src/gallium/auxiliary/tgsi/tgsi_ureg.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * 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
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE, INC AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28
  29 #include "pipe/p_context.h"
  30 #include "pipe/p_state.h"
  31 #include "tgsi/tgsi_ureg.h"
  32 #include "tgsi/tgsi_build.h"
  33 #include "tgsi/tgsi_info.h"
  34 #include "tgsi/tgsi_dump.h"
  35 #include "tgsi/tgsi_sanity.h"
  36 #include "util/u_debug.h"
  37 #include "util/u_memory.h"
  38 #include "util/u_math.h"
  39
  40 union tgsi_any_token {
  41    struct tgsi_header header;
  42    struct tgsi_processor processor;
  43    struct tgsi_token token;
  44    struct tgsi_property prop;
  45    struct tgsi_property_data prop_data;
  46    struct tgsi_declaration decl;
  47    struct tgsi_declaration_range decl_range;
  48    struct tgsi_declaration_dimension decl_dim;
  49    struct tgsi_declaration_semantic decl_semantic;
  50    struct tgsi_immediate imm;
  51    union  tgsi_immediate_data imm_data;
  52    struct tgsi_instruction insn;
  53    struct tgsi_instruction_predicate insn_predicate;
  54    struct tgsi_instruction_label insn_label;
  55    struct tgsi_instruction_texture insn_texture;
  56    struct tgsi_src_register src;
  57    struct tgsi_dimension dim;
  58    struct tgsi_dst_register dst;
  59    unsigned value;
  60 };
  61
  62
  63 struct ureg_tokens {
  64    union tgsi_any_token *tokens;
  65    unsigned size;
  66    unsigned order;
  67    unsigned count;
  68 };
  69
  70 #define UREG_MAX_INPUT PIPE_MAX_ATTRIBS
  71 #define UREG_MAX_SYSTEM_VALUE PIPE_MAX_ATTRIBS
  72 #define UREG_MAX_OUTPUT PIPE_MAX_ATTRIBS
  73 #define UREG_MAX_CONSTANT_RANGE 32
  74 #define UREG_MAX_IMMEDIATE 32
  75 #define UREG_MAX_TEMP 256
  76 #define UREG_MAX_ADDR 2
  77 #define UREG_MAX_PRED 1
  78
  79 struct const_decl {
  80    struct {
  81       unsigned first;
  82       unsigned last;
  83    } constant_range[UREG_MAX_CONSTANT_RANGE];
  84    unsigned nr_constant_ranges;
  85 };
  86
  87 #define DOMAIN_DECL 0
  88 #define DOMAIN_INSN 1
  89
  90 struct ureg_program
  91 {
  92    unsigned processor;
  93    struct pipe_context *pipe;
  94
  95    struct {
  96       unsigned semantic_name;
  97       unsigned semantic_index;
  98       unsigned interp;
  99       unsigned cylindrical_wrap;
 100    } fs_input[UREG_MAX_INPUT];
 101    unsigned nr_fs_inputs;
 102
 103    unsigned vs_inputs[UREG_MAX_INPUT/32];
 104
 105    struct {
 106       unsigned index;
 107       unsigned semantic_name;
 108       unsigned semantic_index;
 109    } gs_input[UREG_MAX_INPUT];
 110    unsigned nr_gs_inputs;
 111
 112    struct {
 113       unsigned index;
 114       unsigned semantic_name;
 115       unsigned semantic_index;
 116    } system_value[UREG_MAX_SYSTEM_VALUE];
 117    unsigned nr_system_values;
 118
 119    struct {
 120       unsigned semantic_name;
 121       unsigned semantic_index;
 122    } output[UREG_MAX_OUTPUT];
 123    unsigned nr_outputs;
 124
 125    struct {
 126       union {
 127          float f[4];
 128          unsigned u[4];
 129          int i[4];
 130       } value;
 131       unsigned nr;
 132       unsigned type;
 133    } immediate[UREG_MAX_IMMEDIATE];
 134    unsigned nr_immediates;
 135
 136    struct ureg_src sampler[PIPE_MAX_SAMPLERS];
 137    unsigned nr_samplers;
 138
 139    unsigned temps_active[UREG_MAX_TEMP / 32];
 140    unsigned nr_temps;
 141
 142    struct const_decl const_decls;
 143    struct const_decl const_decls2D[PIPE_MAX_CONSTANT_BUFFERS];
 144
 145    unsigned property_gs_input_prim;
 146    unsigned property_gs_output_prim;
 147    unsigned property_gs_max_vertices;
 148    unsigned char property_fs_coord_origin; /* = TGSI_FS_COORD_ORIGIN_* */
 149    unsigned char property_fs_coord_pixel_center; /* = TGSI_FS_COORD_PIXEL_CENTER_* */
 150
 151    unsigned nr_addrs;
 152    unsigned nr_preds;
 153    unsigned nr_instructions;
 154
 155    struct ureg_tokens domain[2];
 156 };
 157
 158 static union tgsi_any_token error_tokens[32];
 159
 160 static void tokens_error( struct ureg_tokens *tokens )
 161 {
 162    if (tokens->tokens && tokens->tokens != error_tokens)
 163       FREE(tokens->tokens);
 164
 165    tokens->tokens = error_tokens;
 166    tokens->size = Elements(error_tokens);
 167    tokens->count = 0;
 168 }
 169
 170
 171 static void tokens_expand( struct ureg_tokens *tokens,
 172                            unsigned count )
 173 {
 174    unsigned old_size = tokens->size * sizeof(unsigned);
 175
 176    if (tokens->tokens == error_tokens) {
 177       return;
 178    }
 179
 180    while (tokens->count + count > tokens->size) {
 181       tokens->size = (1 << ++tokens->order);
 182    }
 183
 184    tokens->tokens = REALLOC(tokens->tokens,
 185                             old_size,
 186                             tokens->size * sizeof(unsigned));
 187    if (tokens->tokens == NULL) {
 188       tokens_error(tokens);
 189    }
 190 }
 191
 192 static void set_bad( struct ureg_program *ureg )
 193 {
 194    tokens_error(&ureg->domain[0]);
 195 }
 196
 197
 198
 199 static union tgsi_any_token *get_tokens( struct ureg_program *ureg,
 200                                          unsigned domain,
 201                                          unsigned count )
 202 {
 203    struct ureg_tokens *tokens = &ureg->domain[domain];
 204    union tgsi_any_token *result;
 205
 206    if (tokens->count + count > tokens->size)
 207       tokens_expand(tokens, count);
 208
 209    result = &tokens->tokens[tokens->count];
 210    tokens->count += count;
 211    return result;
 212 }
 213
 214
 215 static union tgsi_any_token *retrieve_token( struct ureg_program *ureg,
 216                                             unsigned domain,
 217                                             unsigned nr )
 218 {
 219    if (ureg->domain[domain].tokens == error_tokens)
 220       return &error_tokens[0];
 221
 222    return &ureg->domain[domain].tokens[nr];
 223 }
 224
 225
 226
 227 static INLINE struct ureg_dst
 228 ureg_dst_register( unsigned file,
 229                    unsigned index )
 230 {
 231    struct ureg_dst dst;
 232
 233    dst.File      = file;
 234    dst.WriteMask = TGSI_WRITEMASK_XYZW;
 235    dst.Indirect  = 0;
 236    dst.IndirectIndex = 0;
 237    dst.IndirectSwizzle = 0;
 238    dst.Saturate  = 0;
 239    dst.Predicate = 0;
 240    dst.PredNegate = 0;
 241    dst.PredSwizzleX = TGSI_SWIZZLE_X;
 242    dst.PredSwizzleY = TGSI_SWIZZLE_Y;
 243    dst.PredSwizzleZ = TGSI_SWIZZLE_Z;
 244    dst.PredSwizzleW = TGSI_SWIZZLE_W;
 245    dst.Index     = index;
 246
 247    return dst;
 248 }
 249
 250
 251 void
 252 ureg_property_gs_input_prim(struct ureg_program *ureg,
 253                             unsigned input_prim)
 254 {
 255    ureg->property_gs_input_prim = input_prim;
 256 }
 257
 258 void
 259 ureg_property_gs_output_prim(struct ureg_program *ureg,
 260                              unsigned output_prim)
 261 {
 262    ureg->property_gs_output_prim = output_prim;
 263 }
 264
 265 void
 266 ureg_property_gs_max_vertices(struct ureg_program *ureg,
 267                               unsigned max_vertices)
 268 {
 269    ureg->property_gs_max_vertices = max_vertices;
 270 }
 271
 272 void
 273 ureg_property_fs_coord_origin(struct ureg_program *ureg,
 274                             unsigned fs_coord_origin)
 275 {
 276    ureg->property_fs_coord_origin = fs_coord_origin;
 277 }
 278
 279 void
 280 ureg_property_fs_coord_pixel_center(struct ureg_program *ureg,
 281                             unsigned fs_coord_pixel_center)
 282 {
 283    ureg->property_fs_coord_pixel_center = fs_coord_pixel_center;
 284 }
 285
 286
 287
 288 struct ureg_src
 289 ureg_DECL_fs_input_cyl(struct ureg_program *ureg,
 290                        unsigned semantic_name,
 291                        unsigned semantic_index,
 292                        unsigned interp_mode,
 293                        unsigned cylindrical_wrap)
 294 {
 295    unsigned i;
 296
 297    for (i = 0; i < ureg->nr_fs_inputs; i++) {
 298       if (ureg->fs_input[i].semantic_name == semantic_name &&
 299           ureg->fs_input[i].semantic_index == semantic_index) {
 300          goto out;
 301       }
 302    }
 303
 304    if (ureg->nr_fs_inputs < UREG_MAX_INPUT) {
 305       ureg->fs_input[i].semantic_name = semantic_name;
 306       ureg->fs_input[i].semantic_index = semantic_index;
 307       ureg->fs_input[i].interp = interp_mode;
 308       ureg->fs_input[i].cylindrical_wrap = cylindrical_wrap;
 309       ureg->nr_fs_inputs++;
 310    } else {
 311       set_bad(ureg);
 312    }
 313
 314 out:
 315    return ureg_src_register(TGSI_FILE_INPUT, i);
 316 }
 317
 318
 319 struct ureg_src
 320 ureg_DECL_vs_input( struct ureg_program *ureg,
 321                     unsigned index )
 322 {
 323    assert(ureg->processor == TGSI_PROCESSOR_VERTEX);
 324
 325    ureg->vs_inputs[index/32] |= 1 << (index % 32);
 326    return ureg_src_register( TGSI_FILE_INPUT, index );
 327 }
 328
 329
 330 struct ureg_src
 331 ureg_DECL_gs_input(struct ureg_program *ureg,
 332                    unsigned index,
 333                    unsigned semantic_name,
 334                    unsigned semantic_index)
 335 {
 336    if (ureg->nr_gs_inputs < UREG_MAX_INPUT) {
 337       ureg->gs_input[ureg->nr_gs_inputs].index = index;
 338       ureg->gs_input[ureg->nr_gs_inputs].semantic_name = semantic_name;
 339       ureg->gs_input[ureg->nr_gs_inputs].semantic_index = semantic_index;
 340       ureg->nr_gs_inputs++;
 341    } else {
 342       set_bad(ureg);
 343    }
 344
 345    /* XXX: Add suport for true 2D input registers. */
 346    return ureg_src_register(TGSI_FILE_INPUT, index);
 347 }
 348
 349
 350 struct ureg_src
 351 ureg_DECL_system_value(struct ureg_program *ureg,
 352                        unsigned index,
 353                        unsigned semantic_name,
 354                        unsigned semantic_index)
 355 {
 356    if (ureg->nr_system_values < UREG_MAX_SYSTEM_VALUE) {
 357       ureg->system_value[ureg->nr_system_values].index = index;
 358       ureg->system_value[ureg->nr_system_values].semantic_name = semantic_name;
 359       ureg->system_value[ureg->nr_system_values].semantic_index = semantic_index;
 360       ureg->nr_system_values++;
 361    } else {
 362       set_bad(ureg);
 363    }
 364
 365    return ureg_src_register(TGSI_FILE_SYSTEM_VALUE, index);
 366 }
 367
 368
 369 struct ureg_dst
 370 ureg_DECL_output( struct ureg_program *ureg,
 371                   unsigned name,
 372                   unsigned index )
 373 {
 374    unsigned i;
 375
 376    for (i = 0; i < ureg->nr_outputs; i++) {
 377       if (ureg->output[i].semantic_name == name &&
 378           ureg->output[i].semantic_index == index)
 379          goto out;
 380    }
 381
 382    if (ureg->nr_outputs < UREG_MAX_OUTPUT) {
 383       ureg->output[i].semantic_name = name;
 384       ureg->output[i].semantic_index = index;
 385       ureg->nr_outputs++;
 386    }
 387    else {
 388       set_bad( ureg );
 389    }
 390
 391 out:
 392    return ureg_dst_register( TGSI_FILE_OUTPUT, i );
 393 }
 394
 395
 396 /* Returns a new constant register.  Keep track of which have been
 397  * referred to so that we can emit decls later.
 398  *
 399  * Constant operands declared with this function must be addressed
 400  * with a two-dimensional index.
 401  *
 402  * There is nothing in this code to bind this constant to any tracked
 403  * value or manage any constant_buffer contents -- that's the
 404  * resposibility of the calling code.
 405  */
 406 void
 407 ureg_DECL_constant2D(struct ureg_program *ureg,
 408                      unsigned first,
 409                      unsigned last,
 410                      unsigned index2D)
 411 {
 412    struct const_decl *decl = &ureg->const_decls2D[index2D];
 413
 414    assert(index2D < PIPE_MAX_CONSTANT_BUFFERS);
 415
 416    if (decl->nr_constant_ranges < UREG_MAX_CONSTANT_RANGE) {
 417       uint i = decl->nr_constant_ranges++;
 418
 419       decl->constant_range[i].first = first;
 420       decl->constant_range[i].last = last;
 421    }
 422 }
 423
 424
 425 /* A one-dimensional, depricated version of ureg_DECL_constant2D().
 426  *
 427  * Constant operands declared with this function must be addressed
 428  * with a one-dimensional index.
 429  */
 430 struct ureg_src
 431 ureg_DECL_constant(struct ureg_program *ureg,
 432                    unsigned index)
 433 {
 434    struct const_decl *decl = &ureg->const_decls;
 435    unsigned minconst = index, maxconst = index;
 436    unsigned i;
 437
 438    /* Inside existing range?
 439     */
 440    for (i = 0; i < decl->nr_constant_ranges; i++) {
 441       if (decl->constant_range[i].first <= index &&
 442           decl->constant_range[i].last >= index) {
 443          goto out;
 444       }
 445    }
 446
 447    /* Extend existing range?
 448     */
 449    for (i = 0; i < decl->nr_constant_ranges; i++) {
 450       if (decl->constant_range[i].last == index - 1) {
 451          decl->constant_range[i].last = index;
 452          goto out;
 453       }
 454
 455       if (decl->constant_range[i].first == index + 1) {
 456          decl->constant_range[i].first = index;
 457          goto out;
 458       }
 459
 460       minconst = MIN2(minconst, decl->constant_range[i].first);
 461       maxconst = MAX2(maxconst, decl->constant_range[i].last);
 462    }
 463
 464    /* Create new range?
 465     */
 466    if (decl->nr_constant_ranges < UREG_MAX_CONSTANT_RANGE) {
 467       i = decl->nr_constant_ranges++;
 468       decl->constant_range[i].first = index;
 469       decl->constant_range[i].last = index;
 470       goto out;
 471    }
 472
 473    /* Collapse all ranges down to one:
 474     */
 475    i = 0;
 476    decl->constant_range[0].first = minconst;
 477    decl->constant_range[0].last = maxconst;
 478    decl->nr_constant_ranges = 1;
 479
 480 out:
 481    assert(i < decl->nr_constant_ranges);
 482    assert(decl->constant_range[i].first <= index);
 483    assert(decl->constant_range[i].last >= index);
 484    return ureg_src_register(TGSI_FILE_CONSTANT, index);
 485 }
 486
 487
 488 /* Allocate a new temporary.  Temporaries greater than UREG_MAX_TEMP
 489  * are legal, but will not be released.
 490  */
 491 struct ureg_dst ureg_DECL_temporary( struct ureg_program *ureg )
 492 {
 493    unsigned i;
 494
 495    for (i = 0; i < UREG_MAX_TEMP; i += 32) {
 496       int bit = ffs(~ureg->temps_active[i/32]);
 497       if (bit != 0) {
 498          i += bit - 1;
 499          goto out;
 500       }
 501    }
 502
 503    /* No reusable temps, so allocate a new one:
 504     */
 505    i = ureg->nr_temps++;
 506
 507 out:
 508    if (i < UREG_MAX_TEMP)
 509       ureg->temps_active[i/32] |= 1 << (i % 32);
 510
 511    if (i >= ureg->nr_temps)
 512       ureg->nr_temps = i + 1;
 513
 514    return ureg_dst_register( TGSI_FILE_TEMPORARY, i );
 515 }
 516
 517
 518 void ureg_release_temporary( struct ureg_program *ureg,
 519                              struct ureg_dst tmp )
 520 {
 521    if(tmp.File == TGSI_FILE_TEMPORARY)
 522       if (tmp.Index < UREG_MAX_TEMP)
 523          ureg->temps_active[tmp.Index/32] &= ~(1 << (tmp.Index % 32));
 524 }
 525
 526
 527 /* Allocate a new address register.
 528  */
 529 struct ureg_dst ureg_DECL_address( struct ureg_program *ureg )
 530 {
 531    if (ureg->nr_addrs < UREG_MAX_ADDR)
 532       return ureg_dst_register( TGSI_FILE_ADDRESS, ureg->nr_addrs++ );
 533
 534    assert( 0 );
 535    return ureg_dst_register( TGSI_FILE_ADDRESS, 0 );
 536 }
 537
 538 /* Allocate a new predicate register.
 539  */
 540 struct ureg_dst
 541 ureg_DECL_predicate(struct ureg_program *ureg)
 542 {
 543    if (ureg->nr_preds < UREG_MAX_PRED) {
 544       return ureg_dst_register(TGSI_FILE_PREDICATE, ureg->nr_preds++);
 545    }
 546
 547    assert(0);
 548    return ureg_dst_register(TGSI_FILE_PREDICATE, 0);
 549 }
 550
 551 /* Allocate a new sampler.
 552  */
 553 struct ureg_src ureg_DECL_sampler( struct ureg_program *ureg,
 554                                    unsigned nr )
 555 {
 556    unsigned i;
 557
 558    for (i = 0; i < ureg->nr_samplers; i++)
 559       if (ureg->sampler[i].Index == nr)
 560          return ureg->sampler[i];
 561
 562    if (i < PIPE_MAX_SAMPLERS) {
 563       ureg->sampler[i] = ureg_src_register( TGSI_FILE_SAMPLER, nr );
 564       ureg->nr_samplers++;
 565       return ureg->sampler[i];
 566    }
 567
 568    assert( 0 );
 569    return ureg->sampler[0];
 570 }
 571
 572
 573 static int
 574 match_or_expand_immediate( const unsigned *v,
 575                            unsigned nr,
 576                            unsigned *v2,
 577                            unsigned *pnr2,
 578                            unsigned *swizzle )
 579 {
 580    unsigned nr2 = *pnr2;
 581    unsigned i, j;
 582
 583    *swizzle = 0;
 584
 585    for (i = 0; i < nr; i++) {
 586       boolean found = FALSE;
 587
 588       for (j = 0; j < nr2 && !found; j++) {
 589          if (v[i] == v2[j]) {
 590             *swizzle |= j << (i * 2);
 591             found = TRUE;
 592          }
 593       }
 594
 595       if (!found) {
 596          if (nr2 >= 4) {
 597             return FALSE;
 598          }
 599
 600          v2[nr2] = v[i];
 601          *swizzle |= nr2 << (i * 2);
 602          nr2++;
 603       }
 604    }
 605
 606    /* Actually expand immediate only when fully succeeded.
 607     */
 608    *pnr2 = nr2;
 609    return TRUE;
 610 }
 611
 612
 613 static struct ureg_src
 614 decl_immediate( struct ureg_program *ureg,
 615                 const unsigned *v,
 616                 unsigned nr,
 617                 unsigned type )
 618 {
 619    unsigned i, j;
 620    unsigned swizzle = 0;
 621
 622    /* Could do a first pass where we examine all existing immediates
 623     * without expanding.
 624     */
 625
 626    for (i = 0; i < ureg->nr_immediates; i++) {
 627       if (ureg->immediate[i].type != type) {
 628          continue;
 629       }
 630       if (match_or_expand_immediate(v,
 631                                     nr,
 632                                     ureg->immediate[i].value.u,
 633                                     &ureg->immediate[i].nr,
 634                                     &swizzle)) {
 635          goto out;
 636       }
 637    }
 638
 639    if (ureg->nr_immediates < UREG_MAX_IMMEDIATE) {
 640       i = ureg->nr_immediates++;
 641       ureg->immediate[i].type = type;
 642       if (match_or_expand_immediate(v,
 643                                     nr,
 644                                     ureg->immediate[i].value.u,
 645                                     &ureg->immediate[i].nr,
 646                                     &swizzle)) {
 647          goto out;
 648       }
 649    }
 650
 651    set_bad(ureg);
 652
 653 out:
 654    /* Make sure that all referenced elements are from this immediate.
 655     * Has the effect of making size-one immediates into scalars.
 656     */
 657    for (j = nr; j < 4; j++) {
 658       swizzle |= (swizzle & 0x3) << (j * 2);
 659    }
 660
 661    return ureg_swizzle(ureg_src_register(TGSI_FILE_IMMEDIATE, i),
 662                        (swizzle >> 0) & 0x3,
 663                        (swizzle >> 2) & 0x3,
 664                        (swizzle >> 4) & 0x3,
 665                        (swizzle >> 6) & 0x3);
 666 }
 667
 668
 669 struct ureg_src
 670 ureg_DECL_immediate( struct ureg_program *ureg,
 671                      const float *v,
 672                      unsigned nr )
 673 {
 674    union {
 675       float f[4];
 676       unsigned u[4];
 677    } fu;
 678    unsigned int i;
 679
 680    for (i = 0; i < nr; i++) {
 681       fu.f[i] = v[i];
 682    }
 683
 684    return decl_immediate(ureg, fu.u, nr, TGSI_IMM_FLOAT32);
 685 }
 686
 687
 688 struct ureg_src
 689 ureg_DECL_immediate_uint( struct ureg_program *ureg,
 690                           const unsigned *v,
 691                           unsigned nr )
 692 {
 693    return decl_immediate(ureg, v, nr, TGSI_IMM_UINT32);
 694 }
 695
 696
 697 struct ureg_src
 698 ureg_DECL_immediate_block_uint( struct ureg_program *ureg,
 699                                 const unsigned *v,
 700                                 unsigned nr )
 701 {
 702    uint index;
 703    uint i;
 704
 705    if (ureg->nr_immediates + (nr + 3) / 4 > UREG_MAX_IMMEDIATE) {
 706       set_bad(ureg);
 707       return ureg_src_register(TGSI_FILE_IMMEDIATE, 0);
 708    }
 709
 710    index = ureg->nr_immediates;
 711    ureg->nr_immediates += (nr + 3) / 4;
 712
 713    for (i = index; i < ureg->nr_immediates; i++) {
 714       ureg->immediate[i].type = TGSI_IMM_UINT32;
 715       ureg->immediate[i].nr = nr > 4 ? 4 : nr;
 716       memcpy(ureg->immediate[i].value.u,
 717              &v[(i - index) * 4],
 718              ureg->immediate[i].nr * sizeof(uint));
 719       nr -= 4;
 720    }
 721
 722    return ureg_src_register(TGSI_FILE_IMMEDIATE, index);
 723 }
 724
 725
 726 struct ureg_src
 727 ureg_DECL_immediate_int( struct ureg_program *ureg,
 728                          const int *v,
 729                          unsigned nr )
 730 {
 731    return decl_immediate(ureg, (const unsigned *)v, nr, TGSI_IMM_INT32);
 732 }
 733
 734
 735 void
 736 ureg_emit_src( struct ureg_program *ureg,
 737                struct ureg_src src )
 738 {
 739    unsigned size = 1 + (src.Indirect ? 1 : 0) + (src.Dimension ? 1 : 0);
 740
 741    union tgsi_any_token *out = get_tokens( ureg, DOMAIN_INSN, size );
 742    unsigned n = 0;
 743
 744    assert(src.File != TGSI_FILE_NULL);
 745    assert(src.File != TGSI_FILE_OUTPUT);
 746    assert(src.File < TGSI_FILE_COUNT);
 747
 748    out[n].value = 0;
 749    out[n].src.File = src.File;
 750    out[n].src.SwizzleX = src.SwizzleX;
 751    out[n].src.SwizzleY = src.SwizzleY;
 752    out[n].src.SwizzleZ = src.SwizzleZ;
 753    out[n].src.SwizzleW = src.SwizzleW;
 754    out[n].src.Index = src.Index;
 755    out[n].src.Negate = src.Negate;
 756    out[0].src.Absolute = src.Absolute;
 757    n++;
 758
 759    if (src.Indirect) {
 760       out[0].src.Indirect = 1;
 761       out[n].value = 0;
 762       out[n].src.File = src.IndirectFile;
 763       out[n].src.SwizzleX = src.IndirectSwizzle;
 764       out[n].src.SwizzleY = src.IndirectSwizzle;
 765       out[n].src.SwizzleZ = src.IndirectSwizzle;
 766       out[n].src.SwizzleW = src.IndirectSwizzle;
 767       out[n].src.Index = src.IndirectIndex;
 768       n++;
 769    }
 770
 771    if (src.Dimension) {
 772       out[0].src.Dimension = 1;
 773       out[n].dim.Indirect = 0;
 774       out[n].dim.Dimension = 0;
 775       out[n].dim.Padding = 0;
 776       out[n].dim.Index = src.DimensionIndex;
 777       n++;
 778    }
 779
 780    assert(n == size);
 781 }
 782
 783
 784 void
 785 ureg_emit_dst( struct ureg_program *ureg,
 786                struct ureg_dst dst )
 787 {
 788    unsigned size = (1 +
 789                     (dst.Indirect ? 1 : 0));
 790
 791    union tgsi_any_token *out = get_tokens( ureg, DOMAIN_INSN, size );
 792    unsigned n = 0;
 793
 794    assert(dst.File != TGSI_FILE_NULL);
 795    assert(dst.File != TGSI_FILE_CONSTANT);
 796    assert(dst.File != TGSI_FILE_INPUT);
 797    assert(dst.File != TGSI_FILE_SAMPLER);
 798    assert(dst.File != TGSI_FILE_IMMEDIATE);
 799    assert(dst.File < TGSI_FILE_COUNT);
 800
 801    out[n].value = 0;
 802    out[n].dst.File = dst.File;
 803    out[n].dst.WriteMask = dst.WriteMask;
 804    out[n].dst.Indirect = dst.Indirect;
 805    out[n].dst.Index = dst.Index;
 806    n++;
 807
 808    if (dst.Indirect) {
 809       out[n].value = 0;
 810       out[n].src.File = TGSI_FILE_ADDRESS;
 811       out[n].src.SwizzleX = dst.IndirectSwizzle;
 812       out[n].src.SwizzleY = dst.IndirectSwizzle;
 813       out[n].src.SwizzleZ = dst.IndirectSwizzle;
 814       out[n].src.SwizzleW = dst.IndirectSwizzle;
 815       out[n].src.Index = dst.IndirectIndex;
 816       n++;
 817    }
 818
 819    assert(n == size);
 820 }
 821
 822
 823 static void validate( unsigned opcode,
 824                       unsigned nr_dst,
 825                       unsigned nr_src )
 826 {
 827 #ifdef DEBUG
 828    const struct tgsi_opcode_info *info = tgsi_get_opcode_info( opcode );
 829    assert(info);
 830    if(info) {
 831       assert(nr_dst == info->num_dst);
 832       assert(nr_src == info->num_src);
 833    }
 834 #endif
 835 }
 836
 837 struct ureg_emit_insn_result
 838 ureg_emit_insn(struct ureg_program *ureg,
 839                unsigned opcode,
 840                boolean saturate,
 841                boolean predicate,
 842                boolean pred_negate,
 843                unsigned pred_swizzle_x,
 844                unsigned pred_swizzle_y,
 845                unsigned pred_swizzle_z,
 846                unsigned pred_swizzle_w,
 847                unsigned num_dst,
 848                unsigned num_src )
 849 {
 850    union tgsi_any_token *out;
 851    uint count = predicate ? 2 : 1;
 852    struct ureg_emit_insn_result result;
 853
 854    validate( opcode, num_dst, num_src );
 855
 856    out = get_tokens( ureg, DOMAIN_INSN, count );
 857    out[0].insn = tgsi_default_instruction();
 858    out[0].insn.Opcode = opcode;
 859    out[0].insn.Saturate = saturate;
 860    out[0].insn.NumDstRegs = num_dst;
 861    out[0].insn.NumSrcRegs = num_src;
 862
 863    result.insn_token = ureg->domain[DOMAIN_INSN].count - count;
 864    result.extended_token = result.insn_token;
 865
 866    if (predicate) {
 867       out[0].insn.Predicate = 1;
 868       out[1].insn_predicate = tgsi_default_instruction_predicate();
 869       out[1].insn_predicate.Negate = pred_negate;
 870       out[1].insn_predicate.SwizzleX = pred_swizzle_x;
 871       out[1].insn_predicate.SwizzleY = pred_swizzle_y;
 872       out[1].insn_predicate.SwizzleZ = pred_swizzle_z;
 873       out[1].insn_predicate.SwizzleW = pred_swizzle_w;
 874    }
 875
 876    ureg->nr_instructions++;
 877
 878    return result;
 879 }
 880
 881
 882 void
 883 ureg_emit_label(struct ureg_program *ureg,
 884                 unsigned extended_token,
 885                 unsigned *label_token )
 886 {
 887    union tgsi_any_token *out, *insn;
 888
 889    if(!label_token)
 890       return;
 891
 892    out = get_tokens( ureg, DOMAIN_INSN, 1 );
 893    out[0].value = 0;
 894
 895    insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
 896    insn->insn.Label = 1;
 897
 898    *label_token = ureg->domain[DOMAIN_INSN].count - 1;
 899 }
 900
 901 /* Will return a number which can be used in a label to point to the
 902  * next instruction to be emitted.
 903  */
 904 unsigned
 905 ureg_get_instruction_number( struct ureg_program *ureg )
 906 {
 907    return ureg->nr_instructions;
 908 }
 909
 910 /* Patch a given label (expressed as a token number) to point to a
 911  * given instruction (expressed as an instruction number).
 912  */
 913 void
 914 ureg_fixup_label(struct ureg_program *ureg,
 915                  unsigned label_token,
 916                  unsigned instruction_number )
 917 {
 918    union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_INSN, label_token );
 919
 920    out->insn_label.Label = instruction_number;
 921 }
 922
 923
 924 void
 925 ureg_emit_texture(struct ureg_program *ureg,
 926                   unsigned extended_token,
 927                   unsigned target )
 928 {
 929    union tgsi_any_token *out, *insn;
 930
 931    out = get_tokens( ureg, DOMAIN_INSN, 1 );
 932    insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
 933
 934    insn->insn.Texture = 1;
 935
 936    out[0].value = 0;
 937    out[0].insn_texture.Texture = target;
 938 }
 939
 940
 941 void
 942 ureg_fixup_insn_size(struct ureg_program *ureg,
 943                      unsigned insn )
 944 {
 945    union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_INSN, insn );
 946
 947    assert(out->insn.Type == TGSI_TOKEN_TYPE_INSTRUCTION);
 948    out->insn.NrTokens = ureg->domain[DOMAIN_INSN].count - insn - 1;
 949 }
 950
 951
 952 void
 953 ureg_insn(struct ureg_program *ureg,
 954           unsigned opcode,
 955           const struct ureg_dst *dst,
 956           unsigned nr_dst,
 957           const struct ureg_src *src,
 958           unsigned nr_src )
 959 {
 960    struct ureg_emit_insn_result insn;
 961    unsigned i;
 962    boolean saturate;
 963    boolean predicate;
 964    boolean negate = FALSE;
 965    unsigned swizzle[4] = { 0 };
 966
 967    saturate = nr_dst ? dst[0].Saturate : FALSE;
 968    predicate = nr_dst ? dst[0].Predicate : FALSE;
 969    if (predicate) {
 970       negate = dst[0].PredNegate;
 971       swizzle[0] = dst[0].PredSwizzleX;
 972       swizzle[1] = dst[0].PredSwizzleY;
 973       swizzle[2] = dst[0].PredSwizzleZ;
 974       swizzle[3] = dst[0].PredSwizzleW;
 975    }
 976
 977    insn = ureg_emit_insn(ureg,
 978                          opcode,
 979                          saturate,
 980                          predicate,
 981                          negate,
 982                          swizzle[0],
 983                          swizzle[1],
 984                          swizzle[2],
 985                          swizzle[3],
 986                          nr_dst,
 987                          nr_src);
 988
 989    for (i = 0; i < nr_dst; i++)
 990       ureg_emit_dst( ureg, dst[i] );
 991
 992    for (i = 0; i < nr_src; i++)
 993       ureg_emit_src( ureg, src[i] );
 994
 995    ureg_fixup_insn_size( ureg, insn.insn_token );
 996 }
 997
 998 void
 999 ureg_tex_insn(struct ureg_program *ureg,
1000               unsigned opcode,
1001               const struct ureg_dst *dst,
1002               unsigned nr_dst,
1003               unsigned target,
1004               const struct ureg_src *src,
1005               unsigned nr_src )
1006 {
1007    struct ureg_emit_insn_result insn;
1008    unsigned i;
1009    boolean saturate;
1010    boolean predicate;
1011    boolean negate = FALSE;
1012    unsigned swizzle[4] = { 0 };
1013
1014    saturate = nr_dst ? dst[0].Saturate : FALSE;
1015    predicate = nr_dst ? dst[0].Predicate : FALSE;
1016    if (predicate) {
1017       negate = dst[0].PredNegate;
1018       swizzle[0] = dst[0].PredSwizzleX;
1019       swizzle[1] = dst[0].PredSwizzleY;
1020       swizzle[2] = dst[0].PredSwizzleZ;
1021       swizzle[3] = dst[0].PredSwizzleW;
1022    }
1023
1024    insn = ureg_emit_insn(ureg,
1025                          opcode,
1026                          saturate,
1027                          predicate,
1028                          negate,
1029                          swizzle[0],
1030                          swizzle[1],
1031                          swizzle[2],
1032                          swizzle[3],
1033                          nr_dst,
1034                          nr_src);
1035
1036    ureg_emit_texture( ureg, insn.extended_token, target );
1037
1038    for (i = 0; i < nr_dst; i++)
1039       ureg_emit_dst( ureg, dst[i] );
1040
1041    for (i = 0; i < nr_src; i++)
1042       ureg_emit_src( ureg, src[i] );
1043
1044    ureg_fixup_insn_size( ureg, insn.insn_token );
1045 }
1046
1047
1048 void
1049 ureg_label_insn(struct ureg_program *ureg,
1050                 unsigned opcode,
1051                 const struct ureg_src *src,
1052                 unsigned nr_src,
1053                 unsigned *label_token )
1054 {
1055    struct ureg_emit_insn_result insn;
1056    unsigned i;
1057
1058    insn = ureg_emit_insn(ureg,
1059                          opcode,
1060                          FALSE,
1061                          FALSE,
1062                          FALSE,
1063                          TGSI_SWIZZLE_X,
1064                          TGSI_SWIZZLE_Y,
1065                          TGSI_SWIZZLE_Z,
1066                          TGSI_SWIZZLE_W,
1067                          0,
1068                          nr_src);
1069
1070    ureg_emit_label( ureg, insn.extended_token, label_token );
1071
1072    for (i = 0; i < nr_src; i++)
1073       ureg_emit_src( ureg, src[i] );
1074
1075    ureg_fixup_insn_size( ureg, insn.insn_token );
1076 }
1077
1078
1079 static void
1080 emit_decl_semantic(struct ureg_program *ureg,
1081                    unsigned file,
1082                    unsigned index,
1083                    unsigned semantic_name,
1084                    unsigned semantic_index)
1085 {
1086    union tgsi_any_token *out = get_tokens(ureg, DOMAIN_DECL, 3);
1087
1088    out[0].value = 0;
1089    out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
1090    out[0].decl.NrTokens = 3;
1091    out[0].decl.File = file;
1092    out[0].decl.UsageMask = TGSI_WRITEMASK_XYZW; /* FIXME! */
1093    out[0].decl.Semantic = 1;
1094
1095    out[1].value = 0;
1096    out[1].decl_range.First = index;
1097    out[1].decl_range.Last = index;
1098
1099    out[2].value = 0;
1100    out[2].decl_semantic.Name = semantic_name;
1101    out[2].decl_semantic.Index = semantic_index;
1102 }
1103
1104
1105 static void
1106 emit_decl_fs(struct ureg_program *ureg,
1107              unsigned file,
1108              unsigned index,
1109              unsigned semantic_name,
1110              unsigned semantic_index,
1111              unsigned interpolate,
1112              unsigned cylindrical_wrap)
1113 {
1114    union tgsi_any_token *out = get_tokens(ureg, DOMAIN_DECL, 3);
1115
1116    out[0].value = 0;
1117    out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
1118    out[0].decl.NrTokens = 3;
1119    out[0].decl.File = file;
1120    out[0].decl.UsageMask = TGSI_WRITEMASK_XYZW; /* FIXME! */
1121    out[0].decl.Interpolate = interpolate;
1122    out[0].decl.Semantic = 1;
1123    out[0].decl.CylindricalWrap = cylindrical_wrap;
1124
1125    out[1].value = 0;
1126    out[1].decl_range.First = index;
1127    out[1].decl_range.Last = index;
1128
1129    out[2].value = 0;
1130    out[2].decl_semantic.Name = semantic_name;
1131    out[2].decl_semantic.Index = semantic_index;
1132 }
1133
1134
1135 static void emit_decl_range( struct ureg_program *ureg,
1136                              unsigned file,
1137                              unsigned first,
1138                              unsigned count )
1139 {
1140    union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 2 );
1141
1142    out[0].value = 0;
1143    out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
1144    out[0].decl.NrTokens = 2;
1145    out[0].decl.File = file;
1146    out[0].decl.UsageMask = TGSI_WRITEMASK_XYZW;
1147    out[0].decl.Interpolate = TGSI_INTERPOLATE_CONSTANT;
1148    out[0].decl.Semantic = 0;
1149
1150    out[1].value = 0;
1151    out[1].decl_range.First = first;
1152    out[1].decl_range.Last = first + count - 1;
1153 }
1154
1155 static void
1156 emit_decl_range2D(struct ureg_program *ureg,
1157                   unsigned file,
1158                   unsigned first,
1159                   unsigned last,
1160                   unsigned index2D)
1161 {
1162    union tgsi_any_token *out = get_tokens(ureg, DOMAIN_DECL, 3);
1163
1164    out[0].value = 0;
1165    out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
1166    out[0].decl.NrTokens = 3;
1167    out[0].decl.File = file;
1168    out[0].decl.UsageMask = TGSI_WRITEMASK_XYZW;
1169    out[0].decl.Interpolate = TGSI_INTERPOLATE_CONSTANT;
1170    out[0].decl.Dimension = 1;
1171
1172    out[1].value = 0;
1173    out[1].decl_range.First = first;
1174    out[1].decl_range.Last = last;
1175
1176    out[2].value = 0;
1177    out[2].decl_dim.Index2D = index2D;
1178 }
1179
1180 static void
1181 emit_immediate( struct ureg_program *ureg,
1182                 const unsigned *v,
1183                 unsigned type )
1184 {
1185    union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 5 );
1186
1187    out[0].value = 0;
1188    out[0].imm.Type = TGSI_TOKEN_TYPE_IMMEDIATE;
1189    out[0].imm.NrTokens = 5;
1190    out[0].imm.DataType = type;
1191    out[0].imm.Padding = 0;
1192
1193    out[1].imm_data.Uint = v[0];
1194    out[2].imm_data.Uint = v[1];
1195    out[3].imm_data.Uint = v[2];
1196    out[4].imm_data.Uint = v[3];
1197 }
1198
1199 static void
1200 emit_property(struct ureg_program *ureg,
1201               unsigned name,
1202               unsigned data)
1203 {
1204    union tgsi_any_token *out = get_tokens(ureg, DOMAIN_DECL, 2);
1205
1206    out[0].value = 0;
1207    out[0].prop.Type = TGSI_TOKEN_TYPE_PROPERTY;
1208    out[0].prop.NrTokens = 2;
1209    out[0].prop.PropertyName = name;
1210
1211    out[1].prop_data.Data = data;
1212 }
1213
1214
1215 static void emit_decls( struct ureg_program *ureg )
1216 {
1217    unsigned i;
1218
1219    if (ureg->property_gs_input_prim != ~0) {
1220       assert(ureg->processor == TGSI_PROCESSOR_GEOMETRY);
1221
1222       emit_property(ureg,
1223                     TGSI_PROPERTY_GS_INPUT_PRIM,
1224                     ureg->property_gs_input_prim);
1225    }
1226
1227    if (ureg->property_gs_output_prim != ~0) {
1228       assert(ureg->processor == TGSI_PROCESSOR_GEOMETRY);
1229
1230       emit_property(ureg,
1231                     TGSI_PROPERTY_GS_OUTPUT_PRIM,
1232                     ureg->property_gs_output_prim);
1233    }
1234
1235    if (ureg->property_gs_max_vertices != ~0) {
1236       assert(ureg->processor == TGSI_PROCESSOR_GEOMETRY);
1237
1238       emit_property(ureg,
1239                     TGSI_PROPERTY_GS_MAX_VERTICES,
1240                     ureg->property_gs_max_vertices);
1241    }
1242
1243    if (ureg->property_fs_coord_origin) {
1244       assert(ureg->processor == TGSI_PROCESSOR_FRAGMENT);
1245
1246       emit_property(ureg,
1247                     TGSI_PROPERTY_FS_COORD_ORIGIN,
1248                     ureg->property_fs_coord_origin);
1249    }
1250
1251    if (ureg->property_fs_coord_pixel_center) {
1252       assert(ureg->processor == TGSI_PROCESSOR_FRAGMENT);
1253
1254       emit_property(ureg,
1255                     TGSI_PROPERTY_FS_COORD_PIXEL_CENTER,
1256                     ureg->property_fs_coord_pixel_center);
1257    }
1258
1259    if (ureg->processor == TGSI_PROCESSOR_VERTEX) {
1260       for (i = 0; i < UREG_MAX_INPUT; i++) {
1261          if (ureg->vs_inputs[i/32] & (1 << (i%32))) {
1262             emit_decl_range( ureg, TGSI_FILE_INPUT, i, 1 );
1263          }
1264       }
1265    } else if (ureg->processor == TGSI_PROCESSOR_FRAGMENT) {
1266       for (i = 0; i < ureg->nr_fs_inputs; i++) {
1267          emit_decl_fs(ureg,
1268                       TGSI_FILE_INPUT,
1269                       i,
1270                       ureg->fs_input[i].semantic_name,
1271                       ureg->fs_input[i].semantic_index,
1272                       ureg->fs_input[i].interp,
1273                       ureg->fs_input[i].cylindrical_wrap);
1274       }
1275    } else {
1276       for (i = 0; i < ureg->nr_gs_inputs; i++) {
1277          emit_decl_semantic(ureg,
1278                             TGSI_FILE_INPUT,
1279                             ureg->gs_input[i].index,
1280                             ureg->gs_input[i].semantic_name,
1281                             ureg->gs_input[i].semantic_index);
1282       }
1283    }
1284
1285    for (i = 0; i < ureg->nr_system_values; i++) {
1286       emit_decl_semantic(ureg,
1287                          TGSI_FILE_SYSTEM_VALUE,
1288                          ureg->system_value[i].index,
1289                          ureg->system_value[i].semantic_name,
1290                          ureg->system_value[i].semantic_index);
1291    }
1292
1293    for (i = 0; i < ureg->nr_outputs; i++) {
1294       emit_decl_semantic(ureg,
1295                          TGSI_FILE_OUTPUT,
1296                          i,
1297                          ureg->output[i].semantic_name,
1298                          ureg->output[i].semantic_index);
1299    }
1300
1301    for (i = 0; i < ureg->nr_samplers; i++) {
1302       emit_decl_range( ureg,
1303                        TGSI_FILE_SAMPLER,
1304                        ureg->sampler[i].Index, 1 );
1305    }
1306
1307    if (ureg->const_decls.nr_constant_ranges) {
1308       for (i = 0; i < ureg->const_decls.nr_constant_ranges; i++) {
1309          emit_decl_range(ureg,
1310                          TGSI_FILE_CONSTANT,
1311                          ureg->const_decls.constant_range[i].first,
1312                          ureg->const_decls.constant_range[i].last - ureg->const_decls.constant_range[i].first + 1);
1313       }
1314    }
1315
1316    for (i = 0; i < PIPE_MAX_CONSTANT_BUFFERS; i++) {
1317       struct const_decl *decl = &ureg->const_decls2D[i];
1318
1319       if (decl->nr_constant_ranges) {
1320          uint j;
1321
1322          for (j = 0; j < decl->nr_constant_ranges; j++) {
1323             emit_decl_range2D(ureg,
1324                               TGSI_FILE_CONSTANT,
1325                               decl->constant_range[j].first,
1326                               decl->constant_range[j].last,
1327                               i);
1328          }
1329       }
1330    }
1331
1332    if (ureg->nr_temps) {
1333       emit_decl_range( ureg,
1334                        TGSI_FILE_TEMPORARY,
1335                        0, ureg->nr_temps );
1336    }
1337
1338    if (ureg->nr_addrs) {
1339       emit_decl_range( ureg,
1340                        TGSI_FILE_ADDRESS,
1341                        0, ureg->nr_addrs );
1342    }
1343
1344    if (ureg->nr_preds) {
1345       emit_decl_range(ureg,
1346                       TGSI_FILE_PREDICATE,
1347                       0,
1348                       ureg->nr_preds);
1349    }
1350
1351    for (i = 0; i < ureg->nr_immediates; i++) {
1352       emit_immediate( ureg,
1353                       ureg->immediate[i].value.u,
1354                       ureg->immediate[i].type );
1355    }
1356 }
1357
1358 /* Append the instruction tokens onto the declarations to build a
1359  * contiguous stream suitable to send to the driver.
1360  */
1361 static void copy_instructions( struct ureg_program *ureg )
1362 {
1363    unsigned nr_tokens = ureg->domain[DOMAIN_INSN].count;
1364    union tgsi_any_token *out = get_tokens( ureg,
1365                                            DOMAIN_DECL,
1366                                            nr_tokens );
1367
1368    memcpy(out,
1369           ureg->domain[DOMAIN_INSN].tokens,
1370           nr_tokens * sizeof out[0] );
1371 }
1372
1373
1374 static void
1375 fixup_header_size(struct ureg_program *ureg)
1376 {
1377    union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_DECL, 0 );
1378
1379    out->header.BodySize = ureg->domain[DOMAIN_DECL].count - 2;
1380 }
1381
1382
1383 static void
1384 emit_header( struct ureg_program *ureg )
1385 {
1386    union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 2 );
1387
1388    out[0].header.HeaderSize = 2;
1389    out[0].header.BodySize = 0;
1390
1391    out[1].processor.Processor = ureg->processor;
1392    out[1].processor.Padding = 0;
1393 }
1394
1395
1396 const struct tgsi_token *ureg_finalize( struct ureg_program *ureg )
1397 {
1398    const struct tgsi_token *tokens;
1399
1400    emit_header( ureg );
1401    emit_decls( ureg );
1402    copy_instructions( ureg );
1403    fixup_header_size( ureg );
1404
1405    if (ureg->domain[0].tokens == error_tokens ||
1406        ureg->domain[1].tokens == error_tokens) {
1407       debug_printf("%s: error in generated shader\n", __FUNCTION__);
1408       assert(0);
1409       return NULL;
1410    }
1411
1412    tokens = &ureg->domain[DOMAIN_DECL].tokens[0].token;
1413
1414    if (0) {
1415       debug_printf("%s: emitted shader %d tokens:\n", __FUNCTION__,
1416                    ureg->domain[DOMAIN_DECL].count);
1417       tgsi_dump( tokens, 0 );
1418    }
1419
1420 #if DEBUG
1421    if (tokens && !tgsi_sanity_check(tokens)) {
1422       debug_printf("tgsi_ureg.c, sanity check failed on generated tokens:\n");
1423       tgsi_dump(tokens, 0);
1424       assert(0);
1425    }
1426 #endif
1427
1428
1429    return tokens;
1430 }
1431
1432
1433 void *ureg_create_shader( struct ureg_program *ureg,
1434                           struct pipe_context *pipe )
1435 {
1436    struct pipe_shader_state state;
1437
1438    state.tokens = ureg_finalize(ureg);
1439    if(!state.tokens)
1440       return NULL;
1441
1442    if (ureg->processor == TGSI_PROCESSOR_VERTEX)
1443       return pipe->create_vs_state( pipe, &state );
1444    else
1445       return pipe->create_fs_state( pipe, &state );
1446 }
1447
1448
1449 const struct tgsi_token *ureg_get_tokens( struct ureg_program *ureg,
1450                                           unsigned *nr_tokens )
1451 {
1452    const struct tgsi_token *tokens;
1453
1454    ureg_finalize(ureg);
1455
1456    tokens = &ureg->domain[DOMAIN_DECL].tokens[0].token;
1457
1458    if (nr_tokens)
1459       *nr_tokens = ureg->domain[DOMAIN_DECL].size;
1460
1461    ureg->domain[DOMAIN_DECL].tokens = 0;
1462    ureg->domain[DOMAIN_DECL].size = 0;
1463    ureg->domain[DOMAIN_DECL].order = 0;
1464    ureg->domain[DOMAIN_DECL].count = 0;
1465
1466    return tokens;
1467 }
1468
1469
1470 struct ureg_program *ureg_create( unsigned processor )
1471 {
1472    struct ureg_program *ureg = CALLOC_STRUCT( ureg_program );
1473    if (ureg == NULL)
1474       return NULL;
1475
1476    ureg->processor = processor;
1477    ureg->property_gs_input_prim = ~0;
1478    ureg->property_gs_output_prim = ~0;
1479    ureg->property_gs_max_vertices = ~0;
1480    return ureg;
1481 }
1482
1483
1484 void ureg_destroy( struct ureg_program *ureg )
1485 {
1486    unsigned i;
1487
1488    for (i = 0; i < Elements(ureg->domain); i++) {
1489       if (ureg->domain[i].tokens &&
1490           ureg->domain[i].tokens != error_tokens)
1491          FREE(ureg->domain[i].tokens);
1492    }
1493
1494    FREE(ureg);
1495 }