src/mesa/x86/x86_xform4.S

   1 /* $Id: x86_xform4.S,v 1.2 2002/03/07 21:40:08 brianp Exp $ */
   2
   3 /*
   4  * Mesa 3-D graphics library
   5  * Version:  3.5
   6  *
   7  * Copyright (C) 1999-2001  Brian Paul   All Rights Reserved.
   8  *
   9  * Permission is hereby granted, free of charge, to any person obtaining a
  10  * copy of this software and associated documentation files (the "Software"),
  11  * to deal in the Software without restriction, including without limitation
  12  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  13  * and/or sell copies of the Software, and to permit persons to whom the
  14  * Software is furnished to do so, subject to the following conditions:
  15  *
  16  * The above copyright notice and this permission notice shall be included
  17  * in all copies or substantial portions of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  23  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  24  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  */
  26
  27 /*
  28  * NOTE: Avoid using spaces in between '(' ')' and arguments, especially
  29  * with macros like CONST, LLBL that expand to CONCAT(...).  Putting spaces
  30  * in there will break the build on some platforms.
  31  */
  32
  33 #include "matypes.h"
  34 #include "xform_args.h"
  35
  36         SEG_TEXT
  37
  38 #define FP_ONE          1065353216
  39 #define FP_ZERO         0
  40
  41 #define SRC0            REGOFF(0, ESI)
  42 #define SRC1            REGOFF(4, ESI)
  43 #define SRC2            REGOFF(8, ESI)
  44 #define SRC3            REGOFF(12, ESI)
  45 #define DST0            REGOFF(0, EDI)
  46 #define DST1            REGOFF(4, EDI)
  47 #define DST2            REGOFF(8, EDI)
  48 #define DST3            REGOFF(12, EDI)
  49 #define MAT0            REGOFF(0, EDX)
  50 #define MAT1            REGOFF(4, EDX)
  51 #define MAT2            REGOFF(8, EDX)
  52 #define MAT3            REGOFF(12, EDX)
  53 #define MAT4            REGOFF(16, EDX)
  54 #define MAT5            REGOFF(20, EDX)
  55 #define MAT6            REGOFF(24, EDX)
  56 #define MAT7            REGOFF(28, EDX)
  57 #define MAT8            REGOFF(32, EDX)
  58 #define MAT9            REGOFF(36, EDX)
  59 #define MAT10           REGOFF(40, EDX)
  60 #define MAT11           REGOFF(44, EDX)
  61 #define MAT12           REGOFF(48, EDX)
  62 #define MAT13           REGOFF(52, EDX)
  63 #define MAT14           REGOFF(56, EDX)
  64 #define MAT15           REGOFF(60, EDX)
  65
  66
  67 ALIGNTEXT16
  68 GLOBL GLNAME( _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 GLNAME( _mesa_x86_transform_points4_perspective ):
 183
 184 #define FRAME_OFFSET 12
 185         PUSH_L( ESI )
 186         PUSH_L( EDI )
 187         PUSH_L( EBX )
 188
 189         MOV_L( ARG_SOURCE, ESI )
 190         MOV_L( ARG_DEST, EDI )
 191
 192         MOV_L( ARG_MATRIX, EDX )
 193         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 194
 195         TEST_L( ECX, ECX )
 196         JZ( LLBL(x86_p4_pr_done) )
 197
 198         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 199         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 200
 201         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 202         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 203
 204         SHL_L( CONST(4), ECX )
 205         MOV_L( REGOFF(V4F_START, ESI), ESI )
 206
 207         MOV_L( REGOFF(V4F_START, EDI), EDI )
 208         ADD_L( EDI, ECX )
 209
 210 ALIGNTEXT16
 211 LLBL(x86_p4_pr_loop):
 212
 213         FLD_S( SRC0 )                   /* F4 */
 214         FMUL_S( MAT0 )
 215
 216         FLD_S( SRC1 )                   /* F5 F4 */
 217         FMUL_S( MAT5 )
 218
 219         FLD_S( SRC2 )                   /* F0 F5 F4 */
 220         FMUL_S( MAT8 )
 221         FLD_S( SRC2 )                   /* F1 F0 F5 F4 */
 222         FMUL_S( MAT9 )
 223         FLD_S( SRC2 )                   /* F6 F1 F0 F5 F4 */
 224         FMUL_S( MAT10 )
 225
 226         FXCH( ST(2) )                   /* F0 F1 F6 F5 F4 */
 227         FADDP( ST0, ST(4) )             /* F1 F6 F5 F4 */
 228         FADDP( ST0, ST(2) )             /* F6 F5 F4 */
 229
 230         FLD_S( SRC3 )                   /* F2 F6 F5 F4 */
 231         FMUL_S( MAT14 )
 232
 233         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 234
 235         MOV_L( SRC2, EBX )
 236         XOR_L( CONST(-2147483648), EBX )/* change sign */
 237
 238         FXCH( ST(2) )                   /* F4 F5 F6 */
 239         FSTP_S( DST0 )          /* F5 F6 */
 240         FSTP_S( DST1 )          /* F6 */
 241         FSTP_S( DST2 )          /* */
 242         MOV_L( EBX, DST3 )
 243
 244 LLBL(x86_p4_pr_skip):
 245
 246         ADD_L( CONST(16), EDI )
 247         ADD_L( EAX, ESI )
 248         CMP_L( ECX, EDI )
 249         JNE( LLBL(x86_p4_pr_loop) )
 250
 251 LLBL(x86_p4_pr_done):
 252
 253         POP_L( EBX )
 254         POP_L( EDI )
 255         POP_L( ESI )
 256         RET
 257 #undef FRAME_OFFSET
 258
 259
 260
 261
 262 ALIGNTEXT16
 263 GLOBL GLNAME( _mesa_x86_transform_points4_3d )
 264 GLNAME( _mesa_x86_transform_points4_3d ):
 265
 266 #define FRAME_OFFSET 12
 267         PUSH_L( ESI )
 268         PUSH_L( EDI )
 269         PUSH_L( EBX )
 270
 271         MOV_L( ARG_SOURCE, ESI )
 272         MOV_L( ARG_DEST, EDI )
 273
 274         MOV_L( ARG_MATRIX, EDX )
 275         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 276
 277         TEST_L( ECX, ECX )
 278         JZ( LLBL(x86_p4_3dr_done) )
 279
 280         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 281         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 282
 283         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 284         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 285
 286         SHL_L( CONST(4), ECX )
 287         MOV_L( REGOFF(V4F_START, ESI), ESI )
 288
 289         MOV_L( REGOFF(V4F_START, EDI), EDI )
 290         ADD_L( EDI, ECX )
 291
 292 ALIGNTEXT16
 293 LLBL(x86_p4_3dr_loop):
 294
 295         FLD_S( SRC0 )                   /* F4 */
 296         FMUL_S( MAT0 )
 297         FLD_S( SRC0 )                   /* F5 F4 */
 298         FMUL_S( MAT1 )
 299         FLD_S( SRC0 )                   /* F6 F5 F4 */
 300         FMUL_S( MAT2 )
 301
 302         FLD_S( SRC1 )                   /* F0 F6 F5 F4 */
 303         FMUL_S( MAT4 )
 304         FLD_S( SRC1 )                   /* F1 F0 F6 F5 F4 */
 305         FMUL_S( MAT5 )
 306         FLD_S( SRC1 )                   /* F2 F1 F0 F6 F5 F4 */
 307         FMUL_S( MAT6 )
 308
 309         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 310         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 311         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 312         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 313
 314         FLD_S( SRC2 )                   /* F0 F6 F5 F4 */
 315         FMUL_S( MAT8 )
 316         FLD_S( SRC2 )                   /* F1 F0 F6 F5 F4 */
 317         FMUL_S( MAT9 )
 318         FLD_S( SRC2 )                   /* F2 F1 F0 F6 F5 F4 */
 319         FMUL_S( MAT10 )
 320
 321         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 322         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 323         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 324         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 325
 326         FLD_S( SRC3 )                   /* F0 F6 F5 F4 */
 327         FMUL_S( MAT12 )
 328         FLD_S( SRC3 )                   /* F1 F0 F6 F5 F4 */
 329         FMUL_S( MAT13 )
 330         FLD_S( SRC3 )                   /* F2 F1 F0 F6 F5 F4 */
 331         FMUL_S( MAT14 )
 332
 333         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 334         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 335         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 336         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 337
 338         MOV_L( SRC3, EBX )
 339
 340         FXCH( ST(2) )                   /* F4 F5 F6 */
 341         FSTP_S( DST0 )          /* F5 F6 */
 342         FSTP_S( DST1 )          /* F6 */
 343         FSTP_S( DST2 )          /* */
 344         MOV_L( EBX, DST3 )
 345
 346 LLBL(x86_p4_3dr_skip):
 347
 348         ADD_L( CONST(16), EDI )
 349         ADD_L( EAX, ESI )
 350         CMP_L( ECX, EDI )
 351         JNE( LLBL(x86_p4_3dr_loop) )
 352
 353 LLBL(x86_p4_3dr_done):
 354
 355         POP_L( EBX )
 356         POP_L( EDI )
 357         POP_L( ESI )
 358         RET
 359 #undef FRAME_OFFSET
 360
 361
 362
 363
 364 ALIGNTEXT16
 365 GLOBL GLNAME(_mesa_x86_transform_points4_3d_no_rot)
 366 GLNAME(_mesa_x86_transform_points4_3d_no_rot):
 367
 368 #define FRAME_OFFSET 12
 369         PUSH_L( ESI )
 370         PUSH_L( EDI )
 371         PUSH_L( EBX )
 372
 373         MOV_L( ARG_SOURCE, ESI )
 374         MOV_L( ARG_DEST, EDI )
 375
 376         MOV_L( ARG_MATRIX, EDX )
 377         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 378
 379         TEST_L( ECX, ECX )
 380         JZ( LLBL(x86_p4_3dnrr_done) )
 381
 382         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 383         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 384
 385         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 386         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 387
 388         SHL_L( CONST(4), ECX )
 389         MOV_L( REGOFF(V4F_START, ESI), ESI )
 390
 391         MOV_L( REGOFF(V4F_START, EDI), EDI )
 392         ADD_L( EDI, ECX )
 393
 394 ALIGNTEXT16
 395 LLBL(x86_p4_3dnrr_loop):
 396
 397         FLD_S( SRC0 )                   /* F4 */
 398         FMUL_S( MAT0 )
 399
 400         FLD_S( SRC1 )                   /* F5 F4 */
 401         FMUL_S( MAT5 )
 402
 403         FLD_S( SRC2 )                   /* F6 F5 F4 */
 404         FMUL_S( MAT10 )
 405
 406         FLD_S( SRC3 )                   /* F0 F6 F5 F4 */
 407         FMUL_S( MAT12 )
 408         FLD_S( SRC3 )                   /* F1 F0 F6 F5 F4 */
 409         FMUL_S( MAT13 )
 410         FLD_S( SRC3 )                   /* F2 F1 F0 F6 F5 F4 */
 411         FMUL_S( MAT14 )
 412
 413         FXCH( ST(2) )                   /* F0 F1 F2 F6 F5 F4 */
 414         FADDP( ST0, ST(5) )             /* F1 F2 F6 F5 F4 */
 415         FADDP( ST0, ST(3) )             /* F2 F6 F5 F4 */
 416         FADDP( ST0, ST(1) )             /* F6 F5 F4 */
 417
 418         MOV_L( SRC3, EBX )
 419
 420         FXCH( ST(2) )                   /* F4 F5 F6 */
 421         FSTP_S( DST0   )                /* F5 F6 */
 422         FSTP_S( DST1   )                /* F6 */
 423         FSTP_S( DST2   )                /* */
 424         MOV_L( EBX, DST3 )
 425
 426 LLBL(x86_p4_3dnrr_skip):
 427
 428         ADD_L( CONST(16), EDI )
 429         ADD_L( EAX, ESI )
 430         CMP_L( ECX, EDI )
 431         JNE( LLBL(x86_p4_3dnrr_loop) )
 432
 433 LLBL(x86_p4_3dnrr_done):
 434
 435         POP_L( EBX )
 436         POP_L( EDI )
 437         POP_L( ESI )
 438         RET
 439 #undef FRAME_OFFSET
 440
 441
 442
 443
 444 ALIGNTEXT16
 445 GLOBL GLNAME( _mesa_x86_transform_points4_2d )
 446 GLNAME( _mesa_x86_transform_points4_2d ):
 447
 448 #define FRAME_OFFSET 16
 449         PUSH_L( ESI )
 450         PUSH_L( EDI )
 451         PUSH_L( EBX )
 452         PUSH_L( EBP )
 453
 454         MOV_L( ARG_SOURCE, ESI )
 455         MOV_L( ARG_DEST, EDI )
 456
 457         MOV_L( ARG_MATRIX, EDX )
 458         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 459
 460         TEST_L( ECX, ECX )
 461         JZ( LLBL(x86_p4_2dr_done) )
 462
 463         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 464         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 465
 466         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 467         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 468
 469         SHL_L( CONST(4), ECX )
 470         MOV_L( REGOFF(V4F_START, ESI), ESI )
 471
 472         MOV_L( REGOFF(V4F_START, EDI), EDI )
 473         ADD_L( EDI, ECX )
 474
 475 ALIGNTEXT16
 476 LLBL(x86_p4_2dr_loop):
 477
 478         FLD_S( SRC0 )                   /* F4 */
 479         FMUL_S( MAT0 )
 480         FLD_S( SRC0 )                   /* F5 F4 */
 481         FMUL_S( MAT1 )
 482
 483         FLD_S( SRC1 )                   /* F0 F5 F4 */
 484         FMUL_S( MAT4 )
 485         FLD_S( SRC1 )                   /* F1 F0 F5 F4 */
 486         FMUL_S( MAT5 )
 487
 488         FXCH( ST(1) )                   /* F0 F1 F5 F4 */
 489         FADDP( ST0, ST(3) )             /* F1 F5 F4 */
 490         FADDP( ST0, ST(1) )             /* F5 F4 */
 491
 492         FLD_S( SRC3 )                   /* F0 F5 F4 */
 493         FMUL_S( MAT12 )
 494         FLD_S( SRC3 )                   /* F1 F0 F5 F4 */
 495         FMUL_S( MAT13 )
 496
 497         FXCH( ST(1) )                   /* F0 F1 F5 F4 */
 498         FADDP( ST0, ST(3) )             /* F1 F5 F4 */
 499         FADDP( ST0, ST(1) )             /* F5 F4 */
 500
 501         MOV_L( SRC2, EBX )
 502         MOV_L( SRC3, EBP )
 503
 504         FXCH( ST(1) )                   /* F4 F5 */
 505         FSTP_S( DST0 )          /* F5 */
 506         FSTP_S( DST1 )          /* */
 507         MOV_L( EBX, DST2 )
 508         MOV_L( EBP, DST3 )
 509
 510 LLBL(x86_p4_2dr_skip):
 511
 512         ADD_L( CONST(16), EDI )
 513         ADD_L( EAX, ESI )
 514         CMP_L( ECX, EDI )
 515         JNE( LLBL(x86_p4_2dr_loop) )
 516
 517 LLBL(x86_p4_2dr_done):
 518
 519         POP_L( EBP )
 520         POP_L( EBX )
 521         POP_L( EDI )
 522         POP_L( ESI )
 523         RET
 524 #undef FRAME_OFFSET
 525
 526
 527
 528
 529 ALIGNTEXT16
 530 GLOBL GLNAME( _mesa_x86_transform_points4_2d_no_rot )
 531 GLNAME( _mesa_x86_transform_points4_2d_no_rot ):
 532
 533 #define FRAME_OFFSET 16
 534         PUSH_L( ESI )
 535         PUSH_L( EDI )
 536         PUSH_L( EBX )
 537         PUSH_L( EBP )
 538
 539         MOV_L( ARG_SOURCE, ESI )
 540         MOV_L( ARG_DEST, EDI )
 541
 542         MOV_L( ARG_MATRIX, EDX )
 543         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 544
 545         TEST_L( ECX, ECX )
 546         JZ( LLBL(x86_p4_2dnrr_done) )
 547
 548         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 549         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 550
 551         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 552         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 553
 554         SHL_L( CONST(4), ECX )
 555         MOV_L( REGOFF(V4F_START, ESI), ESI )
 556
 557         MOV_L( REGOFF(V4F_START, EDI), EDI )
 558         ADD_L( EDI, ECX )
 559
 560 ALIGNTEXT16
 561 LLBL(x86_p4_2dnrr_loop):
 562
 563         FLD_S( SRC0 )                   /* F4 */
 564         FMUL_S( MAT0 )
 565
 566         FLD_S( SRC1 )                   /* F5 F4 */
 567         FMUL_S( MAT5 )
 568
 569         FLD_S( SRC3 )                   /* F0 F5 F4 */
 570         FMUL_S( MAT12 )
 571         FLD_S( SRC3 )                   /* F1 F0 F5 F4 */
 572         FMUL_S( MAT13 )
 573
 574         FXCH( ST(1) )                   /* F0 F1 F5 F4 */
 575         FADDP( ST0, ST(3) )             /* F1 F5 F4 */
 576         FADDP( ST0, ST(1) )             /* F5 F4 */
 577
 578         MOV_L( SRC2, EBX )
 579         MOV_L( SRC3, EBP )
 580
 581         FXCH( ST(1) )                   /* F4 F5 */
 582         FSTP_S( DST0   )                /* F5 */
 583         FSTP_S( DST1   )                /* */
 584         MOV_L( EBX, DST2 )
 585         MOV_L( EBP, DST3 )
 586
 587 LLBL(x86_p4_2dnrr_skip):
 588
 589         ADD_L( CONST(16), EDI )
 590         ADD_L( EAX, ESI )
 591         CMP_L( ECX, EDI )
 592         JNE( LLBL(x86_p4_2dnrr_loop) )
 593
 594 LLBL(x86_p4_2dnrr_done):
 595
 596         POP_L( EBP )
 597         POP_L( EBX )
 598         POP_L( EDI )
 599         POP_L( ESI )
 600         RET
 601 #undef FRAME_OFFSET
 602
 603
 604
 605
 606 ALIGNTEXT16
 607 GLOBL GLNAME( _mesa_x86_transform_points4_identity )
 608 GLNAME( _mesa_x86_transform_points4_identity ):
 609
 610 #define FRAME_OFFSET 12
 611         PUSH_L( ESI )
 612         PUSH_L( EDI )
 613         PUSH_L( EBX )
 614
 615         MOV_L( ARG_SOURCE, ESI )
 616         MOV_L( ARG_DEST, EDI )
 617
 618         MOV_L( ARG_MATRIX, EDX )
 619         MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
 620
 621         TEST_L( ECX, ECX )
 622         JZ( LLBL(x86_p4_ir_done) )
 623
 624         MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
 625         OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
 626
 627         MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
 628         MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
 629
 630         SHL_L( CONST(4), ECX )
 631         MOV_L( REGOFF(V4F_START, ESI), ESI )
 632
 633         MOV_L( REGOFF(V4F_START, EDI), EDI )
 634         ADD_L( EDI, ECX )
 635
 636         CMP_L( ESI, EDI )
 637         JE( LLBL(x86_p4_ir_done) )
 638
 639 ALIGNTEXT16
 640 LLBL(x86_p4_ir_loop):
 641
 642         MOV_L( SRC0, EBX )
 643         MOV_L( SRC1, EDX )
 644
 645         MOV_L( EBX, DST0 )
 646         MOV_L( EDX, DST1 )
 647
 648         MOV_L( SRC2, EBX )
 649         MOV_L( SRC3, EDX )
 650
 651         MOV_L( EBX, DST2 )
 652         MOV_L( EDX, DST3 )
 653
 654 LLBL(x86_p4_ir_skip):
 655
 656         ADD_L( CONST(16), EDI )
 657         ADD_L( EAX, ESI )
 658         CMP_L( ECX, EDI )
 659         JNE( LLBL(x86_p4_ir_loop) )
 660
 661 LLBL(x86_p4_ir_done):
 662
 663         POP_L( EBX )
 664         POP_L( EDI )
 665         POP_L( ESI )
 666         RET