src/mesa/x86/x86_xform4.S

   1
   2 /*
   3  * Mesa 3-D graphics library
   4  * Version:  3.5
   5  *
   6  * Copyright (C) 1999-2001  Brian Paul   All Rights Reserved.
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice shall be included
  16  * in all copies or substantial portions of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  21  * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  22  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  23  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  24  */
  25
  26 /*
  27  * NOTE: Avoid using spaces in between '(' ')' and arguments, especially
  28  * with macros like CONST, LLBL that expand to CONCAT(...).  Putting spaces
  29  * in there will break the build on some platforms.
  30  */
  31
  32 #include "matypes.h"
  33 #include "xform_args.h"
  34
  35         SEG_TEXT
  36
  37 #define FP_ONE          1065353216
  38 #define FP_ZERO         0
  39
  40 #define SRC0            REGOFF(0, ESI)
  41 #define SRC1            REGOFF(4, ESI)
  42 #define SRC2            REGOFF(8, ESI)
  43 #define SRC3            REGOFF(12, ESI)
  44 #define DST0            REGOFF(0, EDI)
  45 #define DST1            REGOFF(4, EDI)
  46 #define DST2            REGOFF(8, EDI)
  47 #define DST3            REGOFF(12, EDI)
  48 #define MAT0            REGOFF(0, EDX)
  49 #define MAT1            REGOFF(4, EDX)
  50 #define MAT2            REGOFF(8, EDX)
  51 #define MAT3            REGOFF(12, EDX)
  52 #define MAT4            REGOFF(16, EDX)
  53 #define MAT5            REGOFF(20, EDX)
  54 #define MAT6            REGOFF(24, EDX)
  55 #define MAT7            REGOFF(28, EDX)
  56 #define MAT8            REGOFF(32, EDX)
  57 #define MAT9            REGOFF(36, EDX)
  58 #define MAT10           REGOFF(40, EDX)
  59 #define MAT11           REGOFF(44, EDX)
  60 #define MAT12           REGOFF(48, EDX)
  61 #define MAT13           REGOFF(52, EDX)
  62 #define MAT14           REGOFF(56, EDX)
  63 #define MAT15           REGOFF(60, EDX)
  64
  65
  66 ALIGNTEXT16
  67 GLOBL GLNAME( _mesa_x86_transform_points4_general )
  68 HIDDEN(_mesa_x86_transform_points4_general)
  69 GLNAME( _mesa_x86_transform_points4_general ):
  70
  71 #define FRAME_OFFSET 8
  72         PUSH_L( ESI )
  73         PUSH_L( EDI )
  74
  75         MOV_L( ARG_SOURCE, ESI )
  76         MOV_L( ARG_DEST, EDI )
  77
  78         MOV_L( ARG_MATRIX, EDX )
  79         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
  80
  81         TEST_L( ECX, ECX )
  82         JZ( LLBL(x86_p4_gr_done) )
  83
  84         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
  85         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
  86
  87         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
  88         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
  89
  90         SHL_L( CONST(4), ECX )
  91         MOV_L( REGOFF(V4F_START, ESI), ESI )
  92
  93         MOV_L( REGOFF(V4F_START, EDI), EDI )
  94         ADD_L( EDI, ECX )
  95
  96 ALIGNTEXT16
  97 LLBL(x86_p4_gr_loop):
  98
  99         FLD_S( SRC0 )                   /* F4 */
 100         FMUL_S( MAT0 )
 101         FLD_S( SRC0 )                   /* F5 F4 */
 102         FMUL_S( MAT1 )
 103         FLD_S( SRC0 )                   /* F6 F5 F4 */
 104         FMUL_S( MAT2 )
 105         FLD_S( SRC0 )                   /* F7 F6 F5 F4 */
 106         FMUL_S( MAT3 )
 107
 108         FLD_S( SRC1 )                   /* F0 F7 F6 F5 F4 */
 109         FMUL_S( MAT4 )
 110         FLD_S( SRC1 )                   /* F1 F0 F7 F6 F5 F4 */
 111         FMUL_S( MAT5 )
 112         FLD_S( SRC1 )                   /* F2 F1 F0 F7 F6 F5 F4 */
 113         FMUL_S( MAT6 )
 114         FLD_S( SRC1 )                   /* F3 F2 F1 F0 F7 F6 F5 F4 */
 115         FMUL_S( MAT7 )
 116
 117         FXCH( ST(3) )                   /* F0 F2 F1 F3 F7 F6 F5 F4 */
 118         FADDP( ST0, ST(7) )             /* F2 F1 F3 F7 F6 F5 F4 */
 119         FXCH( ST(1) )                   /* F1 F2 F3 F7 F6 F5 F4 */
 120         FADDP( ST0, ST(5) )             /* F2 F3 F7 F6 F5 F4 */
 121         FADDP( ST0, ST(3) )             /* F3 F7 F6 F5 F4 */
 122         FADDP( ST0, ST(1) )             /* F7 F6 F5 F4 */
 123
 124         FLD_S( SRC2 )                   /* F0 F7 F6 F5 F4 */
 125         FMUL_S( MAT8 )
 126         FLD_S( SRC2 )                   /* F1 F0 F7 F6 F5 F4 */
 127         FMUL_S( MAT9 )
 128         FLD_S( SRC2 )                   /* F2 F1 F0 F7 F6 F5 F4 */
 129         FMUL_S( MAT10 )
 130         FLD_S( SRC2 )                   /* F3 F2 F1 F0 F7 F6 F5 F4 */
 131         FMUL_S( MAT11 )
 132
 133         FXCH( ST(3) )                   /* F0 F2 F1 F3 F7 F6 F5 F4 */
 134         FADDP( ST0, ST(7) )             /* F2 F1 F3 F7 F6 F5 F4 */
 135         FXCH( ST(1) )                   /* F1 F2 F3 F7 F6 F5 F4 */
 136         FADDP( ST0, ST(5) )             /* F2 F3 F7 F6 F5 F4 */
 137         FADDP( ST0, ST(3) )             /* F3 F7 F6 F5 F4 */
 138         FADDP( ST0, ST(1) )             /* F7 F6 F5 F4 */
 139
 140         FLD_S( SRC3 )                   /* F0 F7 F6 F5 F4 */
 141         FMUL_S( MAT12 )
 142         FLD_S( SRC3 )                   /* F1 F0 F7 F6 F5 F4 */
 143         FMUL_S( MAT13 )
 144         FLD_S( SRC3 )                   /* F2 F1 F0 F7 F6 F5 F4 */
 145         FMUL_S( MAT14 )
 146         FLD_S( SRC3 )                   /* F3 F2 F1 F0 F7 F6 F5 F4 */
 147         FMUL_S( MAT15 )
 148
 149         FXCH( ST(3) )                   /* F0 F2 F1 F3 F7 F6 F5 F4 */
 150         FADDP( ST0, ST(7) )             /* F2 F1 F3 F7 F6 F5 F4 */
 151         FXCH( ST(1) )                   /* F1 F2 F3 F7 F6 F5 F4 */
 152         FADDP( ST0, ST(5) )             /* F2 F3 F7 F6 F5 F4 */
 153         FADDP( ST0, ST(3) )             /* F3 F7 F6 F5 F4 */
 154         FADDP( ST0, ST(1) )             /* F7 F6 F5 F4 */
 155
 156         FXCH( ST(3) )                   /* F4 F6 F5 F7 */
 157         FSTP_S( DST0 )          /* F6 F5 F7 */
 158         FXCH( ST(1) )                   /* F5 F6 F7 */
 159         FSTP_S( DST1 )          /* F6 F7 */
 160         FSTP_S( DST2 )          /* F7 */
 161         FSTP_S( DST3 )          /* */
 162
 163 LLBL(x86_p4_gr_skip):
 164
 165         ADD_L( CONST(16), EDI )
 166         ADD_L( EAX, ESI )
 167         CMP_L( ECX, EDI )
 168         JNE( LLBL(x86_p4_gr_loop) )
 169
 170 LLBL(x86_p4_gr_done):
 171
 172         POP_L( EDI )
 173         POP_L( ESI )
 174         RET
 175 #undef FRAME_OFFSET
 176
 177
 178
 179
 180 ALIGNTEXT16
 181 GLOBL GLNAME( _mesa_x86_transform_points4_perspective )
 182 HIDDEN(_mesa_x86_transform_points4_perspective)
 183 GLNAME( _mesa_x86_transform_points4_perspective ):
 184
 185 #define FRAME_OFFSET 12
 186         PUSH_L( ESI )
 187         PUSH_L( EDI )
 188         PUSH_L( EBX )
 189
 190         MOV_L( ARG_SOURCE, ESI )
 191         MOV_L( ARG_DEST, EDI )
 192
 193         MOV_L( ARG_MATRIX, EDX )
 194         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 195
 196         TEST_L( ECX, ECX )
 197         JZ( LLBL(x86_p4_pr_done) )
 198
 199         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 200         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 201
 202         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 203         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 204
 205         SHL_L( CONST(4), ECX )
 206         MOV_L( REGOFF(V4F_START, ESI), ESI )
 207
 208         MOV_L( REGOFF(V4F_START, EDI), EDI )
 209         ADD_L( EDI, ECX )
 210
 211 ALIGNTEXT16
 212 LLBL(x86_p4_pr_loop):
 213
 214         FLD_S( SRC0 )                   /* F4 */
 215         FMUL_S( MAT0 )
 216
 217         FLD_S( SRC1 )                   /* F5 F4 */
 218         FMUL_S( MAT5 )
 219
 220         FLD_S( SRC2 )                   /* F0 F5 F4 */
 221         FMUL_S( MAT8 )
 222         FLD_S( SRC2 )                   /* F1 F0 F5 F4 */
 223         FMUL_S( MAT9 )
 224         FLD_S( SRC2 )                   /* F6 F1 F0 F5 F4 */
 225         FMUL_S( MAT10 )
 226
 227         FXCH( ST(2) )                   /* F0 F1 F6 F5 F4 */
 228         FADDP( ST0, ST(4) )             /* F1 F6 F5 F4 */
 229         FADDP( ST0, ST(2) )             /* F6 F5 F4 */
 230
 231         FLD_S( SRC3 )                   /* F2 F6 F5 F4 */
 232         FMUL_S( MAT14 )
 233
 234         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 235
 236         MOV_L( SRC2, EBX )
 237         XOR_L( CONST(-2147483648), EBX )/* change sign */
 238
 239         FXCH( ST(2) )                   /* F4 F5 F6 */
 240         FSTP_S( DST0 )          /* F5 F6 */
 241         FSTP_S( DST1 )          /* F6 */
 242         FSTP_S( DST2 )          /* */
 243         MOV_L( EBX, DST3 )
 244
 245 LLBL(x86_p4_pr_skip):
 246
 247         ADD_L( CONST(16), EDI )
 248         ADD_L( EAX, ESI )
 249         CMP_L( ECX, EDI )
 250         JNE( LLBL(x86_p4_pr_loop) )
 251
 252 LLBL(x86_p4_pr_done):
 253
 254         POP_L( EBX )
 255         POP_L( EDI )
 256         POP_L( ESI )
 257         RET
 258 #undef FRAME_OFFSET
 259
 260
 261
 262
 263 ALIGNTEXT16
 264 GLOBL GLNAME( _mesa_x86_transform_points4_3d )
 265 HIDDEN(_mesa_x86_transform_points4_3d)
 266 GLNAME( _mesa_x86_transform_points4_3d ):
 267
 268 #define FRAME_OFFSET 12
 269         PUSH_L( ESI )
 270         PUSH_L( EDI )
 271         PUSH_L( EBX )
 272
 273         MOV_L( ARG_SOURCE, ESI )
 274         MOV_L( ARG_DEST, EDI )
 275
 276         MOV_L( ARG_MATRIX, EDX )
 277         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 278
 279         TEST_L( ECX, ECX )
 280         JZ( LLBL(x86_p4_3dr_done) )
 281
 282         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 283         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 284
 285         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 286         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 287
 288         SHL_L( CONST(4), ECX )
 289         MOV_L( REGOFF(V4F_START, ESI), ESI )
 290
 291         MOV_L( REGOFF(V4F_START, EDI), EDI )
 292         ADD_L( EDI, ECX )
 293
 294 ALIGNTEXT16
 295 LLBL(x86_p4_3dr_loop):
 296
 297         FLD_S( SRC0 )                   /* F4 */
 298         FMUL_S( MAT0 )
 299         FLD_S( SRC0 )                   /* F5 F4 */
 300         FMUL_S( MAT1 )
 301         FLD_S( SRC0 )                   /* F6 F5 F4 */
 302         FMUL_S( MAT2 )
 303
 304         FLD_S( SRC1 )                   /* F0 F6 F5 F4 */
 305         FMUL_S( MAT4 )
 306         FLD_S( SRC1 )                   /* F1 F0 F6 F5 F4 */
 307         FMUL_S( MAT5 )
 308         FLD_S( SRC1 )                   /* F2 F1 F0 F6 F5 F4 */
 309         FMUL_S( MAT6 )
 310
 311         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 312         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 313         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 314         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 315
 316         FLD_S( SRC2 )                   /* F0 F6 F5 F4 */
 317         FMUL_S( MAT8 )
 318         FLD_S( SRC2 )                   /* F1 F0 F6 F5 F4 */
 319         FMUL_S( MAT9 )
 320         FLD_S( SRC2 )                   /* F2 F1 F0 F6 F5 F4 */
 321         FMUL_S( MAT10 )
 322
 323         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 324         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 325         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 326         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 327
 328         FLD_S( SRC3 )                   /* F0 F6 F5 F4 */
 329         FMUL_S( MAT12 )
 330         FLD_S( SRC3 )                   /* F1 F0 F6 F5 F4 */
 331         FMUL_S( MAT13 )
 332         FLD_S( SRC3 )                   /* F2 F1 F0 F6 F5 F4 */
 333         FMUL_S( MAT14 )
 334
 335         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 336         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 337         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 338         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 339
 340         MOV_L( SRC3, EBX )
 341
 342         FXCH( ST(2) )                   /* F4 F5 F6 */
 343         FSTP_S( DST0 )          /* F5 F6 */
 344         FSTP_S( DST1 )          /* F6 */
 345         FSTP_S( DST2 )          /* */
 346         MOV_L( EBX, DST3 )
 347
 348 LLBL(x86_p4_3dr_skip):
 349
 350         ADD_L( CONST(16), EDI )
 351         ADD_L( EAX, ESI )
 352         CMP_L( ECX, EDI )
 353         JNE( LLBL(x86_p4_3dr_loop) )
 354
 355 LLBL(x86_p4_3dr_done):
 356
 357         POP_L( EBX )
 358         POP_L( EDI )
 359         POP_L( ESI )
 360         RET
 361 #undef FRAME_OFFSET
 362
 363
 364
 365
 366 ALIGNTEXT16
 367 GLOBL GLNAME(_mesa_x86_transform_points4_3d_no_rot)
 368 HIDDEN(_mesa_x86_transform_points4_3d_no_rot)
 369 GLNAME(_mesa_x86_transform_points4_3d_no_rot):
 370
 371 #define FRAME_OFFSET 12
 372         PUSH_L( ESI )
 373         PUSH_L( EDI )
 374         PUSH_L( EBX )
 375
 376         MOV_L( ARG_SOURCE, ESI )
 377         MOV_L( ARG_DEST, EDI )
 378
 379         MOV_L( ARG_MATRIX, EDX )
 380         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 381
 382         TEST_L( ECX, ECX )
 383         JZ( LLBL(x86_p4_3dnrr_done) )
 384
 385         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 386         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 387
 388         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 389         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 390
 391         SHL_L( CONST(4), ECX )
 392         MOV_L( REGOFF(V4F_START, ESI), ESI )
 393
 394         MOV_L( REGOFF(V4F_START, EDI), EDI )
 395         ADD_L( EDI, ECX )
 396
 397 ALIGNTEXT16
 398 LLBL(x86_p4_3dnrr_loop):
 399
 400         FLD_S( SRC0 )                   /* F4 */
 401         FMUL_S( MAT0 )
 402
 403         FLD_S( SRC1 )                   /* F5 F4 */
 404         FMUL_S( MAT5 )
 405
 406         FLD_S( SRC2 )                   /* F6 F5 F4 */
 407         FMUL_S( MAT10 )
 408
 409         FLD_S( SRC3 )                   /* F0 F6 F5 F4 */
 410         FMUL_S( MAT12 )
 411         FLD_S( SRC3 )                   /* F1 F0 F6 F5 F4 */
 412         FMUL_S( MAT13 )
 413         FLD_S( SRC3 )                   /* F2 F1 F0 F6 F5 F4 */
 414         FMUL_S( MAT14 )
 415
 416         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 417         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 418         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 419         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 420
 421         MOV_L( SRC3, EBX )
 422
 423         FXCH( ST(2) )                   /* F4 F5 F6 */
 424         FSTP_S( DST0   )                /* F5 F6 */
 425         FSTP_S( DST1   )                /* F6 */
 426         FSTP_S( DST2   )                /* */
 427         MOV_L( EBX, DST3 )
 428
 429 LLBL(x86_p4_3dnrr_skip):
 430
 431         ADD_L( CONST(16), EDI )
 432         ADD_L( EAX, ESI )
 433         CMP_L( ECX, EDI )
 434         JNE( LLBL(x86_p4_3dnrr_loop) )
 435
 436 LLBL(x86_p4_3dnrr_done):
 437
 438         POP_L( EBX )
 439         POP_L( EDI )
 440         POP_L( ESI )
 441         RET
 442 #undef FRAME_OFFSET
 443
 444
 445
 446
 447 ALIGNTEXT16
 448 GLOBL GLNAME( _mesa_x86_transform_points4_2d )
 449 HIDDEN(_mesa_x86_transform_points4_2d)
 450 GLNAME( _mesa_x86_transform_points4_2d ):
 451
 452 #define FRAME_OFFSET 16
 453         PUSH_L( ESI )
 454         PUSH_L( EDI )
 455         PUSH_L( EBX )
 456         PUSH_L( EBP )
 457
 458         MOV_L( ARG_SOURCE, ESI )
 459         MOV_L( ARG_DEST, EDI )
 460
 461         MOV_L( ARG_MATRIX, EDX )
 462         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 463
 464         TEST_L( ECX, ECX )
 465         JZ( LLBL(x86_p4_2dr_done) )
 466
 467         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 468         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 469
 470         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 471         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 472
 473         SHL_L( CONST(4), ECX )
 474         MOV_L( REGOFF(V4F_START, ESI), ESI )
 475
 476         MOV_L( REGOFF(V4F_START, EDI), EDI )
 477         ADD_L( EDI, ECX )
 478
 479 ALIGNTEXT16
 480 LLBL(x86_p4_2dr_loop):
 481
 482         FLD_S( SRC0 )                   /* F4 */
 483         FMUL_S( MAT0 )
 484         FLD_S( SRC0 )                   /* F5 F4 */
 485         FMUL_S( MAT1 )
 486
 487         FLD_S( SRC1 )                   /* F0 F5 F4 */
 488         FMUL_S( MAT4 )
 489         FLD_S( SRC1 )                   /* F1 F0 F5 F4 */
 490         FMUL_S( MAT5 )
 491
 492         FXCH( ST(1) )                   /* F0 F1 F5 F4 */
 493         FADDP( ST0, ST(3) )             /* F1 F5 F4 */
 494         FADDP( ST0, ST(1) )             /* F5 F4 */
 495
 496         FLD_S( SRC3 )                   /* F0 F5 F4 */
 497         FMUL_S( MAT12 )
 498         FLD_S( SRC3 )                   /* F1 F0 F5 F4 */
 499         FMUL_S( MAT13 )
 500
 501         FXCH( ST(1) )                   /* F0 F1 F5 F4 */
 502         FADDP( ST0, ST(3) )             /* F1 F5 F4 */
 503         FADDP( ST0, ST(1) )             /* F5 F4 */
 504
 505         MOV_L( SRC2, EBX )
 506         MOV_L( SRC3, EBP )
 507
 508         FXCH( ST(1) )                   /* F4 F5 */
 509         FSTP_S( DST0 )          /* F5 */
 510         FSTP_S( DST1 )          /* */
 511         MOV_L( EBX, DST2 )
 512         MOV_L( EBP, DST3 )
 513
 514 LLBL(x86_p4_2dr_skip):
 515
 516         ADD_L( CONST(16), EDI )
 517         ADD_L( EAX, ESI )
 518         CMP_L( ECX, EDI )
 519         JNE( LLBL(x86_p4_2dr_loop) )
 520
 521 LLBL(x86_p4_2dr_done):
 522
 523         POP_L( EBP )
 524         POP_L( EBX )
 525         POP_L( EDI )
 526         POP_L( ESI )
 527         RET
 528 #undef FRAME_OFFSET
 529
 530
 531
 532
 533 ALIGNTEXT16
 534 GLOBL GLNAME( _mesa_x86_transform_points4_2d_no_rot )
 535 HIDDEN(_mesa_x86_transform_points4_2d_no_rot)
 536 GLNAME( _mesa_x86_transform_points4_2d_no_rot ):
 537
 538 #define FRAME_OFFSET 16
 539         PUSH_L( ESI )
 540         PUSH_L( EDI )
 541         PUSH_L( EBX )
 542         PUSH_L( EBP )
 543
 544         MOV_L( ARG_SOURCE, ESI )
 545         MOV_L( ARG_DEST, EDI )
 546
 547         MOV_L( ARG_MATRIX, EDX )
 548         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 549
 550         TEST_L( ECX, ECX )
 551         JZ( LLBL(x86_p4_2dnrr_done) )
 552
 553         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 554         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 555
 556         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 557         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 558
 559         SHL_L( CONST(4), ECX )
 560         MOV_L( REGOFF(V4F_START, ESI), ESI )
 561
 562         MOV_L( REGOFF(V4F_START, EDI), EDI )
 563         ADD_L( EDI, ECX )
 564
 565 ALIGNTEXT16
 566 LLBL(x86_p4_2dnrr_loop):
 567
 568         FLD_S( SRC0 )                   /* F4 */
 569         FMUL_S( MAT0 )
 570
 571         FLD_S( SRC1 )                   /* F5 F4 */
 572         FMUL_S( MAT5 )
 573
 574         FLD_S( SRC3 )                   /* F0 F5 F4 */
 575         FMUL_S( MAT12 )
 576         FLD_S( SRC3 )                   /* F1 F0 F5 F4 */
 577         FMUL_S( MAT13 )
 578
 579         FXCH( ST(1) )                   /* F0 F1 F5 F4 */
 580         FADDP( ST0, ST(3) )             /* F1 F5 F4 */
 581         FADDP( ST0, ST(1) )             /* F5 F4 */
 582
 583         MOV_L( SRC2, EBX )
 584         MOV_L( SRC3, EBP )
 585
 586         FXCH( ST(1) )                   /* F4 F5 */
 587         FSTP_S( DST0   )                /* F5 */
 588         FSTP_S( DST1   )                /* */
 589         MOV_L( EBX, DST2 )
 590         MOV_L( EBP, DST3 )
 591
 592 LLBL(x86_p4_2dnrr_skip):
 593
 594         ADD_L( CONST(16), EDI )
 595         ADD_L( EAX, ESI )
 596         CMP_L( ECX, EDI )
 597         JNE( LLBL(x86_p4_2dnrr_loop) )
 598
 599 LLBL(x86_p4_2dnrr_done):
 600
 601         POP_L( EBP )
 602         POP_L( EBX )
 603         POP_L( EDI )
 604         POP_L( ESI )
 605         RET
 606 #undef FRAME_OFFSET
 607
 608
 609
 610
 611 ALIGNTEXT16
 612 GLOBL GLNAME( _mesa_x86_transform_points4_identity )
 613 HIDDEN(_mesa_x86_transform_points4_identity)
 614 GLNAME( _mesa_x86_transform_points4_identity ):
 615
 616 #define FRAME_OFFSET 12
 617         PUSH_L( ESI )
 618         PUSH_L( EDI )
 619         PUSH_L( EBX )
 620
 621         MOV_L( ARG_SOURCE, ESI )
 622         MOV_L( ARG_DEST, EDI )
 623
 624         MOV_L( ARG_MATRIX, EDX )
 625         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 626
 627         TEST_L( ECX, ECX )
 628         JZ( LLBL(x86_p4_ir_done) )
 629
 630         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 631         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 632
 633         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 634         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 635
 636         SHL_L( CONST(4), ECX )
 637         MOV_L( REGOFF(V4F_START, ESI), ESI )
 638
 639         MOV_L( REGOFF(V4F_START, EDI), EDI )
 640         ADD_L( EDI, ECX )
 641
 642         CMP_L( ESI, EDI )
 643         JE( LLBL(x86_p4_ir_done) )
 644
 645 ALIGNTEXT16
 646 LLBL(x86_p4_ir_loop):
 647
 648         MOV_L( SRC0, EBX )
 649         MOV_L( SRC1, EDX )
 650
 651         MOV_L( EBX, DST0 )
 652         MOV_L( EDX, DST1 )
 653
 654         MOV_L( SRC2, EBX )
 655         MOV_L( SRC3, EDX )
 656
 657         MOV_L( EBX, DST2 )
 658         MOV_L( EDX, DST3 )
 659
 660 LLBL(x86_p4_ir_skip):
 661
 662         ADD_L( CONST(16), EDI )
 663         ADD_L( EAX, ESI )
 664         CMP_L( ECX, EDI )
 665         JNE( LLBL(x86_p4_ir_loop) )
 666
 667 LLBL(x86_p4_ir_done):
 668
 669         POP_L( EBX )
 670         POP_L( EDI )
 671         POP_L( ESI )
 672         RET
 673
 674 #if defined (__ELF__) && defined (__linux__)
 675         .section .note.GNU-stack,"",%progbits
 676 #endif