src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * 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
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * TGSI to LLVM IR translation -- SoA.
  32  *
  33  * @author Jose Fonseca <jfonseca@vmware.com>
  34  *
  35  * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
  36  * Brian Paul, and others.
  37  */
  38
  39 #include "pipe/p_config.h"
  40 #include "pipe/p_shader_tokens.h"
  41 #include "util/u_debug.h"
  42 #include "util/u_math.h"
  43 #include "util/u_memory.h"
  44 #include "tgsi/tgsi_dump.h"
  45 #include "tgsi/tgsi_info.h"
  46 #include "tgsi/tgsi_parse.h"
  47 #include "tgsi/tgsi_util.h"
  48 #include "tgsi/tgsi_exec.h"
  49 #include "lp_bld_type.h"
  50 #include "lp_bld_const.h"
  51 #include "lp_bld_arit.h"
  52 #include "lp_bld_logic.h"
  53 #include "lp_bld_swizzle.h"
  54 #include "lp_bld_flow.h"
  55 #include "lp_bld_tgsi.h"
  56 #include "lp_bld_debug.h"
  57
  58
  59 #define LP_MAX_TEMPS 256
  60 #define LP_MAX_IMMEDIATES 256
  61
  62
  63 #define FOR_EACH_CHANNEL( CHAN )\
  64    for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
  65
  66 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  67    ((INST)->Dst[0].Register.WriteMask & (1 << (CHAN)))
  68
  69 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  70    if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
  71
  72 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
  73    FOR_EACH_CHANNEL( CHAN )\
  74       IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
  75
  76 #define CHAN_X 0
  77 #define CHAN_Y 1
  78 #define CHAN_Z 2
  79 #define CHAN_W 3
  80
  81 #define QUAD_TOP_LEFT     0
  82 #define QUAD_TOP_RIGHT    1
  83 #define QUAD_BOTTOM_LEFT  2
  84 #define QUAD_BOTTOM_RIGHT 3
  85
  86 #define LP_TGSI_MAX_NESTING 16
  87
  88 struct lp_exec_mask {
  89    struct lp_build_context *bld;
  90
  91    boolean has_mask;
  92
  93    LLVMTypeRef int_vec_type;
  94
  95    LLVMValueRef cond_stack[LP_TGSI_MAX_NESTING];
  96    int cond_stack_size;
  97    LLVMValueRef cond_mask;
  98
  99    LLVMValueRef break_stack[LP_TGSI_MAX_NESTING];
 100    int break_stack_size;
 101    LLVMValueRef break_mask;
 102
 103    LLVMValueRef cont_stack[LP_TGSI_MAX_NESTING];
 104    int cont_stack_size;
 105    LLVMValueRef cont_mask;
 106
 107    LLVMBasicBlockRef loop_stack[LP_TGSI_MAX_NESTING];
 108    int loop_stack_size;
 109    LLVMBasicBlockRef loop_block;
 110
 111
 112    LLVMValueRef exec_mask;
 113 };
 114
 115 struct lp_build_tgsi_soa_context
 116 {
 117    struct lp_build_context base;
 118
 119    LLVMValueRef consts_ptr;
 120    const LLVMValueRef *pos;
 121    const LLVMValueRef (*inputs)[NUM_CHANNELS];
 122    LLVMValueRef (*outputs)[NUM_CHANNELS];
 123
 124    struct lp_build_sampler_soa *sampler;
 125
 126    LLVMValueRef immediates[LP_MAX_IMMEDIATES][NUM_CHANNELS];
 127    LLVMValueRef temps[LP_MAX_TEMPS][NUM_CHANNELS];
 128
 129    struct lp_build_mask_context *mask;
 130    struct lp_exec_mask exec_mask;
 131 };
 132
 133 static const unsigned char
 134 swizzle_left[4] = {
 135    QUAD_TOP_LEFT,     QUAD_TOP_LEFT,
 136    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_LEFT
 137 };
 138
 139 static const unsigned char
 140 swizzle_right[4] = {
 141    QUAD_TOP_RIGHT,    QUAD_TOP_RIGHT,
 142    QUAD_BOTTOM_RIGHT, QUAD_BOTTOM_RIGHT
 143 };
 144
 145 static const unsigned char
 146 swizzle_top[4] = {
 147    QUAD_TOP_LEFT,     QUAD_TOP_RIGHT,
 148    QUAD_TOP_LEFT,     QUAD_TOP_RIGHT
 149 };
 150
 151 static const unsigned char
 152 swizzle_bottom[4] = {
 153    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT,
 154    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT
 155 };
 156
 157 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
 158 {
 159    mask->bld = bld;
 160    mask->has_mask = FALSE;
 161    mask->cond_stack_size = 0;
 162    mask->loop_stack_size = 0;
 163    mask->break_stack_size = 0;
 164    mask->cont_stack_size = 0;
 165
 166    mask->int_vec_type = lp_build_int_vec_type(mask->bld->type);
 167 }
 168
 169 static void lp_exec_mask_update(struct lp_exec_mask *mask)
 170 {
 171    if (mask->loop_stack_size) {
 172       /*for loops we need to update the entire mask at
 173        * runtime */
 174       LLVMValueRef tmp;
 175       tmp = LLVMBuildAnd(mask->bld->builder,
 176                          mask->cont_mask,
 177                          mask->break_mask,
 178                          "maskcb");
 179       mask->exec_mask = LLVMBuildAnd(mask->bld->builder,
 180                                      mask->cond_mask,
 181                                      tmp,
 182                                      "maskfull");
 183    } else
 184       mask->exec_mask = mask->cond_mask;
 185
 186
 187    mask->has_mask = (mask->cond_stack_size > 0 ||
 188                      mask->loop_stack_size > 0);
 189 }
 190
 191 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 192                                    LLVMValueRef val)
 193 {
 194    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 195    mask->cond_mask = LLVMBuildBitCast(mask->bld->builder, val,
 196                                       mask->int_vec_type, "");
 197
 198    lp_exec_mask_update(mask);
 199 }
 200
 201 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 202 {
 203    LLVMValueRef prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 204    LLVMValueRef inv_mask = LLVMBuildNot(mask->bld->builder,
 205                                         mask->cond_mask, "");
 206
 207    /* means that we didn't have any mask before and that
 208     * we were fully enabled */
 209    if (mask->cond_stack_size <= 1) {
 210       prev_mask = LLVMConstAllOnes(mask->int_vec_type);
 211    }
 212
 213    mask->cond_mask = LLVMBuildAnd(mask->bld->builder,
 214                                   inv_mask,
 215                                   prev_mask, "");
 216    lp_exec_mask_update(mask);
 217 }
 218
 219 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 220 {
 221    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 222    lp_exec_mask_update(mask);
 223 }
 224
 225 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 226 {
 227
 228    if (mask->cont_stack_size == 0)
 229       mask->cont_mask = LLVMConstAllOnes(mask->int_vec_type);
 230    if (mask->cont_stack_size == 0)
 231       mask->break_mask = LLVMConstAllOnes(mask->int_vec_type);
 232    if (mask->cond_stack_size == 0)
 233       mask->cond_mask = LLVMConstAllOnes(mask->int_vec_type);
 234    mask->loop_stack[mask->loop_stack_size++] = mask->loop_block;
 235    mask->loop_block = lp_build_insert_new_block(mask->bld->builder, "bgnloop");
 236    LLVMBuildBr(mask->bld->builder, mask->loop_block);
 237    LLVMPositionBuilderAtEnd(mask->bld->builder, mask->loop_block);
 238
 239    lp_exec_mask_update(mask);
 240 }
 241
 242 static void lp_exec_break(struct lp_exec_mask *mask)
 243 {
 244    LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
 245                                          mask->exec_mask,
 246                                          "break");
 247
 248    mask->break_stack[mask->break_stack_size++] = mask->break_mask;
 249    if (mask->break_stack_size > 1) {
 250       mask->break_mask = LLVMBuildAnd(mask->bld->builder,
 251                                       mask->break_mask,
 252                                       exec_mask, "break_full");
 253    } else
 254       mask->break_mask = exec_mask;
 255
 256    lp_exec_mask_update(mask);
 257 }
 258
 259 static void lp_exec_continue(struct lp_exec_mask *mask)
 260 {
 261    LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
 262                                          mask->exec_mask,
 263                                          "");
 264
 265    mask->cont_stack[mask->cont_stack_size++] = mask->cont_mask;
 266    if (mask->cont_stack_size > 1) {
 267       mask->cont_mask = LLVMBuildAnd(mask->bld->builder,
 268                                      mask->cont_mask,
 269                                      exec_mask, "");
 270    } else
 271       mask->cont_mask = exec_mask;
 272
 273    lp_exec_mask_update(mask);
 274 }
 275
 276
 277 static void lp_exec_endloop(struct lp_exec_mask *mask)
 278 {
 279    LLVMBasicBlockRef endloop;
 280    LLVMTypeRef reg_type = LLVMIntType(mask->bld->type.width*
 281                                       mask->bld->type.length);
 282    /* i1cond = (mask == 0) */
 283    LLVMValueRef i1cond = LLVMBuildICmp(
 284       mask->bld->builder,
 285       LLVMIntNE,
 286       LLVMBuildBitCast(mask->bld->builder, mask->break_mask, reg_type, ""),
 287       LLVMConstNull(reg_type), "");
 288
 289    endloop = lp_build_insert_new_block(mask->bld->builder, "endloop");
 290
 291    LLVMBuildCondBr(mask->bld->builder,
 292                    i1cond, mask->loop_block, endloop);
 293
 294    LLVMPositionBuilderAtEnd(mask->bld->builder, endloop);
 295
 296    mask->loop_block = mask->loop_stack[--mask->loop_stack_size];
 297    /* pop the break mask */
 298    if (mask->cont_stack_size) {
 299       mask->cont_mask = mask->cont_stack[--mask->cont_stack_size];
 300    }
 301    if (mask->break_stack_size) {
 302       mask->break_mask = mask->cont_stack[--mask->break_stack_size];
 303    }
 304
 305    lp_exec_mask_update(mask);
 306 }
 307
 308 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 309                                LLVMValueRef val,
 310                                LLVMValueRef dst)
 311 {
 312    if (mask->has_mask) {
 313       LLVMValueRef real_val, dst_val;
 314
 315       dst_val = LLVMBuildLoad(mask->bld->builder, dst, "");
 316       real_val = lp_build_select(mask->bld,
 317                                  mask->exec_mask,
 318                                  val, dst_val);
 319
 320       LLVMBuildStore(mask->bld->builder, real_val, dst);
 321    } else
 322       LLVMBuildStore(mask->bld->builder, val, dst);
 323 }
 324
 325
 326 static LLVMValueRef
 327 emit_ddx(struct lp_build_tgsi_soa_context *bld,
 328          LLVMValueRef src)
 329 {
 330    LLVMValueRef src_left  = lp_build_swizzle1_aos(&bld->base, src, swizzle_left);
 331    LLVMValueRef src_right = lp_build_swizzle1_aos(&bld->base, src, swizzle_right);
 332    return lp_build_sub(&bld->base, src_right, src_left);
 333 }
 334
 335
 336 static LLVMValueRef
 337 emit_ddy(struct lp_build_tgsi_soa_context *bld,
 338          LLVMValueRef src)
 339 {
 340    LLVMValueRef src_top    = lp_build_swizzle1_aos(&bld->base, src, swizzle_top);
 341    LLVMValueRef src_bottom = lp_build_swizzle1_aos(&bld->base, src, swizzle_bottom);
 342    return lp_build_sub(&bld->base, src_top, src_bottom);
 343 }
 344
 345
 346 /**
 347  * Register fetch.
 348  */
 349 static LLVMValueRef
 350 emit_fetch(
 351    struct lp_build_tgsi_soa_context *bld,
 352    const struct tgsi_full_instruction *inst,
 353    unsigned index,
 354    const unsigned chan_index )
 355 {
 356    const struct tgsi_full_src_register *reg = &inst->Src[index];
 357    unsigned swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 358    LLVMValueRef res;
 359
 360    switch (swizzle) {
 361    case TGSI_SWIZZLE_X:
 362    case TGSI_SWIZZLE_Y:
 363    case TGSI_SWIZZLE_Z:
 364    case TGSI_SWIZZLE_W:
 365
 366       switch (reg->Register.File) {
 367       case TGSI_FILE_CONSTANT: {
 368          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), reg->Register.Index*4 + swizzle, 0);
 369          LLVMValueRef scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr, &index, 1, "");
 370          LLVMValueRef scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
 371          res = lp_build_broadcast_scalar(&bld->base, scalar);
 372          break;
 373       }
 374
 375       case TGSI_FILE_IMMEDIATE:
 376          res = bld->immediates[reg->Register.Index][swizzle];
 377          assert(res);
 378          break;
 379
 380       case TGSI_FILE_INPUT:
 381          res = bld->inputs[reg->Register.Index][swizzle];
 382          assert(res);
 383          break;
 384
 385       case TGSI_FILE_TEMPORARY:
 386          res = LLVMBuildLoad(bld->base.builder, bld->temps[reg->Register.Index][swizzle], "");
 387          if(!res)
 388             return bld->base.undef;
 389          break;
 390
 391       default:
 392          assert( 0 );
 393          return bld->base.undef;
 394       }
 395       break;
 396
 397    default:
 398       assert( 0 );
 399       return bld->base.undef;
 400    }
 401
 402    switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
 403    case TGSI_UTIL_SIGN_CLEAR:
 404       res = lp_build_abs( &bld->base, res );
 405       break;
 406
 407    case TGSI_UTIL_SIGN_SET:
 408       /* TODO: Use bitwese OR for floating point */
 409       res = lp_build_abs( &bld->base, res );
 410       res = LLVMBuildNeg( bld->base.builder, res, "" );
 411       break;
 412
 413    case TGSI_UTIL_SIGN_TOGGLE:
 414       res = LLVMBuildNeg( bld->base.builder, res, "" );
 415       break;
 416
 417    case TGSI_UTIL_SIGN_KEEP:
 418       break;
 419    }
 420
 421    return res;
 422 }
 423
 424
 425 /**
 426  * Register fetch with derivatives.
 427  */
 428 static void
 429 emit_fetch_deriv(
 430    struct lp_build_tgsi_soa_context *bld,
 431    const struct tgsi_full_instruction *inst,
 432    unsigned index,
 433    const unsigned chan_index,
 434    LLVMValueRef *res,
 435    LLVMValueRef *ddx,
 436    LLVMValueRef *ddy)
 437 {
 438    LLVMValueRef src;
 439
 440    src = emit_fetch(bld, inst, index, chan_index);
 441
 442    if(res)
 443       *res = src;
 444
 445    /* TODO: use interpolation coeffs for inputs */
 446
 447    if(ddx)
 448       *ddx = emit_ddx(bld, src);
 449
 450    if(ddy)
 451       *ddy = emit_ddy(bld, src);
 452 }
 453
 454
 455 /**
 456  * Register store.
 457  */
 458 static void
 459 emit_store(
 460    struct lp_build_tgsi_soa_context *bld,
 461    const struct tgsi_full_instruction *inst,
 462    unsigned index,
 463    unsigned chan_index,
 464    LLVMValueRef value)
 465 {
 466    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 467
 468    switch( inst->Instruction.Saturate ) {
 469    case TGSI_SAT_NONE:
 470       break;
 471
 472    case TGSI_SAT_ZERO_ONE:
 473       value = lp_build_max(&bld->base, value, bld->base.zero);
 474       value = lp_build_min(&bld->base, value, bld->base.one);
 475       break;
 476
 477    case TGSI_SAT_MINUS_PLUS_ONE:
 478       value = lp_build_max(&bld->base, value, lp_build_const_vec(bld->base.type, -1.0));
 479       value = lp_build_min(&bld->base, value, bld->base.one);
 480       break;
 481
 482    default:
 483       assert(0);
 484    }
 485
 486    switch( reg->Register.File ) {
 487    case TGSI_FILE_OUTPUT:
 488       lp_exec_mask_store(&bld->exec_mask, value,
 489                          bld->outputs[reg->Register.Index][chan_index]);
 490       break;
 491
 492    case TGSI_FILE_TEMPORARY:
 493       lp_exec_mask_store(&bld->exec_mask, value,
 494                          bld->temps[reg->Register.Index][chan_index]);
 495       break;
 496
 497    case TGSI_FILE_ADDRESS:
 498       /* FIXME */
 499       assert(0);
 500       break;
 501
 502    case TGSI_FILE_PREDICATE:
 503       /* FIXME */
 504       assert(0);
 505       break;
 506
 507    default:
 508       assert( 0 );
 509    }
 510 }
 511
 512
 513 /**
 514  * High-level instruction translators.
 515  */
 516
 517
 518 static void
 519 emit_tex( struct lp_build_tgsi_soa_context *bld,
 520           const struct tgsi_full_instruction *inst,
 521           boolean apply_lodbias,
 522           boolean projected,
 523           LLVMValueRef *texel)
 524 {
 525    const uint unit = inst->Src[1].Register.Index;
 526    LLVMValueRef lodbias;
 527    LLVMValueRef oow = NULL;
 528    LLVMValueRef coords[3];
 529    unsigned num_coords;
 530    unsigned i;
 531
 532    switch (inst->Texture.Texture) {
 533    case TGSI_TEXTURE_1D:
 534       num_coords = 1;
 535       break;
 536    case TGSI_TEXTURE_2D:
 537    case TGSI_TEXTURE_RECT:
 538       num_coords = 2;
 539       break;
 540    case TGSI_TEXTURE_SHADOW1D:
 541    case TGSI_TEXTURE_SHADOW2D:
 542    case TGSI_TEXTURE_SHADOWRECT:
 543    case TGSI_TEXTURE_3D:
 544    case TGSI_TEXTURE_CUBE:
 545       num_coords = 3;
 546       break;
 547    default:
 548       assert(0);
 549       return;
 550    }
 551
 552    if(apply_lodbias)
 553       lodbias = emit_fetch( bld, inst, 0, 3 );
 554    else
 555       lodbias = bld->base.zero;
 556
 557    if (projected) {
 558       oow = emit_fetch( bld, inst, 0, 3 );
 559       oow = lp_build_rcp(&bld->base, oow);
 560    }
 561
 562    for (i = 0; i < num_coords; i++) {
 563       coords[i] = emit_fetch( bld, inst, 0, i );
 564       if (projected)
 565          coords[i] = lp_build_mul(&bld->base, coords[i], oow);
 566    }
 567    for (i = num_coords; i < 3; i++) {
 568       coords[i] = bld->base.undef;
 569    }
 570
 571    bld->sampler->emit_fetch_texel(bld->sampler,
 572                                   bld->base.builder,
 573                                   bld->base.type,
 574                                   unit, num_coords, coords, lodbias,
 575                                   texel);
 576 }
 577
 578
 579 static void
 580 emit_kil(
 581    struct lp_build_tgsi_soa_context *bld,
 582    const struct tgsi_full_instruction *inst )
 583 {
 584    const struct tgsi_full_src_register *reg = &inst->Src[0];
 585    LLVMValueRef terms[NUM_CHANNELS];
 586    LLVMValueRef mask;
 587    unsigned chan_index;
 588
 589    memset(&terms, 0, sizeof terms);
 590
 591    FOR_EACH_CHANNEL( chan_index ) {
 592       unsigned swizzle;
 593
 594       /* Unswizzle channel */
 595       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 596
 597       /* Check if the component has not been already tested. */
 598       assert(swizzle < NUM_CHANNELS);
 599       if( !terms[swizzle] )
 600          /* TODO: change the comparison operator instead of setting the sign */
 601          terms[swizzle] =  emit_fetch(bld, inst, 0, chan_index );
 602    }
 603
 604    mask = NULL;
 605    FOR_EACH_CHANNEL( chan_index ) {
 606       if(terms[chan_index]) {
 607          LLVMValueRef chan_mask;
 608
 609          chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero);
 610
 611          if(mask)
 612             mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, "");
 613          else
 614             mask = chan_mask;
 615       }
 616    }
 617
 618    if(mask)
 619       lp_build_mask_update(bld->mask, mask);
 620 }
 621
 622
 623 /**
 624  * Check if inst src/dest regs use indirect addressing into temporary
 625  * register file.
 626  */
 627 static boolean
 628 indirect_temp_reference(const struct tgsi_full_instruction *inst)
 629 {
 630    uint i;
 631    for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
 632       const struct tgsi_full_src_register *reg = &inst->Src[i];
 633       if (reg->Register.File == TGSI_FILE_TEMPORARY &&
 634           reg->Register.Indirect)
 635          return TRUE;
 636    }
 637    for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
 638       const struct tgsi_full_dst_register *reg = &inst->Dst[i];
 639       if (reg->Register.File == TGSI_FILE_TEMPORARY &&
 640           reg->Register.Indirect)
 641          return TRUE;
 642    }
 643    return FALSE;
 644 }
 645
 646 static int
 647 emit_declaration(
 648    struct lp_build_tgsi_soa_context *bld,
 649    const struct tgsi_full_declaration *decl)
 650 {
 651    unsigned first = decl->Range.First;
 652    unsigned last = decl->Range.Last;
 653    unsigned idx, i;
 654
 655    for (idx = first; idx <= last; ++idx) {
 656       boolean ok;
 657
 658       switch (decl->Declaration.File) {
 659       case TGSI_FILE_TEMPORARY:
 660          for (i = 0; i < NUM_CHANNELS; i++)
 661             bld->temps[idx][i] = lp_build_alloca(&bld->base);
 662          ok = TRUE;
 663          break;
 664
 665       case TGSI_FILE_OUTPUT:
 666          for (i = 0; i < NUM_CHANNELS; i++)
 667             bld->outputs[idx][i] = lp_build_alloca(&bld->base);
 668          ok = TRUE;
 669          break;
 670
 671       default:
 672          /* don't need to declare other vars */
 673          ok = TRUE;
 674       }
 675
 676       if (!ok)
 677          return FALSE;
 678    }
 679
 680    return TRUE;
 681 }
 682
 683 static int
 684 emit_instruction(
 685    struct lp_build_tgsi_soa_context *bld,
 686    const struct tgsi_full_instruction *inst,
 687    const struct tgsi_opcode_info *info)
 688 {
 689    unsigned chan_index;
 690    LLVMValueRef src0, src1, src2;
 691    LLVMValueRef tmp0, tmp1, tmp2;
 692    LLVMValueRef tmp3 = NULL;
 693    LLVMValueRef tmp4 = NULL;
 694    LLVMValueRef tmp5 = NULL;
 695    LLVMValueRef tmp6 = NULL;
 696    LLVMValueRef tmp7 = NULL;
 697    LLVMValueRef res;
 698    LLVMValueRef dst0[NUM_CHANNELS];
 699
 700    /* we can't handle indirect addressing into temp register file yet */
 701    if (indirect_temp_reference(inst))
 702       return FALSE;
 703
 704    /*
 705     * Stores and write masks are handled in a general fashion after the long
 706     * instruction opcode switch statement.
 707     *
 708     * Although not stricitly necessary, we avoid generating instructions for
 709     * channels which won't be stored, in cases where's that easy. For some
 710     * complex instructions, like texture sampling, it is more convenient to
 711     * assume a full writemask and then let LLVM optimization passes eliminate
 712     * redundant code.
 713     */
 714
 715    assert(info->num_dst <= 1);
 716    if(info->num_dst) {
 717       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 718          dst0[chan_index] = bld->base.undef;
 719       }
 720    }
 721
 722    switch (inst->Instruction.Opcode) {
 723 #if 0
 724    case TGSI_OPCODE_ARL:
 725       /* FIXME */
 726       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 727          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 728          emit_flr(bld, 0, 0);
 729          emit_f2it( bld, 0 );
 730          dst0[chan_index] = tmp0;
 731       }
 732       break;
 733 #endif
 734
 735    case TGSI_OPCODE_MOV:
 736       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 737          dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
 738       }
 739       break;
 740
 741    case TGSI_OPCODE_LIT:
 742       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 743          dst0[CHAN_X] = bld->base.one;
 744       }
 745       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 746          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 747          dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
 748       }
 749       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 750          /* XMM[1] = SrcReg[0].yyyy */
 751          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 752          /* XMM[1] = max(XMM[1], 0) */
 753          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 754          /* XMM[2] = SrcReg[0].wwww */
 755          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 756          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 757          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 758          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 759          dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 760       }
 761       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 762          dst0[CHAN_W] = bld->base.one;
 763       }
 764       break;
 765
 766    case TGSI_OPCODE_RCP:
 767    /* TGSI_OPCODE_RECIP */
 768       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 769       res = lp_build_rcp(&bld->base, src0);
 770       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 771          dst0[chan_index] = res;
 772       }
 773       break;
 774
 775    case TGSI_OPCODE_RSQ:
 776    /* TGSI_OPCODE_RECIPSQRT */
 777       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 778       src0 = lp_build_abs(&bld->base, src0);
 779       res = lp_build_rsqrt(&bld->base, src0);
 780       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 781          dst0[chan_index] = res;
 782       }
 783       break;
 784
 785    case TGSI_OPCODE_EXP:
 786       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 787           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 788           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 789          LLVMValueRef *p_exp2_int_part = NULL;
 790          LLVMValueRef *p_frac_part = NULL;
 791          LLVMValueRef *p_exp2 = NULL;
 792
 793          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 794
 795          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 796             p_exp2_int_part = &tmp0;
 797          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 798             p_frac_part = &tmp1;
 799          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 800             p_exp2 = &tmp2;
 801
 802          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 803
 804          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 805             dst0[CHAN_X] = tmp0;
 806          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 807             dst0[CHAN_Y] = tmp1;
 808          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 809             dst0[CHAN_Z] = tmp2;
 810       }
 811       /* dst.w = 1.0 */
 812       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 813          dst0[CHAN_W] = bld->base.one;
 814       }
 815       break;
 816
 817    case TGSI_OPCODE_LOG:
 818       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 819           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 820           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 821          LLVMValueRef *p_floor_log2 = NULL;
 822          LLVMValueRef *p_exp = NULL;
 823          LLVMValueRef *p_log2 = NULL;
 824
 825          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 826          src0 = lp_build_abs( &bld->base, src0 );
 827
 828          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 829             p_floor_log2 = &tmp0;
 830          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 831             p_exp = &tmp1;
 832          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 833             p_log2 = &tmp2;
 834
 835          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 836
 837          /* dst.x = floor(lg2(abs(src.x))) */
 838          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 839             dst0[CHAN_X] = tmp0;
 840          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 841          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 842             dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
 843          }
 844          /* dst.z = lg2(abs(src.x)) */
 845          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 846             dst0[CHAN_Z] = tmp2;
 847       }
 848       /* dst.w = 1.0 */
 849       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 850          dst0[CHAN_W] = bld->base.one;
 851       }
 852       break;
 853
 854    case TGSI_OPCODE_MUL:
 855       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 856          src0 = emit_fetch( bld, inst, 0, chan_index );
 857          src1 = emit_fetch( bld, inst, 1, chan_index );
 858          dst0[chan_index] = lp_build_mul(&bld->base, src0, src1);
 859       }
 860       break;
 861
 862    case TGSI_OPCODE_ADD:
 863       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 864          src0 = emit_fetch( bld, inst, 0, chan_index );
 865          src1 = emit_fetch( bld, inst, 1, chan_index );
 866          dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
 867       }
 868       break;
 869
 870    case TGSI_OPCODE_DP3:
 871    /* TGSI_OPCODE_DOT3 */
 872       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 873       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 874       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 875       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 876       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 877       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 878       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 879       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 880       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 881       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 882       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 883       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 884          dst0[chan_index] = tmp0;
 885       }
 886       break;
 887
 888    case TGSI_OPCODE_DP4:
 889    /* TGSI_OPCODE_DOT4 */
 890       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 891       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 892       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 893       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 894       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 895       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 896       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 897       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 898       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 899       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 900       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 901       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
 902       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
 903       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 904       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 905       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 906          dst0[chan_index] = tmp0;
 907       }
 908       break;
 909
 910    case TGSI_OPCODE_DST:
 911       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
 912          dst0[CHAN_X] = bld->base.one;
 913       }
 914       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
 915          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
 916          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
 917          dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
 918       }
 919       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
 920          dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
 921       }
 922       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
 923          dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
 924       }
 925       break;
 926
 927    case TGSI_OPCODE_MIN:
 928       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 929          src0 = emit_fetch( bld, inst, 0, chan_index );
 930          src1 = emit_fetch( bld, inst, 1, chan_index );
 931          dst0[chan_index] = lp_build_min( &bld->base, src0, src1 );
 932       }
 933       break;
 934
 935    case TGSI_OPCODE_MAX:
 936       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 937          src0 = emit_fetch( bld, inst, 0, chan_index );
 938          src1 = emit_fetch( bld, inst, 1, chan_index );
 939          dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
 940       }
 941       break;
 942
 943    case TGSI_OPCODE_SLT:
 944    /* TGSI_OPCODE_SETLT */
 945       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 946          src0 = emit_fetch( bld, inst, 0, chan_index );
 947          src1 = emit_fetch( bld, inst, 1, chan_index );
 948          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
 949          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 950       }
 951       break;
 952
 953    case TGSI_OPCODE_SGE:
 954    /* TGSI_OPCODE_SETGE */
 955       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 956          src0 = emit_fetch( bld, inst, 0, chan_index );
 957          src1 = emit_fetch( bld, inst, 1, chan_index );
 958          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
 959          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 960       }
 961       break;
 962
 963    case TGSI_OPCODE_MAD:
 964    /* TGSI_OPCODE_MADD */
 965       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 966          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 967          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 968          tmp2 = emit_fetch( bld, inst, 2, chan_index );
 969          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 970          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
 971          dst0[chan_index] = tmp0;
 972       }
 973       break;
 974
 975    case TGSI_OPCODE_SUB:
 976       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 977          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 978          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 979          dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
 980       }
 981       break;
 982
 983    case TGSI_OPCODE_LRP:
 984       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 985          src0 = emit_fetch( bld, inst, 0, chan_index );
 986          src1 = emit_fetch( bld, inst, 1, chan_index );
 987          src2 = emit_fetch( bld, inst, 2, chan_index );
 988          tmp0 = lp_build_sub( &bld->base, src1, src2 );
 989          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
 990          dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
 991       }
 992       break;
 993
 994    case TGSI_OPCODE_CND:
 995       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 996          src0 = emit_fetch( bld, inst, 0, chan_index );
 997          src1 = emit_fetch( bld, inst, 1, chan_index );
 998          src2 = emit_fetch( bld, inst, 2, chan_index );
 999          tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1000          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1001          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1002       }
1003       break;
1004
1005    case TGSI_OPCODE_DP2A:
1006       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1007       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1008       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1009       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1010       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1011       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1012       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1013       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
1014       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1015       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1016          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1017       }
1018       break;
1019
1020    case TGSI_OPCODE_FRC:
1021       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1022          src0 = emit_fetch( bld, inst, 0, chan_index );
1023          tmp0 = lp_build_floor(&bld->base, src0);
1024          tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1025          dst0[chan_index] = tmp0;
1026       }
1027       break;
1028
1029    case TGSI_OPCODE_CLAMP:
1030       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1031          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1032          src1 = emit_fetch( bld, inst, 1, chan_index );
1033          src2 = emit_fetch( bld, inst, 2, chan_index );
1034          tmp0 = lp_build_max(&bld->base, tmp0, src1);
1035          tmp0 = lp_build_min(&bld->base, tmp0, src2);
1036          dst0[chan_index] = tmp0;
1037       }
1038       break;
1039
1040    case TGSI_OPCODE_FLR:
1041       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1042          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1043          dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1044       }
1045       break;
1046
1047    case TGSI_OPCODE_ROUND:
1048       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1049          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1050          dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1051       }
1052       break;
1053
1054    case TGSI_OPCODE_EX2: {
1055       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1056       tmp0 = lp_build_exp2( &bld->base, tmp0);
1057       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1058          dst0[chan_index] = tmp0;
1059       }
1060       break;
1061    }
1062
1063    case TGSI_OPCODE_LG2:
1064       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1065       tmp0 = lp_build_log2( &bld->base, tmp0);
1066       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1067          dst0[chan_index] = tmp0;
1068       }
1069       break;
1070
1071    case TGSI_OPCODE_POW:
1072       src0 = emit_fetch( bld, inst, 0, CHAN_X );
1073       src1 = emit_fetch( bld, inst, 1, CHAN_X );
1074       res = lp_build_pow( &bld->base, src0, src1 );
1075       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1076          dst0[chan_index] = res;
1077       }
1078       break;
1079
1080    case TGSI_OPCODE_XPD:
1081       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1082           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1083          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1084          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1085       }
1086       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1087           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1088          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1089          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1090       }
1091       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1092          tmp2 = tmp0;
1093          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1094          tmp5 = tmp3;
1095          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1096          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1097          dst0[CHAN_X] = tmp2;
1098       }
1099       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1100           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1101          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1102          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1103       }
1104       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1105          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1106          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1107          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1108          dst0[CHAN_Y] = tmp3;
1109       }
1110       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1111          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1112          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1113          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1114          dst0[CHAN_Z] = tmp5;
1115       }
1116       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1117          dst0[CHAN_W] = bld->base.one;
1118       }
1119       break;
1120
1121    case TGSI_OPCODE_ABS:
1122       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1123          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1124          dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1125       }
1126       break;
1127
1128    case TGSI_OPCODE_RCC:
1129       /* deprecated? */
1130       assert(0);
1131       return 0;
1132
1133    case TGSI_OPCODE_DPH:
1134       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1135       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1136       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1137       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1138       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1139       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1140       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1141       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1142       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1143       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1144       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1145       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1146       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1147       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1148          dst0[chan_index] = tmp0;
1149       }
1150       break;
1151
1152    case TGSI_OPCODE_COS:
1153       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1154       tmp0 = lp_build_cos( &bld->base, tmp0 );
1155       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1156          dst0[chan_index] = tmp0;
1157       }
1158       break;
1159
1160    case TGSI_OPCODE_DDX:
1161       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1162          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1163       }
1164       break;
1165
1166    case TGSI_OPCODE_DDY:
1167       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1168          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1169       }
1170       break;
1171
1172    case TGSI_OPCODE_KILP:
1173       /* predicated kill */
1174       /* FIXME */
1175       return 0;
1176       break;
1177
1178    case TGSI_OPCODE_KIL:
1179       /* conditional kill */
1180       emit_kil( bld, inst );
1181       break;
1182
1183    case TGSI_OPCODE_PK2H:
1184       return 0;
1185       break;
1186
1187    case TGSI_OPCODE_PK2US:
1188       return 0;
1189       break;
1190
1191    case TGSI_OPCODE_PK4B:
1192       return 0;
1193       break;
1194
1195    case TGSI_OPCODE_PK4UB:
1196       return 0;
1197       break;
1198
1199    case TGSI_OPCODE_RFL:
1200       return 0;
1201       break;
1202
1203    case TGSI_OPCODE_SEQ:
1204       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1205          src0 = emit_fetch( bld, inst, 0, chan_index );
1206          src1 = emit_fetch( bld, inst, 1, chan_index );
1207          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1208          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1209       }
1210       break;
1211
1212    case TGSI_OPCODE_SFL:
1213       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1214          dst0[chan_index] = bld->base.zero;
1215       }
1216       break;
1217
1218    case TGSI_OPCODE_SGT:
1219       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1220          src0 = emit_fetch( bld, inst, 0, chan_index );
1221          src1 = emit_fetch( bld, inst, 1, chan_index );
1222          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1223          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1224       }
1225       break;
1226
1227    case TGSI_OPCODE_SIN:
1228       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1229       tmp0 = lp_build_sin( &bld->base, tmp0 );
1230       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1231          dst0[chan_index] = tmp0;
1232       }
1233       break;
1234
1235    case TGSI_OPCODE_SLE:
1236       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1237          src0 = emit_fetch( bld, inst, 0, chan_index );
1238          src1 = emit_fetch( bld, inst, 1, chan_index );
1239          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
1240          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1241       }
1242       break;
1243
1244    case TGSI_OPCODE_SNE:
1245       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1246          src0 = emit_fetch( bld, inst, 0, chan_index );
1247          src1 = emit_fetch( bld, inst, 1, chan_index );
1248          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1249          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1250       }
1251       break;
1252
1253    case TGSI_OPCODE_STR:
1254       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1255          dst0[chan_index] = bld->base.one;
1256       }
1257       break;
1258
1259    case TGSI_OPCODE_TEX:
1260       emit_tex( bld, inst, FALSE, FALSE, dst0 );
1261       break;
1262
1263    case TGSI_OPCODE_TXD:
1264       /* FIXME */
1265       return 0;
1266       break;
1267
1268    case TGSI_OPCODE_UP2H:
1269       /* deprecated */
1270       assert (0);
1271       return 0;
1272       break;
1273
1274    case TGSI_OPCODE_UP2US:
1275       /* deprecated */
1276       assert(0);
1277       return 0;
1278       break;
1279
1280    case TGSI_OPCODE_UP4B:
1281       /* deprecated */
1282       assert(0);
1283       return 0;
1284       break;
1285
1286    case TGSI_OPCODE_UP4UB:
1287       /* deprecated */
1288       assert(0);
1289       return 0;
1290       break;
1291
1292    case TGSI_OPCODE_X2D:
1293       /* deprecated? */
1294       assert(0);
1295       return 0;
1296       break;
1297
1298    case TGSI_OPCODE_ARA:
1299       /* deprecated */
1300       assert(0);
1301       return 0;
1302       break;
1303
1304 #if 0
1305    case TGSI_OPCODE_ARR:
1306       /* FIXME */
1307       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1308          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1309          emit_rnd( bld, 0, 0 );
1310          emit_f2it( bld, 0 );
1311          dst0[chan_index] = tmp0;
1312       }
1313       break;
1314 #endif
1315
1316    case TGSI_OPCODE_BRA:
1317       /* deprecated */
1318       assert(0);
1319       return 0;
1320       break;
1321
1322    case TGSI_OPCODE_CAL:
1323       /* FIXME */
1324       return 0;
1325       break;
1326
1327    case TGSI_OPCODE_RET:
1328       /* FIXME */
1329       return 0;
1330       break;
1331
1332    case TGSI_OPCODE_END:
1333       break;
1334
1335    case TGSI_OPCODE_SSG:
1336    /* TGSI_OPCODE_SGN */
1337       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1338          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1339          dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1340       }
1341       break;
1342
1343    case TGSI_OPCODE_CMP:
1344       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1345          src0 = emit_fetch( bld, inst, 0, chan_index );
1346          src1 = emit_fetch( bld, inst, 1, chan_index );
1347          src2 = emit_fetch( bld, inst, 2, chan_index );
1348          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1349          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1350       }
1351       break;
1352
1353    case TGSI_OPCODE_SCS:
1354       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1355          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1356          dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1357       }
1358       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1359          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1360          dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1361       }
1362       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1363          dst0[CHAN_Z] = bld->base.zero;
1364       }
1365       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1366          dst0[CHAN_W] = bld->base.one;
1367       }
1368       break;
1369
1370    case TGSI_OPCODE_TXB:
1371       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1372       break;
1373
1374    case TGSI_OPCODE_NRM:
1375       /* fall-through */
1376    case TGSI_OPCODE_NRM4:
1377       /* 3 or 4-component normalization */
1378       {
1379          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1380
1381          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1382              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1383              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1384              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1385
1386             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1387
1388             /* xmm4 = src.x */
1389             /* xmm0 = src.x * src.x */
1390             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1391             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1392                tmp4 = tmp0;
1393             }
1394             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1395
1396             /* xmm5 = src.y */
1397             /* xmm0 = xmm0 + src.y * src.y */
1398             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1399             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1400                tmp5 = tmp1;
1401             }
1402             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1403             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1404
1405             /* xmm6 = src.z */
1406             /* xmm0 = xmm0 + src.z * src.z */
1407             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1408             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1409                tmp6 = tmp1;
1410             }
1411             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1412             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1413
1414             if (dims == 4) {
1415                /* xmm7 = src.w */
1416                /* xmm0 = xmm0 + src.w * src.w */
1417                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1418                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1419                   tmp7 = tmp1;
1420                }
1421                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1422                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1423             }
1424
1425             /* xmm1 = 1 / sqrt(xmm0) */
1426             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1427
1428             /* dst.x = xmm1 * src.x */
1429             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1430                dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1431             }
1432
1433             /* dst.y = xmm1 * src.y */
1434             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1435                dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1436             }
1437
1438             /* dst.z = xmm1 * src.z */
1439             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1440                dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1441             }
1442
1443             /* dst.w = xmm1 * src.w */
1444             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1445                dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1446             }
1447          }
1448
1449          /* dst.w = 1.0 */
1450          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1451             dst0[CHAN_W] = bld->base.one;
1452          }
1453       }
1454       break;
1455
1456    case TGSI_OPCODE_DIV:
1457       /* deprecated */
1458       assert( 0 );
1459       return 0;
1460       break;
1461
1462    case TGSI_OPCODE_DP2:
1463       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1464       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1465       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1466       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1467       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1468       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1469       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1470       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1471          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1472       }
1473       break;
1474
1475    case TGSI_OPCODE_TXL:
1476       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1477       break;
1478
1479    case TGSI_OPCODE_TXP:
1480       emit_tex( bld, inst, FALSE, TRUE, dst0 );
1481       break;
1482
1483    case TGSI_OPCODE_BRK:
1484       lp_exec_break(&bld->exec_mask);
1485       break;
1486
1487    case TGSI_OPCODE_IF:
1488       tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1489       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1490                           tmp0, bld->base.zero);
1491       lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1492       break;
1493
1494    case TGSI_OPCODE_BGNFOR:
1495       /* deprecated */
1496       assert(0);
1497       return 0;
1498       break;
1499
1500    case TGSI_OPCODE_BGNLOOP:
1501       lp_exec_bgnloop(&bld->exec_mask);
1502       break;
1503
1504    case TGSI_OPCODE_REP:
1505       /* deprecated */
1506       assert(0);
1507       return 0;
1508       break;
1509
1510    case TGSI_OPCODE_ELSE:
1511       lp_exec_mask_cond_invert(&bld->exec_mask);
1512       break;
1513
1514    case TGSI_OPCODE_ENDIF:
1515       lp_exec_mask_cond_pop(&bld->exec_mask);
1516       break;
1517
1518    case TGSI_OPCODE_ENDFOR:
1519       /* deprecated */
1520       assert(0);
1521       return 0;
1522       break;
1523
1524    case TGSI_OPCODE_ENDLOOP:
1525       lp_exec_endloop(&bld->exec_mask);
1526       break;
1527
1528    case TGSI_OPCODE_ENDREP:
1529       /* deprecated */
1530       assert(0);
1531       return 0;
1532       break;
1533
1534    case TGSI_OPCODE_PUSHA:
1535       /* deprecated? */
1536       assert(0);
1537       return 0;
1538       break;
1539
1540    case TGSI_OPCODE_POPA:
1541       /* deprecated? */
1542       assert(0);
1543       return 0;
1544       break;
1545
1546    case TGSI_OPCODE_CEIL:
1547       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1548          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1549          dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1550       }
1551       break;
1552
1553    case TGSI_OPCODE_I2F:
1554       /* deprecated? */
1555       assert(0);
1556       return 0;
1557       break;
1558
1559    case TGSI_OPCODE_NOT:
1560       /* deprecated? */
1561       assert(0);
1562       return 0;
1563       break;
1564
1565    case TGSI_OPCODE_TRUNC:
1566       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1567          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1568          dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1569       }
1570       break;
1571
1572    case TGSI_OPCODE_SHL:
1573       /* deprecated? */
1574       assert(0);
1575       return 0;
1576       break;
1577
1578    case TGSI_OPCODE_ISHR:
1579       /* deprecated? */
1580       assert(0);
1581       return 0;
1582       break;
1583
1584    case TGSI_OPCODE_AND:
1585       /* deprecated? */
1586       assert(0);
1587       return 0;
1588       break;
1589
1590    case TGSI_OPCODE_OR:
1591       /* deprecated? */
1592       assert(0);
1593       return 0;
1594       break;
1595
1596    case TGSI_OPCODE_MOD:
1597       /* deprecated? */
1598       assert(0);
1599       return 0;
1600       break;
1601
1602    case TGSI_OPCODE_XOR:
1603       /* deprecated? */
1604       assert(0);
1605       return 0;
1606       break;
1607
1608    case TGSI_OPCODE_SAD:
1609       /* deprecated? */
1610       assert(0);
1611       return 0;
1612       break;
1613
1614    case TGSI_OPCODE_TXF:
1615       /* deprecated? */
1616       assert(0);
1617       return 0;
1618       break;
1619
1620    case TGSI_OPCODE_TXQ:
1621       /* deprecated? */
1622       assert(0);
1623       return 0;
1624       break;
1625
1626    case TGSI_OPCODE_CONT:
1627       lp_exec_continue(&bld->exec_mask);
1628       break;
1629
1630    case TGSI_OPCODE_EMIT:
1631       return 0;
1632       break;
1633
1634    case TGSI_OPCODE_ENDPRIM:
1635       return 0;
1636       break;
1637
1638    case TGSI_OPCODE_NOP:
1639       break;
1640
1641    default:
1642       return 0;
1643    }
1644
1645    if(info->num_dst) {
1646       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1647          emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
1648       }
1649    }
1650
1651    return 1;
1652 }
1653
1654
1655 void
1656 lp_build_tgsi_soa(LLVMBuilderRef builder,
1657                   const struct tgsi_token *tokens,
1658                   struct lp_type type,
1659                   struct lp_build_mask_context *mask,
1660                   LLVMValueRef consts_ptr,
1661                   const LLVMValueRef *pos,
1662                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1663                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1664                   struct lp_build_sampler_soa *sampler)
1665 {
1666    struct lp_build_tgsi_soa_context bld;
1667    struct tgsi_parse_context parse;
1668    uint num_immediates = 0;
1669    unsigned i;
1670
1671    /* Setup build context */
1672    memset(&bld, 0, sizeof bld);
1673    lp_build_context_init(&bld.base, builder, type);
1674    bld.mask = mask;
1675    bld.pos = pos;
1676    bld.inputs = inputs;
1677    bld.outputs = outputs;
1678    bld.consts_ptr = consts_ptr;
1679    bld.sampler = sampler;
1680
1681    lp_exec_mask_init(&bld.exec_mask, &bld.base);
1682
1683    tgsi_parse_init( &parse, tokens );
1684
1685    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1686       tgsi_parse_token( &parse );
1687
1688       switch( parse.FullToken.Token.Type ) {
1689       case TGSI_TOKEN_TYPE_DECLARATION:
1690          /* Inputs already interpolated */
1691          {
1692             if (!emit_declaration( &bld, &parse.FullToken.FullDeclaration ))
1693                _debug_printf("warning: failed to define LLVM variable\n");
1694          }
1695          break;
1696
1697       case TGSI_TOKEN_TYPE_INSTRUCTION:
1698          {
1699             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1700             const struct tgsi_opcode_info *info = tgsi_get_opcode_info(opcode);
1701             if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, info ))
1702                _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1703                              info ? info->mnemonic : "<invalid>");
1704          }
1705
1706          break;
1707
1708       case TGSI_TOKEN_TYPE_IMMEDIATE:
1709          /* simply copy the immediate values into the next immediates[] slot */
1710          {
1711             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1712             assert(size <= 4);
1713             assert(num_immediates < LP_MAX_IMMEDIATES);
1714             for( i = 0; i < size; ++i )
1715                bld.immediates[num_immediates][i] =
1716                   lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
1717             for( i = size; i < 4; ++i )
1718                bld.immediates[num_immediates][i] = bld.base.undef;
1719             num_immediates++;
1720          }
1721          break;
1722
1723       case TGSI_TOKEN_TYPE_PROPERTY:
1724          break;
1725
1726       default:
1727          assert( 0 );
1728       }
1729    }
1730    if (0) {
1731       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1732       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1733       debug_printf("11111111111111111111111111111 \n");
1734       tgsi_dump(tokens, 0);
1735       LLVMDumpValue(function);
1736       debug_printf("2222222222222222222222222222 \n");
1737    }
1738    tgsi_parse_free( &parse );
1739 }
1740