src/mesa/drivers/dri/r200/r200_vtxtmp_x86.S

   1 /* $XFree86: xc/lib/GL/mesa/src/drv/r200/r200_vtxtmp_x86.S,v 1.2 2002/11/07 18:31:59 tsi Exp $ */
   2 /**************************************************************************
   3
   4 Copyright 2002 Tungsten Graphics Inc., Cedar Park, Texas.
   5
   6 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 on the rights to use, copy, modify, merge, publish, distribute, sub
  12 license, and/or sell copies of the Software, and to permit persons to whom
  13 the Software is furnished to do so, subject to the following conditions:
  14
  15 The above copyright notice and this permission notice (including the next
  16 paragraph) shall be included in all copies or substantial portions of the
  17 Software.
  18
  19 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20 IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21 FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  22 ATI, TUNGSTEN GRAPHICS AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  23 DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  24 OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  25 USE OR OTHER DEALINGS IN THE SOFTWARE.
  26
  27 **************************************************************************/
  28
  29 #define GLOBL( x )      \
  30 .globl x;               \
  31 x:
  32
  33 .data
  34 .align 4
  35
  36 /*
  37         vertex 3f vertex size 4
  38 */
  39
  40 GLOBL ( _x86_Vertex3f_4 )
  41         movl    (0), %ecx
  42         movl    4(%esp), %eax
  43         movl    8(%esp), %edx
  44         movl    %eax, (%ecx)
  45         movl    %edx, 4(%ecx)
  46         movl    12(%esp), %eax
  47         movl    (0), %edx
  48         movl    %eax, 8(%ecx)
  49         movl    %edx, 12(%ecx)
  50         movl    (0), %eax
  51         addl    $16, %ecx
  52         dec     %eax
  53         movl    %ecx, (0)
  54         movl    %eax, (0)
  55         je      .1
  56         ret
  57 .1:     jmp     *0
  58
  59 GLOBL ( _x86_Vertex3f_4_end )
  60
  61 /*
  62         vertex 3f vertex size 6
  63 */
  64 GLOBL ( _x86_Vertex3f_6 )
  65         push    %edi
  66         movl    (0), %edi
  67         movl    8(%esp), %eax
  68         movl    12(%esp), %edx
  69         movl    16(%esp), %ecx
  70         movl    %eax, (%edi)
  71         movl    %edx, 4(%edi)
  72         movl    %ecx, 8(%edi)
  73         movl    (0), %eax
  74         movl    (0), %edx
  75         movl    (0), %ecx
  76         movl    %eax, 12(%edi)
  77         movl    %edx, 16(%edi)
  78         movl    %ecx, 20(%edi)
  79         addl    $24, %edi
  80         movl    (0), %eax
  81         movl    %edi, (0)
  82         dec     %eax
  83         pop     %edi
  84         movl    %eax, (0)
  85         je      .2
  86         ret
  87 .2:     jmp     *0
  88 GLOBL ( _x86_Vertex3f_6_end )
  89 /*
  90         vertex 3f generic size
  91 */
  92 GLOBL ( _x86_Vertex3f )
  93         push    %edi
  94         push    %esi
  95         movl    $0, %esi
  96         movl    (0), %edi
  97         movl    12(%esp), %eax
  98         movl    16(%esp), %edx
  99         movl    20(%esp), %ecx
 100         movl    %eax, (%edi)
 101         movl    %edx, 4(%edi)
 102         movl    %ecx, 8(%edi)
 103         addl    $12, %edi
 104         movl    $0, %ecx
 105         repz
 106         movsl %ds:(%esi), %es:(%edi)
 107         movl    (0), %eax
 108         movl    %edi, (0)
 109         dec     %eax
 110         movl    %eax, (0)
 111         pop     %esi
 112         pop     %edi
 113         je      .3
 114         ret
 115 .3:     jmp     *0
 116
 117 GLOBL ( _x86_Vertex3f_end )
 118
 119 /*
 120         Vertex 3fv vertex size 6
 121 */
 122 GLOBL ( _x86_Vertex3fv_6 )
 123         movl    (0), %eax
 124         movl    4(%esp), %ecx
 125         movl    (%ecx), %edx
 126         movl    %edx, (%eax)
 127         movl    4(%ecx), %edx
 128         movl    8(%ecx), %ecx
 129         movl    %edx, 4(%eax)
 130         movl    %ecx, 8(%eax)
 131         movl    (28), %edx
 132         movl    (32), %ecx
 133         movl    %edx, 12(%eax)
 134         movl    %ecx, 16(%eax)
 135         movl    (36), %edx
 136         movl    %edx, 20(%eax)
 137         addl    $24, %eax
 138         movl    %eax, 0
 139         movl    4, %eax
 140         dec     %eax
 141         movl    %eax, 4
 142         je      .4
 143         ret
 144 .4:     jmp    *8
 145
 146 GLOBL ( _x86_Vertex3fv_6_end )
 147
 148 /*
 149         Vertex 3fv vertex size 8
 150 */
 151 GLOBL ( _x86_Vertex3fv_8 )
 152         movl    (0), %eax
 153         movl    4(%esp), %ecx
 154         movl    (%ecx), %edx
 155         movl    %edx ,(%eax)
 156         movl    4(%ecx) ,%edx
 157         movl    8(%ecx) ,%ecx
 158         movl    %edx, 4(%eax)
 159         movl    %ecx, 8(%eax)
 160         movl    (28), %edx
 161         movl    (32), %ecx
 162         movl    %edx, 12(%eax)
 163         movl    %ecx, 16(%eax)
 164         movl    (28), %edx
 165         movl    (32), %ecx
 166         movl    %edx, 20(%eax)
 167         movl    %ecx, 24(%eax)
 168         movl    (36), %edx
 169         movl    %edx, 28(%eax)
 170         addl    $32, %eax
 171         movl    %eax, (0)
 172         movl    4, %eax
 173         dec     %eax
 174         movl    %eax, (4)
 175         je      .5
 176         ret
 177 .5:     jmp    *8
 178
 179 GLOBL ( _x86_Vertex3fv_8_end )
 180
 181 /*
 182         Vertex 3fv generic vertex size
 183 */
 184 GLOBL ( _x86_Vertex3fv )
 185         movl    4(%esp), %edx
 186         push    %edi
 187         push    %esi
 188         movl    (0x1010101), %edi
 189         movl    (%edx), %eax
 190         movl    4(%edx), %ecx
 191         movl    8(%edx), %esi
 192         movl    %eax, (%edi)
 193         movl    %ecx, 4(%edi)
 194         movl    %esi, 8(%edi)
 195         addl    $12, %edi
 196         movl    $6, %ecx
 197         movl    $0x58, %esi
 198         repz
 199         movsl %ds:(%esi), %es:(%edi)
 200         movl    %edi, (0x1010101)
 201         movl    (0x2020202), %eax
 202         pop     %esi
 203         pop     %edi
 204         dec     %eax
 205         movl    %eax, (0x2020202)
 206         je      .6
 207         ret
 208 .6:     jmp    *0
 209 GLOBL ( _x86_Vertex3fv_end )
 210
 211
 212 /**
 213  * Generic handler for 2 float format data.  This can be used for
 214  * TexCoord2f and possibly other functions.
 215  */
 216
 217 GLOBL ( _x86_Attribute2f )
 218         movl    $0x0, %edx
 219         movl    4(%esp), %eax
 220         movl    8(%esp), %ecx
 221         movl    %eax, (%edx)
 222         movl    %ecx, 4(%edx)
 223         ret
 224 GLOBL ( _x86_Attribute2f_end )
 225
 226
 227 /**
 228  * Generic handler for 2 float vector format data.  This can be used for
 229  * TexCoord2fv and possibly other functions.
 230  */
 231
 232 GLOBL( _x86_Attribute2fv)
 233         movl 4(%esp), %eax      /* load 'v' off stack */
 234         movl (%eax), %ecx       /* load v[0] */
 235         movl 4(%eax), %eax      /* load v[1] */
 236         movl %ecx, 0            /* store v[0] to current vertex */
 237         movl %eax, 4            /* store v[1] to current vertex */
 238         ret
 239 GLOBL ( _x86_Attribute2fv_end )
 240
 241
 242 /**
 243  * Generic handler for 3 float format data.  This can be used for
 244  * Normal3f, Color3f (when the color target is also float), or
 245  * TexCoord3f.
 246  */
 247
 248 GLOBL ( _x86_Attribute3f )
 249         movl    4(%esp), %ecx
 250         movl    8(%esp), %edx
 251         movl    12(%esp), %eax
 252         movl    %ecx, 0
 253         movl    %edx, 4
 254         movl    %eax, 8
 255         ret
 256 GLOBL ( _x86_Attribute3f_end )
 257
 258 /**
 259  * Generic handler for 3 float vector format data.  This can be used for
 260  * Normal3f, Color3f (when the color target is also float), or
 261  * TexCoord3f.
 262  */
 263
 264 GLOBL( _x86_Attribute3fv)
 265         movl 4(%esp), %eax      /* load 'v' off stack */
 266         movl (%eax), %ecx       /* load v[0] */
 267         movl 4(%eax), %edx      /* load v[1] */
 268         movl 8(%eax), %eax      /* load v[2] */
 269         movl %ecx, 0            /* store v[0] to current vertex */
 270         movl %edx, 4            /* store v[1] to current vertex */
 271         movl %eax, 8            /* store v[2] to current vertex */
 272         ret
 273 GLOBL ( _x86_Attribute3fv_end )
 274
 275
 276 /*
 277         Color 4ubv_ub
 278 */
 279 GLOBL ( _x86_Color4ubv_ub )
 280         movl 4(%esp), %eax
 281         movl $0x12345678, %edx
 282         movl (%eax), %eax
 283         movl %eax, (%edx)
 284         ret
 285 GLOBL ( _x86_Color4ubv_ub_end )
 286
 287 /*
 288         Color 4ubv 4f
 289 */
 290 GLOBL ( _x86_Color4ubv_4f )
 291         push    %ebx
 292         movl    $0, %edx
 293         xor     %eax, %eax
 294         xor     %ecx, %ecx
 295         movl    8(%esp), %ebx
 296         movl    (%ebx), %ebx
 297         mov     %bl, %al
 298         mov     %bh, %cl
 299         movl    (%edx,%eax,4),%eax
 300         movl    (%edx,%ecx,4),%ecx
 301         movl    %eax, (0xdeadbeaf)
 302         movl    %ecx, (0xdeadbeaf)
 303         xor     %eax, %eax
 304         xor     %ecx, %ecx
 305         shr     $16, %ebx
 306         mov     %bl, %al
 307         mov     %bh, %cl
 308         movl    (%edx,%eax,4), %eax
 309         movl    (%edx,%ecx,4), %ecx
 310         movl    %eax, (0xdeadbeaf)
 311         movl    %ecx, (0xdeadbeaf)
 312         pop     %ebx
 313         ret
 314 GLOBL ( _x86_Color4ubv_4f_end )
 315
 316 /*
 317
 318         Color4ub_ub
 319 */
 320 GLOBL( _x86_Color4ub_ub )
 321         push    %ebx
 322         movl    8(%esp), %eax
 323         movl    12(%esp), %edx
 324         movl    16(%esp), %ecx
 325         movl    20(%esp), %ebx
 326         mov     %al, (0)
 327         mov     %dl, (0)
 328         mov     %cl, (0)
 329         mov     %bl, (0)
 330         pop     %ebx
 331         ret
 332 GLOBL( _x86_Color4ub_ub_end )
 333
 334
 335 /* \todo: change the "and $7, %eax" to something like "target & 4 ? target & 5 : target & 3)" */
 336 /*
 337         MultiTexCoord2fv st0/st1
 338 */
 339 GLOBL( _x86_MultiTexCoord2fv )
 340         movl    4(%esp), %eax
 341         movl    8(%esp), %ecx
 342         and     $7, %eax
 343         movl    (%ecx), %edx
 344         shl     $3, %eax
 345         movl    4(%ecx), %ecx
 346         movl    %edx, 0xdeadbeef(%eax)
 347         movl    %ecx, 0xdeadbeef(%eax)
 348         ret
 349 GLOBL( _x86_MultiTexCoord2fv_end )
 350
 351 /*
 352         MultiTexCoord2fv
 353 */
 354
 355 GLOBL( _x86_MultiTexCoord2fv_2 )
 356         movl    4(%esp,1), %eax
 357         movl    8(%esp,1), %ecx
 358         and     $0x7, %eax
 359         movl    0(,%eax,4), %edx
 360         movl    (%ecx), %eax
 361         movl    %eax, (%edx)
 362         movl    4(%ecx), %eax
 363         movl    %eax, 4(%edx)
 364         ret
 365 GLOBL( _x86_MultiTexCoord2fv_2_end )
 366
 367 /*
 368         MultiTexCoord2f st0/st1
 369 */
 370 GLOBL( _x86_MultiTexCoord2f )
 371         movl    4(%esp), %eax
 372         movl    8(%esp), %edx
 373         movl    12(%esp), %ecx
 374         and     $7, %eax
 375         shl     $3, %eax
 376         movl    %edx, 0xdeadbeef(%eax)
 377         movl    %ecx, 0xdeadbeef(%eax)
 378         ret
 379 GLOBL( _x86_MultiTexCoord2f_end )
 380
 381 /*
 382         MultiTexCoord2f
 383 */
 384 GLOBL( _x86_MultiTexCoord2f_2 )
 385         movl    4(%esp), %eax
 386         movl    8(%esp), %edx
 387         movl    12(%esp,1), %ecx
 388         and     $7,%eax
 389         movl    0(,%eax,4), %eax
 390         movl    %edx, (%eax)
 391         movl    %ecx, 4(%eax)
 392         ret
 393 GLOBL( _x86_MultiTexCoord2f_2_end )
 394
 395 #if defined(USE_SSE_ASM)
 396 /**
 397  * This can be used as a template for either Color3fv (when the color
 398  * target is also a 3f) or Normal3fv.
 399  */
 400
 401 GLOBL( _sse_Attribute3fv )
 402         movl    4(%esp), %eax
 403         movlps  (%eax), %xmm0
 404         movl    8(%eax), %eax
 405         movlps  %xmm0, 0
 406         movl    %eax, 8
 407         ret
 408 GLOBL( _sse_Attribute3fv_end )
 409
 410 /**
 411  * This can be used as a template for either Color3f (when the color
 412  * target is also a 3f) or Normal3f.
 413  */
 414
 415 GLOBL( _sse_Attribute3f )
 416         movlps  4(%esp), %xmm0
 417         movl    12(%esp), %eax
 418         movlps  %xmm0, 0
 419         movl    %eax, 8
 420         ret
 421 GLOBL( _sse_Attribute3f_end )
 422
 423
 424 /**
 425  * Generic handler for 2 float vector format data.  This can be used for
 426  * TexCoord2fv and possibly other functions.
 427  */
 428
 429 GLOBL( _sse_Attribute2fv )
 430         movl    4(%esp), %eax
 431         movlps  (%eax), %xmm0
 432         movlps  %xmm0, 0
 433         ret
 434 GLOBL( _sse_Attribute2fv_end )
 435
 436
 437 /**
 438  * Generic handler for 2 float format data.  This can be used for
 439  * TexCoord2f and possibly other functions.
 440  */
 441
 442 GLOBL( _sse_Attribute2f )
 443         movlps  4(%esp), %xmm0
 444         movlps  %xmm0, 0
 445         ret
 446 GLOBL( _sse_Attribute2f_end )
 447
 448 /*
 449         MultiTexCoord2fv st0/st1
 450 */
 451 GLOBL( _sse_MultiTexCoord2fv )
 452         movl    4(%esp), %eax
 453         movl    8(%esp), %ecx
 454         and     $7, %eax
 455         movlps  (%ecx), %xmm0
 456         movlps  %xmm0, 0xdeadbeef(,%eax,8)
 457         ret
 458 GLOBL( _sse_MultiTexCoord2fv_end )
 459
 460 /*
 461         MultiTexCoord2fv
 462 */
 463 GLOBL( _sse_MultiTexCoord2fv_2 )
 464         movl    4(%esp), %eax
 465         movl    8(%esp), %ecx
 466         and     $0x7, %eax
 467         movl    0(,%eax,4), %edx
 468         movlps  (%ecx), %xmm0
 469         movlps  %xmm0, (%edx)
 470         ret
 471 GLOBL( _sse_MultiTexCoord2fv_2_end )
 472
 473 /*
 474         MultiTexCoord2f st0/st1
 475 */
 476 GLOBL( _sse_MultiTexCoord2f )
 477         movl    4(%esp), %eax
 478         and     $7, %eax
 479         movlps  8(%esp), %xmm0
 480         movlps  %xmm0, 0xdeadbeef(,%eax,8)
 481         ret
 482 GLOBL( _sse_MultiTexCoord2f_end )
 483
 484 /*
 485         MultiTexCoord2f
 486 */
 487 GLOBL( _sse_MultiTexCoord2f_2 )
 488         movl    4(%esp), %eax
 489         movlps  8(%esp), %xmm0
 490         and     $7,%eax
 491         movl    0(,%eax,4), %eax
 492         movlps  %xmm0, (%eax)
 493         ret
 494 GLOBL( _sse_MultiTexCoord2f_2_end )
 495 #endif