r4_9)
  65         .global GLOBAL(ashiftrt_r4_10)
  66         .global GLOBAL(ashiftrt_r4_11)
  67         .global GLOBAL(ashiftrt_r4_12)
  68         .global GLOBAL(ashiftrt_r4_13)
  69         .global GLOBAL(ashiftrt_r4_14)
  70         .global GLOBAL(ashiftrt_r4_15)
  71         .global GLOBAL(ashiftrt_r4_16)
  72         .global GLOBAL(ashiftrt_r4_17)
  73         .global GLOBAL(ashiftrt_r4_18)
  74         .global GLOBAL(ashiftrt_r4_19)
  75         .global GLOBAL(ashiftrt_r4_20)
  76         .global GLOBAL(ashiftrt_r4_21)
  77         .global GLOBAL(ashiftrt_r4_22)
  78         .global GLOBAL(ashiftrt_r4_23)
  79         .global GLOBAL(ashiftrt_r4_24)
  80         .global GLOBAL(ashiftrt_r4_25)
  81         .global GLOBAL(ashiftrt_r4_26)
  82         .global GLOBAL(ashiftrt_r4_27)
  83         .global GLOBAL(ashiftrt_r4_28)
  84         .global GLOBAL(ashiftrt_r4_29)
  85         .global GLOBAL(ashiftrt_r4_30)
  86         .global GLOBAL(ashiftrt_r4_31)
  87         .global GLOBAL(ashiftrt_r4_32)
  88
  89         .align  1
  90 GLOBAL(ashiftrt_r4_32):
  91 GLOBAL(ashiftrt_r4_31):
  92         rotcl   r4
  93         rts
  94         subc    r4,r4
  95
  96 GLOBAL(ashiftrt_r4_30):
  97         shar    r4
  98 GLOBAL(ashiftrt_r4_29):
  99         shar    r4
 100 GLOBAL(ashiftrt_r4_28):
 101         shar    r4
 102 GLOBAL(ashiftrt_r4_27):
 103         shar    r4
 104 GLOBAL(ashiftrt_r4_26):
 105         shar    r4
 106 GLOBAL(ashiftrt_r4_25):
 107         shar    r4
 108 GLOBAL(ashiftrt_r4_24):
 109         shlr16  r4
 110         shlr8   r4
 111         rts
 112         exts.b  r4,r4
 113
 114 GLOBAL(ashiftrt_r4_23):
 115         shar    r4
 116 GLOBAL(ashiftrt_r4_22):
 117         shar    r4
 118 GLOBAL(ashiftrt_r4_21):
 119         shar    r4
 120 GLOBAL(ashiftrt_r4_20):
 121         shar    r4
 122 GLOBAL(ashiftrt_r4_19):
 123         shar    r4
 124 GLOBAL(ashiftrt_r4_18):
 125         shar    r4
 126 GLOBAL(ashiftrt_r4_17):
 127         shar    r4
 128 GLOBAL(ashiftrt_r4_16):
 129         shlr16  r4
 130         rts
 131         exts.w  r4,r4
 132
 133 GLOBAL(ashiftrt_r4_15):
 134         shar    r4
 135 GLOBAL(ashiftrt_r4_14):
 136         shar    r4
 137 GLOBAL(ashiftrt_r4_13):
 138         shar    r4
 139 GLOBAL(ashiftrt_r4_12):
 140         shar    r4
 141 GLOBAL(ashiftrt_r4_11):
 142         shar    r4
 143 GLOBAL(ashiftrt_r4_10):
 144         shar    r4
 145 GLOBAL(ashiftrt_r4_9):
 146         shar    r4
 147 GLOBAL(ashiftrt_r4_8):
 148         shar    r4
 149 GLOBAL(ashiftrt_r4_7):
 150         shar    r4
 151 GLOBAL(ashiftrt_r4_6):
 152         shar    r4
 153 GLOBAL(ashiftrt_r4_5):
 154         shar    r4
 155 GLOBAL(ashiftrt_r4_4):
 156         shar    r4
 157 GLOBAL(ashiftrt_r4_3):
 158         shar    r4
 159 GLOBAL(ashiftrt_r4_2):
 160         shar    r4
 161 GLOBAL(ashiftrt_r4_1):
 162         rts
 163         shar    r4
 164
 165 GLOBAL(ashiftrt_r4_0):
 166         rts
 167         nop
 168 #endif
 169
 170 #ifdef L_ashiftrt_n
 171
 172 !
 173 ! GLOBAL(ashrsi3)
 174 !
 175 ! Entry:
 176 !
 177 ! r4: Value to shift
 178 ! r5: Shifts
 179 !
 180 ! Exit:
 181 !
 182 ! r0: Result
 183 !
 184 ! Destroys:
 185 !
 186 ! (none)
 187 !
 188
 189         .global GLOBAL(ashrsi3)
 190         .align  2
 191 GLOBAL(ashrsi3):
 192         mov     #31,r0
 193         and     r0,r5
 194         mova    LOCAL(ashrsi3_table),r0
 195         mov.b   @(r0,r5),r5
 196 #ifdef __sh1__
 197         add     r5,r0
 198         jmp     @r0
 199 #else
 200         braf    r5
 201 #endif
 202         mov     r4,r0
 203
 204         .align  2
 205 LOCAL(ashrsi3_table):
 206         .byte           LOCAL(ashrsi3_0)-LOCAL(ashrsi3_table)
 207         .byte           LOCAL(ashrsi3_1)-LOCAL(ashrsi3_table)
 208         .byte           LOCAL(ashrsi3_2)-LOCAL(ashrsi3_table)
 209         .byte           LOCAL(ashrsi3_3)-LOCAL(ashrsi3_table)
 210         .byte           LOCAL(ashrsi3_4)-LOCAL(ashrsi3_table)
 211         .byte           LOCAL(ashrsi3_5)-LOCAL(ashrsi3_table)
 212         .byte           LOCAL(ashrsi3_6)-LOCAL(ashrsi3_table)
 213         .byte           LOCAL(ashrsi3_7)-LOCAL(ashrsi3_table)
 214         .byte           LOCAL(ashrsi3_8)-LOCAL(ashrsi3_table)
 215         .byte           LOCAL(ashrsi3_9)-LOCAL(ashrsi3_table)
 216         .byte           LOCAL(ashrsi3_10)-LOCAL(ashrsi3_table)
 217         .byte           LOCAL(ashrsi3_11)-LOCAL(ashrsi3_table)
 218         .byte           LOCAL(ashrsi3_12)-LOCAL(ashrsi3_table)
 219         .byte           LOCAL(ashrsi3_13)-LOCAL(ashrsi3_table)
 220         .byte           LOCAL(ashrsi3_14)-LOCAL(ashrsi3_table)
 221         .byte           LOCAL(ashrsi3_15)-LOCAL(ashrsi3_table)
 222         .byte           LOCAL(ashrsi3_16)-LOCAL(ashrsi3_table)
 223         .byte           LOCAL(ashrsi3_17)-LOCAL(ashrsi3_table)
 224         .byte           LOCAL(ashrsi3_18)-LOCAL(ashrsi3_table)
 225         .byte           LOCAL(ashrsi3_19)-LOCAL(ashrsi3_table)
 226         .byte           LOCAL(ashrsi3_20)-LOCAL(ashrsi3_table)
 227         .byte           LOCAL(ashrsi3_21)-LOCAL(ashrsi3_table)
 228         .byte           LOCAL(ashrsi3_22)-LOCAL(ashrsi3_table)
 229         .byte           LOCAL(ashrsi3_23)-LOCAL(ashrsi3_table)
 230         .byte           LOCAL(ashrsi3_24)-LOCAL(ashrsi3_table)
 231         .byte           LOCAL(ashrsi3_25)-LOCAL(ashrsi3_table)
 232         .byte           LOCAL(ashrsi3_26)-LOCAL(ashrsi3_table)
 233         .byte           LOCAL(ashrsi3_27)-LOCAL(ashrsi3_table)
 234         .byte           LOCAL(ashrsi3_28)-LOCAL(ashrsi3_table)
 235         .byte           LOCAL(ashrsi3_29)-LOCAL(ashrsi3_table)
 236         .byte           LOCAL(ashrsi3_30)-LOCAL(ashrsi3_table)
 237         .byte           LOCAL(ashrsi3_31)-LOCAL(ashrsi3_table)
 238
 239 LOCAL(ashrsi3_31):
 240         rotcl   r0
 241         rts
 242         subc    r0,r0
 243
 244 LOCAL(ashrsi3_30):
 245         shar    r0
 246 LOCAL(ashrsi3_29):
 247         shar    r0
 248 LOCAL(ashrsi3_28):
 249         shar    r0
 250 LOCAL(ashrsi3_27):
 251         shar    r0
 252 LOCAL(ashrsi3_26):
 253         shar    r0
 254 LOCAL(ashrsi3_25):
 255         shar    r0
 256 LOCAL(ashrsi3_24):
 257         shlr16  r0
 258         shlr8   r0
 259         rts
 260         exts.b  r0,r0
 261
 262 LOCAL(ashrsi3_23):
 263         shar    r0
 264 LOCAL(ashrsi3_22):
 265         shar    r0
 266 LOCAL(ashrsi3_21):
 267         shar    r0
 268 LOCAL(ashrsi3_20):
 269         shar    r0
 270 LOCAL(ashrsi3_19):
 271         shar    r0
 272 LOCAL(ashrsi3_18):
 273         shar    r0
 274 LOCAL(ashrsi3_17):
 275         shar    r0
 276 LOCAL(ashrsi3_16):
 277         shlr16  r0
 278         rts
 279         exts.w  r0,r0
 280
 281 LOCAL(ashrsi3_15):
 282         shar    r0
 283 LOCAL(ashrsi3_14):
 284         shar    r0
 285 LOCAL(ashrsi3_13):
 286         shar    r0
 287 LOCAL(ashrsi3_12):
 288         shar    r0
 289 LOCAL(ashrsi3_11):
 290         shar    r0
 291 LOCAL(ashrsi3_10):
 292         shar    r0
 293 LOCAL(ashrsi3_9):
 294         shar    r0
 295 LOCAL(ashrsi3_8):
 296         shar    r0
 297 LOCAL(ashrsi3_7):
 298         shar    r0
 299 LOCAL(ashrsi3_6):
 300         shar    r0
 301 LOCAL(ashrsi3_5):
 302         shar    r0
 303 LOCAL(ashrsi3_4):
 304         shar    r0
 305 LOCAL(ashrsi3_3):
 306         shar    r0
 307 LOCAL(ashrsi3_2):
 308         shar    r0
 309 LOCAL(ashrsi3_1):
 310         rts
 311         shar    r0
 312
 313 LOCAL(ashrsi3_0):
 314         rts
 315         nop
 316
 317 #endif
 318
 319 #ifdef L_ashiftlt
 320
 321 !
 322 ! GLOBAL(ashlsi3)
 323 !
 324 ! Entry:
 325 !
 326 ! r4: Value to shift
 327 ! r5: Shifts
 328 !
 329 ! Exit:
 330 !
 331 ! r0: Result
 332 !
 333 ! Destroys:
 334 !
 335 ! (none)
 336 !
 337         .global GLOBAL(ashlsi3)
 338         .align  2
 339 GLOBAL(ashlsi3):
 340         mov     #31,r0
 341         and     r0,r5
 342         mova    LOCAL(ashlsi3_table),r0
 343         mov.b   @(r0,r5),r5
 344 #ifdef __sh1__
 345         add     r5,r0
 346         jmp     @r0
 347 #else
 348         braf    r5
 349 #endif
 350         mov     r4,r0
 351
 352         .align  2
 353 LOCAL(ashlsi3_table):
 354         .byte           LOCAL(ashlsi3_0)-LOCAL(ashlsi3_table)
 355         .byte           LOCAL(ashlsi3_1)-LOCAL(ashlsi3_table)
 356         .byte           LOCAL(ashlsi3_2)-LOCAL(ashlsi3_table)
 357         .byte           LOCAL(ashlsi3_3)-LOCAL(ashlsi3_table)
 358         .byte           LOCAL(ashlsi3_4)-LOCAL(ashlsi3_table)
 359         .byte           LOCAL(ashlsi3_5)-LOCAL(ashlsi3_table)
 360         .byte           LOCAL(ashlsi3_6)-LOCAL(ashlsi3_table)
 361         .byte           LOCAL(ashlsi3_7)-LOCAL(ashlsi3_table)
 362         .byte           LOCAL(ashlsi3_8)-LOCAL(ashlsi3_table)
 363         .byte           LOCAL(ashlsi3_9)-LOCAL(ashlsi3_table)
 364         .byte           LOCAL(ashlsi3_10)-LOCAL(ashlsi3_table)
 365         .byte           LOCAL(ashlsi3_11)-LOCAL(ashlsi3_table)
 366         .byte           LOCAL(ashlsi3_12)-LOCAL(ashlsi3_table)
 367         .byte           LOCAL(ashlsi3_13)-LOCAL(ashlsi3_table)
 368         .byte           LOCAL(ashlsi3_14)-LOCAL(ashlsi3_table)
 369         .byte           LOCAL(ashlsi3_15)-LOCAL(ashlsi3_table)
 370         .byte           LOCAL(ashlsi3_16)-LOCAL(ashlsi3_table)
 371         .byte           LOCAL(ashlsi3_17)-LOCAL(ashlsi3_table)
 372         .byte           LOCAL(ashlsi3_18)-LOCAL(ashlsi3_table)
 373         .byte           LOCAL(ashlsi3_19)-LOCAL(ashlsi3_table)
 374         .byte           LOCAL(ashlsi3_20)-LOCAL(ashlsi3_table)
 375         .byte           LOCAL(ashlsi3_21)-LOCAL(ashlsi3_table)
 376         .byte           LOCAL(ashlsi3_22)-LOCAL(ashlsi3_table)
 377         .byte           LOCAL(ashlsi3_23)-LOCAL(ashlsi3_table)
 378         .byte           LOCAL(ashlsi3_24)-LOCAL(ashlsi3_table)
 379         .byte           LOCAL(ashlsi3_25)-LOCAL(ashlsi3_table)
 380         .byte           LOCAL(ashlsi3_26)-LOCAL(ashlsi3_table)
 381         .byte           LOCAL(ashlsi3_27)-LOCAL(ashlsi3_table)
 382         .byte           LOCAL(ashlsi3_28)-LOCAL(ashlsi3_table)
 383         .byte           LOCAL(ashlsi3_29)-LOCAL(ashlsi3_table)
 384         .byte           LOCAL(ashlsi3_30)-LOCAL(ashlsi3_table)
 385         .byte           LOCAL(ashlsi3_31)-LOCAL(ashlsi3_table)
 386
 387 LOCAL(ashlsi3_6):
 388         shll2   r0
 389 LOCAL(ashlsi3_4):
 390         shll2   r0
 391 LOCAL(ashlsi3_2):
 392         rts
 393         shll2   r0
 394
 395 LOCAL(ashlsi3_7):
 396         shll2   r0
 397 LOCAL(ashlsi3_5):
 398         shll2   r0
 399 LOCAL(ashlsi3_3):
 400         shll2   r0
 401 LOCAL(ashlsi3_1):
 402         rts
 403         shll    r0
 404
 405 LOCAL(ashlsi3_14):
 406         shll2   r0
 407 LOCAL(ashlsi3_12):
 408         shll2   r0
 409 LOCAL(ashlsi3_10):
 410         shll2   r0
 411 LOCAL(ashlsi3_8):
 412         rts
 413         shll8   r0
 414
 415 LOCAL(ashlsi3_15):
 416         shll2   r0
 417 LOCAL(ashlsi3_13):
 418         shll2   r0
 419 LOCAL(ashlsi3_11):
 420         shll2   r0
 421 LOCAL(ashlsi3_9):
 422         shll8   r0
 423         rts
 424         shll    r0
 425
 426 LOCAL(ashlsi3_22):
 427         shll2   r0
 428 LOCAL(ashlsi3_20):
 429         shll2   r0
 430 LOCAL(ashlsi3_18):
 431         shll2   r0
 432 LOCAL(ashlsi3_16):
 433         rts
 434         shll16  r0
 435
 436 LOCAL(ashlsi3_23):
 437         shll2   r0
 438 LOCAL(ashlsi3_21):
 439         shll2   r0
 440 LOCAL(ashlsi3_19):
 441         shll2   r0
 442 LOCAL(ashlsi3_17):
 443         shll16  r0
 444         rts
 445         shll    r0
 446
 447 LOCAL(ashlsi3_30):
 448         shll2   r0
 449 LOCAL(ashlsi3_28):
 450         shll2   r0
 451 LOCAL(ashlsi3_26):
 452         shll2   r0
 453 LOCAL(ashlsi3_24):
 454         shll16  r0
 455         rts
 456         shll8   r0
 457
 458 LOCAL(ashlsi3_31):
 459         shll2   r0
 460 LOCAL(ashlsi3_29):
 461         shll2   r0
 462 LOCAL(ashlsi3_27):
 463         shll2   r0
 464 LOCAL(ashlsi3_25):
 465         shll16  r0
 466         shll8   r0
 467         rts
 468         shll    r0
 469
 470 LOCAL(ashlsi3_0):
 471         rts
 472         nop
 473
 474 #endif
 475
 476 #ifdef L_lshiftrt
 477
 478 !
 479 ! GLOBAL(lshrsi3)
 480 !
 481 ! Entry:
 482 !
 483 ! r4: Value to shift
 484 ! r5: Shifts
 485 !
 486 ! Exit:
 487 !
 488 ! r0: Result
 489 !
 490 ! Destroys:
 491 !
 492 ! (none)
 493 !
 494         .global GLOBAL(lshrsi3)
 495         .align  2
 496 GLOBAL(lshrsi3):
 497         mov     #31,r0
 498         and     r0,r5
 499         mova    LOCAL(lshrsi3_table),r0
 500         mov.b   @(r0,r5),r5
 501 #ifdef __sh1__
 502         add     r5,r0
 503         jmp     @r0
 504 #else
 505         braf    r5
 506 #endif
 507         mov     r4,r0
 508
 509         .align  2
 510 LOCAL(lshrsi3_table):
 511         .byte           LOCAL(lshrsi3_0)-LOCAL(lshrsi3_table)
 512         .byte           LOCAL(lshrsi3_1)-LOCAL(lshrsi3_table)
 513         .byte           LOCAL(lshrsi3_2)-LOCAL(lshrsi3_table)
 514         .byte           LOCAL(lshrsi3_3)-LOCAL(lshrsi3_table)
 515         .byte           LOCAL(lshrsi3_4)-LOCAL(lshrsi3_table)
 516         .byte           LOCAL(lshrsi3_5)-LOCAL(lshrsi3_table)
 517         .byte           LOCAL(lshrsi3_6)-LOCAL(lshrsi3_table)
 518         .byte           LOCAL(lshrsi3_7)-LOCAL(lshrsi3_table)
 519         .byte           LOCAL(lshrsi3_8)-LOCAL(lshrsi3_table)
 520         .byte           LOCAL(lshrsi3_9)-LOCAL(lshrsi3_table)
 521         .byte           LOCAL(lshrsi3_10)-LOCAL(lshrsi3_table)
 522         .byte           LOCAL(lshrsi3_11)-LOCAL(lshrsi3_table)
 523         .byte           LOCAL(lshrsi3_12)-LOCAL(lshrsi3_table)
 524         .byte           LOCAL(lshrsi3_13)-LOCAL(lshrsi3_table)
 525         .byte           LOCAL(lshrsi3_14)-LOCAL(lshrsi3_table)
 526         .byte           LOCAL(lshrsi3_15)-LOCAL(lshrsi3_table)
 527         .byte           LOCAL(lshrsi3_16)-LOCAL(lshrsi3_table)
 528         .byte           LOCAL(lshrsi3_17)-LOCAL(lshrsi3_table)
 529         .byte           LOCAL(lshrsi3_18)-LOCAL(lshrsi3_table)
 530         .byte           LOCAL(lshrsi3_19)-LOCAL(lshrsi3_table)
 531         .byte           LOCAL(lshrsi3_20)-LOCAL(lshrsi3_table)
 532         .byte           LOCAL(lshrsi3_21)-LOCAL(lshrsi3_table)
 533         .byte           LOCAL(lshrsi3_22)-LOCAL(lshrsi3_table)
 534         .byte           LOCAL(lshrsi3_23)-LOCAL(lshrsi3_table)
 535         .byte           LOCAL(lshrsi3_24)-LOCAL(lshrsi3_table)
 536         .byte           LOCAL(lshrsi3_25)-LOCAL(lshrsi3_table)
 537         .byte           LOCAL(lshrsi3_26)-LOCAL(lshrsi3_table)
 538         .byte           LOCAL(lshrsi3_27)-LOCAL(lshrsi3_table)
 539         .byte           LOCAL(lshrsi3_28)-LOCAL(lshrsi3_table)
 540         .byte           LOCAL(lshrsi3_29)-LOCAL(lshrsi3_table)
 541         .byte           LOCAL(lshrsi3_30)-LOCAL(lshrsi3_table)
 542         .byte           LOCAL(lshrsi3_31)-LOCAL(lshrsi3_table)
 543
 544 LOCAL(lshrsi3_6):
 545         shlr2   r0
 546 LOCAL(lshrsi3_4):
 547         shlr2   r0
 548 LOCAL(lshrsi3_2):
 549         rts
 550         shlr2   r0
 551
 552 LOCAL(lshrsi3_7):
 553         shlr2   r0
 554 LOCAL(lshrsi3_5):
 555         shlr2   r0
 556 LOCAL(lshrsi3_3):
 557         shlr2   r0
 558 LOCAL(lshrsi3_1):
 559         rts
 560         shlr    r0
 561
 562 LOCAL(lshrsi3_14):
 563         shlr2   r0
 564 LOCAL(lshrsi3_12):
 565         shlr2   r0
 566 LOCAL(lshrsi3_10):
 567         shlr2   r0
 568 LOCAL(lshrsi3_8):
 569         rts
 570         shlr8   r0
 571
 572 LOCAL(lshrsi3_15):
 573         shlr2   r0
 574 LOCAL(lshrsi3_13):
 575         shlr2   r0
 576 LOCAL(lshrsi3_11):
 577         shlr2   r0
 578 LOCAL(lshrsi3_9):
 579         shlr8   r0
 580         rts
 581         shlr    r0
 582
 583 LOCAL(lshrsi3_22):
 584         shlr2   r0
 585 LOCAL(lshrsi3_20):
 586         shlr2   r0
 587 LOCAL(lshrsi3_18):
 588         shlr2   r0
 589 LOCAL(lshrsi3_16):
 590         rts
 591         shlr16  r0
 592
 593 LOCAL(lshrsi3_23):
 594         shlr2   r0
 595 LOCAL(lshrsi3_21):
 596         shlr2   r0
 597 LOCAL(lshrsi3_19):
 598         shlr2   r0
 599 LOCAL(lshrsi3_17):
 600         shlr16  r0
 601         rts
 602         shlr    r0
 603
 604 LOCAL(lshrsi3_30):
 605         shlr2   r0
 606 LOCAL(lshrsi3_28):
 607         shlr2   r0
 608 LOCAL(lshrsi3_26):
 609         shlr2   r0
 610 LOCAL(lshrsi3_24):
 611         shlr16  r0
 612         rts
 613         shlr8   r0
 614
 615 LOCAL(lshrsi3_31):
 616         shlr2   r0
 617 LOCAL(lshrsi3_29):
 618         shlr2   r0
 619 LOCAL(lshrsi3_27):
 620         shlr2   r0
 621 LOCAL(lshrsi3_25):
 622         shlr16  r0
 623         shlr8   r0
 624         rts
 625         shlr    r0
 626
 627 LOCAL(lshrsi3_0):
 628         rts
 629         nop
 630
 631 #endif
 632
 633 #ifdef L_movstr
 634         .text
 635 ! done all the large groups, do the remainder
 636
 637 ! jump to movstr+
 638 done:
 639         add     #64,r5
 640         mova    GLOBAL(movstrSI0),r0
 641         shll2   r6
 642         add     r6,r0
 643         jmp     @r0
 644         add     #64,r4
 645         .align  4
 646         .global GLOBAL(movstrSI64)
 647 GLOBAL(movstrSI64):
 648         mov.l   @(60,r5),r0
 649         mov.l   r0,@(60,r4)
 650         .global GLOBAL(movstrSI60)
 651 GLOBAL(movstrSI60):
 652         mov.l   @(56,r5),r0
 653         mov.l   r0,@(56,r4)
 654         .global GLOBAL(movstrSI56)
 655 GLOBAL(movstrSI56):
 656         mov.l   @(52,r5),r0
 657         mov.l   r0,@(52,r4)
 658         .global GLOBAL(movstrSI52)
 659 GLOBAL(movstrSI52):
 660         mov.l   @(48,r5),r0
 661         mov.l   r0,@(48,r4)
 662         .global GLOBAL(movstrSI48)
 663 GLOBAL(movstrSI48):
 664         mov.l   @(44,r5),r0
 665         mov.l   r0,@(44,r4)
 666         .global GLOBAL(movstrSI44)
 667 GLOBAL(movstrSI44):
 668         mov.l   @(40,r5),r0
 669         mov.l   r0,@(40,r4)
 670         .global GLOBAL(movstrSI40)
 671 GLOBAL(movstrSI40):
 672         mov.l   @(36,r5),r0
 673         mov.l   r0,@(36,r4)
 674         .global GLOBAL(movstrSI36)
 675 GLOBAL(movstrSI36):
 676         mov.l   @(32,r5),r0
 677         mov.l   r0,@(32,r4)
 678         .global GLOBAL(movstrSI32)
 679 GLOBAL(movstrSI32):
 680         mov.l   @(28,r5),r0
 681         mov.l   r0,@(28,r4)
 682         .global GLOBAL(movstrSI28)
 683 GLOBAL(movstrSI28):
 684         mov.l   @(24,r5),r0
 685         mov.l   r0,@(24,r4)
 686         .global GLOBAL(movstrSI24)
 687 GLOBAL(movstrSI24):
 688         mov.l   @(20,r5),r0
 689         mov.l   r0,@(20,r4)
 690         .global GLOBAL(movstrSI20)
 691 GLOBAL(movstrSI20):
 692         mov.l   @(16,r5),r0
 693         mov.l   r0,@(16,r4)
 694         .global GLOBAL(movstrSI16)
 695 GLOBAL(movstrSI16):
 696         mov.l   @(12,r5),r0
 697         mov.l   r0,@(12,r4)
 698         .global GLOBAL(movstrSI12)
 699 GLOBAL(movstrSI12):
 700         mov.l   @(8,r5),r0
 701         mov.l   r0,@(8,r4)
 702         .global GLOBAL(movstrSI8)
 703 GLOBAL(movstrSI8):
 704         mov.l   @(4,r5),r0
 705         mov.l   r0,@(4,r4)
 706         .global GLOBAL(movstrSI4)
 707 GLOBAL(movstrSI4):
 708         mov.l   @(0,r5),r0
 709         mov.l   r0,@(0,r4)
 710 GLOBAL(movstrSI0):
 711         rts
 712         nop
 713
 714         .align  4
 715
 716         .global GLOBAL(movstr)
 717 GLOBAL(movstr):
 718         mov.l   @(60,r5),r0
 719         mov.l   r0,@(60,r4)
 720
 721         mov.l   @(56,r5),r0
 722         mov.l   r0,@(56,r4)
 723
 724         mov.l   @(52,r5),r0
 725         mov.l   r0,@(52,r4)
 726
 727         mov.l   @(48,r5),r0
 728         mov.l   r0,@(48,r4)
 729
 730         mov.l   @(44,r5),r0
 731         mov.l   r0,@(44,r4)
 732
 733         mov.l   @(40,r5),r0
 734         mov.l   r0,@(40,r4)
 735
 736         mov.l   @(36,r5),r0
 737         mov.l   r0,@(36,r4)
 738
 739         mov.l   @(32,r5),r0
 740         mov.l   r0,@(32,r4)
 741
 742         mov.l   @(28,r5),r0
 743         mov.l   r0,@(28,r4)
 744
 745         mov.l   @(24,r5),r0
 746         mov.l   r0,@(24,r4)
 747
 748         mov.l   @(20,r5),r0
 749         mov.l   r0,@(20,r4)
 750
 751         mov.l   @(16,r5),r0
 752         mov.l   r0,@(16,r4)
 753
 754         mov.l   @(12,r5),r0
 755         mov.l   r0,@(12,r4)
 756
 757         mov.l   @(8,r5),r0
 758         mov.l   r0,@(8,r4)
 759
 760         mov.l   @(4,r5),r0
 761         mov.l   r0,@(4,r4)
 762
 763         mov.l   @(0,r5),r0
 764         mov.l   r0,@(0,r4)
 765
 766         add     #-16,r6
 767         cmp/pl  r6
 768         bf      done
 769
 770         add     #64,r5
 771         bra     GLOBAL(movstr)
 772         add     #64,r4
 773 #endif
 774
 775 #ifdef L_movstr_i4
 776         .text
 777         .global GLOBAL(movstr_i4_even)
 778         .global GLOBAL(movstr_i4_odd)
 779         .global GLOBAL(movstrSI12_i4)
 780
 781         .p2align        5
 782 L_movstr_2mod4_end:
 783         mov.l   r0,@(16,r4)
 784         rts
 785         mov.l   r1,@(20,r4)
 786
 787         .p2align        2
 788
 789 GLOBAL(movstr_i4_odd):
 790         mov.l   @r5+,r1
 791         add     #-4,r4
 792         mov.l   @r5+,r2
 793         mov.l   @r5+,r3
 794         mov.l   r1,@(4,r4)
 795         mov.l   r2,@(8,r4)
 796
 797 L_movstr_loop:
 798         mov.l   r3,@(12,r4)
 799         dt      r6
 800         mov.l   @r5+,r0
 801         bt/s    L_movstr_2mod4_end
 802         mov.l   @r5+,r1
 803         add     #16,r4
 804 L_movstr_start_even:
 805         mov.l   @r5+,r2
 806         mov.l   @r5+,r3
 807         mov.l   r0,@r4
 808         dt      r6
 809         mov.l   r1,@(4,r4)
 810         bf/s    L_movstr_loop
 811         mov.l   r2,@(8,r4)
 812         rts
 813         mov.l   r3,@(12,r4)
 814
 815 GLOBAL(movstr_i4_even):
 816         mov.l   @r5+,r0
 817         bra     L_movstr_start_even
 818         mov.l   @r5+,r1
 819
 820         .p2align        4
 821 GLOBAL(movstrSI12_i4):
 822         mov.l   @r5,r0
 823         mov.l   @(4,r5),r1
 824         mov.l   @(8,r5),r2
 825         mov.l   r0,@r4
 826         mov.l   r1,@(4,r4)
 827         rts
 828         mov.l   r2,@(8,r4)
 829 #endif
 830
 831 #ifdef L_mulsi3
 832
 833
 834         .global GLOBAL(mulsi3)
 835
 836 ! r4 =       aabb
 837 ! r5 =       ccdd
 838 ! r0 = aabb*ccdd  via partial products
 839 !
 840 ! if aa == 0 and cc = 0
 841 ! r0 = bb*dd
 842 !
 843 ! else
 844 ! aa = bb*dd + (aa*dd*65536) + (cc*bb*65536)
 845 !
 846
 847 GLOBAL(mulsi3):
 848         mulu.w  r4,r5           ! multiply the lsws  macl=bb*dd
 849         mov     r5,r3           ! r3 = ccdd
 850         swap.w  r4,r2           ! r2 = bbaa
 851         xtrct   r2,r3           ! r3 = aacc
 852         tst     r3,r3           ! msws zero ?
 853         bf      hiset
 854         rts                     ! yes - then we have the answer
 855         sts     macl,r0
 856
 857 hiset:  sts     macl,r0         ! r0 = bb*dd
 858         mulu.w  r2,r5           ! brewing macl = aa*dd
 859         sts     macl,r1
 860         mulu.w  r3,r4           ! brewing macl = cc*bb
 861         sts     macl,r2
 862         add     r1,r2
 863         shll16  r2
 864         rts
 865         add     r2,r0
 866
 867
 868 #endif
 869 #ifdef L_sdivsi3_i4
 870         .title "SH DIVIDE"
 871 !! 4 byte integer Divide code for the Hitachi SH
 872 #ifdef __SH4__
 873 !! args in r4 and r5, result in fpul, clobber dr0, dr2
 874
 875         .global GLOBAL(sdivsi3_i4)
 876 GLOBAL(sdivsi3_i4):
 877         lds r4,fpul
 878         float fpul,dr0
 879         lds r5,fpul
 880         float fpul,dr2
 881         fdiv dr2,dr0
 882         rts
 883         ftrc dr0,fpul
 884
 885 #elif defined(__SH4_SINGLE__) || defined(__SH4_SINGLE_ONLY__)
 886 !! args in r4 and r5, result in fpul, clobber r2, dr0, dr2
 887
 888         .global GLOBAL(sdivsi3_i4)
 889 GLOBAL(sdivsi3_i4):
 890         sts.l fpscr,@-r15
 891         mov #8,r2
 892         swap.w r2,r2
 893         lds r2,fpscr
 894         lds r4,fpul
 895         float fpul,dr0
 896         lds r5,fpul
 897         float fpul,dr2
 898         fdiv dr2,dr0
 899         ftrc dr0,fpul
 900         rts
 901         lds.l @r15+,fpscr
 902
 903 #endif /* ! __SH4__ */
 904 #endif
 905
 906 #ifdef L_sdivsi3
 907 /* __SH4_SINGLE_ONLY__ keeps this part for link compatibility with
 908    sh3e code.  */
 909 #if (! defined(__SH4__) && ! defined (__SH4_SINGLE__)) || defined (__linux__)
 910 !!
 911 !! Steve Chamberlain
 912 !! sac@cygnus.com
 913 !!
 914 !!
 915
 916 !! args in r4 and r5, result in r0 clobber r1,r2,r3
 917
 918         .global GLOBAL(sdivsi3)
 919 GLOBAL(sdivsi3):
 920         mov     r4,r1
 921         mov     r5,r0
 922
 923         tst     r0,r0
 924         bt      div0
 925         mov     #0,r2
 926         div0s   r2,r1
 927         subc    r3,r3
 928         subc    r2,r1
 929         div0s   r0,r3
 930         rotcl   r1
 931         div1    r0,r3
 932         rotcl   r1
 933         div1    r0,r3
 934         rotcl   r1
 935         div1    r0,r3
 936         rotcl   r1
 937         div1    r0,r3
 938         rotcl   r1
 939         div1    r0,r3
 940         rotcl   r1
 941         div1    r0,r3
 942         rotcl   r1
 943         div1    r0,r3
 944         rotcl   r1
 945         div1    r0,r3
 946         rotcl   r1
 947         div1    r0,r3
 948         rotcl   r1
 949         div1    r0,r3
 950         rotcl   r1
 951         div1    r0,r3
 952         rotcl   r1
 953         div1    r0,r3
 954         rotcl   r1
 955         div1    r0,r3
 956         rotcl   r1
 957         div1    r0,r3
 958         rotcl   r1
 959         div1    r0,r3
 960         rotcl   r1
 961         div1    r0,r3
 962         rotcl   r1
 963         div1    r0,r3
 964         rotcl   r1
 965         div1    r0,r3
 966         rotcl   r1
 967         div1    r0,r3
 968         rotcl   r1
 969         div1    r0,r3
 970         rotcl   r1
 971         div1    r0,r3
 972         rotcl   r1
 973         div1    r0,r3
 974         rotcl   r1
 975         div1    r0,r3
 976         rotcl   r1
 977         div1    r0,r3
 978         rotcl   r1
 979         div1    r0,r3
 980         rotcl   r1
 981         div1    r0,r3
 982         rotcl   r1
 983         div1    r0,r3
 984         rotcl   r1
 985         div1    r0,r3
 986         rotcl   r1
 987         div1    r0,r3
 988         rotcl   r1
 989         div1    r0,r3
 990         rotcl   r1
 991         div1    r0,r3
 992         rotcl   r1
 993         div1    r0,r3
 994         rotcl   r1
 995         addc    r2,r1
 996         rts
 997         mov     r1,r0
 998
 999
1000 div0:   rts
1001         mov     #0,r0
1002
1003 #endif /* ! __SH4__ */
1004 #endif
1005 #ifdef L_udivsi3_i4
1006
1007         .title "SH DIVIDE"
1008 !! 4 byte integer Divide code for the Hitachi SH
1009 #ifdef __SH4__
1010 !! args in r4 and r5, result in fpul, clobber r0, r1, r4, r5, dr0, dr2, dr4
1011
1012         .global GLOBAL(udivsi3_i4)
1013 GLOBAL(udivsi3_i4):
1014         mov #1,r1
1015         cmp/hi r1,r5
1016         bf trivial
1017         rotr r1
1018         xor r1,r4
1019         lds r4,fpul
1020         mova L1,r0
1021 #ifdef FMOVD_WORKS
1022         fmov.d @r0+,dr4
1023 #else
1024 #ifdef __LITTLE_ENDIAN__
1025         fmov.s @r0+,fr5
1026         fmov.s @r0,fr4
1027 #else
1028         fmov.s @r0+,fr4
1029         fmov.s @r0,fr5
1030 #endif
1031 #endif
1032         float fpul,dr0
1033         xor r1,r5
1034         lds r5,fpul
1035         float fpul,dr2
1036         fadd dr4,dr0
1037         fadd dr4,dr2
1038         fdiv dr2,dr0
1039         rts
1040         ftrc dr0,fpul
1041
1042 trivial:
1043         rts
1044         lds r4,fpul
1045
1046         .align 2
1047 #ifdef FMOVD_WORKS
1048         .align 3        ! make double below 8 byte aligned.
1049 #endif
1050 L1:
1051         .double 2147483648
1052
1053 #elif defined(__SH4_SINGLE__) || defined(__SH4_SINGLE_ONLY__)
1054 !! args in r4 and r5, result in fpul, clobber r0, r1, r4, r5, dr0, dr2, dr4
1055
1056         .global GLOBAL(udivsi3_i4)
1057 GLOBAL(udivsi3_i4):
1058         mov #1,r1
1059         cmp/hi r1,r5
1060         bf trivial
1061         sts.l fpscr,@-r15
1062         mova L1,r0
1063         lds.l @r0+,fpscr
1064         rotr r1
1065         xor r1,r4
1066         lds r4,fpul
1067 #ifdef FMOVD_WORKS
1068         fmov.d @r0+,dr4
1069 #else
1070 #ifdef __LITTLE_ENDIAN__
1071         fmov.s @r0+,fr5
1072         fmov.s @r0,fr4
1073 #else
1074         fmov.s @r0+,fr4
1075         fmov.s @r0,fr5
1076 #endif
1077 #endif
1078         float fpul,dr0
1079         xor r1,r5
1080         lds r5,fpul
1081         float fpul,dr2
1082         fadd dr4,dr0
1083         fadd dr4,dr2
1084         fdiv dr2,dr0
1085         ftrc dr0,fpul
1086         rts
1087         lds.l @r15+,fpscr
1088
1089 #ifdef FMOVD_WORKS
1090         .align 3        ! make double below 8 byte aligned.
1091 #endif
1092 trivial:
1093         rts
1094         lds r4,fpul
1095
1096         .align 2
1097 L1:
1098 #ifndef FMOVD_WORKS
1099         .long 0x80000
1100 #else
1101         .long 0x180000
1102 #endif
1103         .double 2147483648
1104
1105 #endif /* ! __SH4__ */
1106 #endif
1107
1108 #ifdef L_udivsi3
1109 /* __SH4_SINGLE_ONLY__ keeps this part for link compatibility with
1110    sh3e code.  */
1111 #if (! defined(__SH4__) && ! defined (__SH4_SINGLE__)) || defined (__linux__)
1112 !!
1113 !! Steve Chamberlain
1114 !! sac@cygnus.com
1115 !!
1116 !!
1117
1118 !! args in r4 and r5, result in r0, clobbers r4, pr, and t bit
1119         .global GLOBAL(udivsi3)
1120
1121 GLOBAL(udivsi3):
1122 longway:
1123         mov     #0,r0
1124         div0u
1125         ! get one bit from the msb of the numerator into the T
1126         ! bit and divide it by whats in r5.  Put the answer bit
1127         ! into the T bit so it can come out again at the bottom
1128
1129         rotcl   r4 ; div1 r5,r0
1130         rotcl   r4 ; div1 r5,r0
1131         rotcl   r4 ; div1 r5,r0
1132         rotcl   r4 ; div1 r5,r0
1133         rotcl   r4 ; div1 r5,r0
1134         rotcl   r4 ; div1 r5,r0
1135         rotcl   r4 ; div1 r5,r0
1136         rotcl   r4 ; div1 r5,r0
1137
1138         rotcl   r4 ; div1 r5,r0
1139         rotcl   r4 ; div1 r5,r0
1140         rotcl   r4 ; div1 r5,r0
1141         rotcl   r4 ; div1 r5,r0
1142         rotcl   r4 ; div1 r5,r0
1143         rotcl   r4 ; div1 r5,r0
1144         rotcl   r4 ; div1 r5,r0
1145         rotcl   r4 ; div1 r5,r0
1146 shortway:
1147         rotcl   r4 ; div1 r5,r0
1148         rotcl   r4 ; div1 r5,r0
1149         rotcl   r4 ; div1 r5,r0
1150         rotcl   r4 ; div1 r5,r0
1151         rotcl   r4 ; div1 r5,r0
1152         rotcl   r4 ; div1 r5,r0
1153         rotcl   r4 ; div1 r5,r0
1154         rotcl   r4 ; div1 r5,r0
1155
1156 vshortway:
1157         rotcl   r4 ; div1 r5,r0
1158         rotcl   r4 ; div1 r5,r0
1159         rotcl   r4 ; div1 r5,r0
1160         rotcl   r4 ; div1 r5,r0
1161         rotcl   r4 ; div1 r5,r0
1162         rotcl   r4 ; div1 r5,r0
1163         rotcl   r4 ; div1 r5,r0
1164         rotcl   r4 ; div1 r5,r0
1165         rotcl   r4
1166 ret:    rts
1167         mov     r4,r0
1168
1169 #endif /* __SH4__ */
1170 #endif
1171 #ifdef L_set_fpscr
1172 #if defined (__SH3E__) || defined(__SH4_SINGLE__) || defined(__SH4__) || defined(__SH4_SINGLE_ONLY__)
1173         .global GLOBAL(set_fpscr)
1174 GLOBAL(set_fpscr):
1175         lds r4,fpscr
1176         mov.l LOCAL(set_fpscr_L1),r1
1177         swap.w r4,r0
1178         or #24,r0
1179 #ifndef FMOVD_WORKS
1180         xor #16,r0
1181 #endif
1182 #if defined(__SH4__)
1183         swap.w r0,r3
1184         mov.l r3,@(4,r1)
1185 #else /* defined(__SH3E__) || defined(__SH4_SINGLE*__) */
1186         swap.w r0,r2
1187         mov.l r2,@r1
1188 #endif
1189 #ifndef FMOVD_WORKS
1190         xor #8,r0
1191 #else
1192         xor #24,r0
1193 #endif
1194 #if defined(__SH4__)
1195         swap.w r0,r2
1196         rts
1197         mov.l r2,@r1
1198 #else /* defined(__SH3E__) || defined(__SH4_SINGLE*__) */
1199         swap.w r0,r3
1200         rts
1201         mov.l r3,@(4,r1)
1202 #endif
1203         .align 2
1204 LOCAL(set_fpscr_L1):
1205         .long GLOBAL(fpscr_values)
1206 #ifdef __ELF__
1207         .comm   GLOBAL(fpscr_values),8,4
1208 #else
1209         .comm   GLOBAL(fpscr_values),8
1210 #endif /* ELF */
1211 #endif /* SH3E / SH4 */
1212 #endif /* L_set_fpscr */
1213 #ifdef L_ic_invalidate
1214 #if defined(__SH4_SINGLE__) || defined(__SH4__) || defined(__SH4_SINGLE_ONLY__)
1215         .global GLOBAL(ic_invalidate)
1216 GLOBAL(ic_invalidate):
1217         ocbwb   @r4
1218         mova    0f,r0
1219         mov.w   1f,r1
1220 /* Compute how many cache lines 0f is away from r4.  */
1221         sub     r0,r4
1222         and     r1,r4
1223 /* Prepare to branch to 0f plus the cache-line offset.  */
1224         add     # 0f - 1f,r4
1225         braf    r4
1226         nop
1227 1:
1228         .short  0x1fe0
1229         .p2align 5
1230 /* This must be aligned to the beginning of a cache line.  */
1231 0:
1232         .rept   256 /* There are 256 cache lines of 32 bytes.  */
1233         rts
1234         .rept   15
1235         nop
1236         .endr
1237         .endr
1238 #endif /* SH4 */
1239 #endif /* L_ic_invalidate */