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    LLVMBasicBlockRef current_block =
 655       LLVMGetInsertBlock(bld->base.builder);
 656    LLVMBasicBlockRef first_block =
 657       LLVMGetEntryBasicBlock(
 658          LLVMGetBasicBlockParent(current_block));
 659    LLVMValueRef first_inst =
 660       LLVMGetFirstInstruction(first_block);
 661
 662    /* we want alloca's to be the first instruction
 663     * in the function so we need to rewind the builder
 664     * to the very beginning */
 665    LLVMPositionBuilderBefore(bld->base.builder,
 666                              first_inst);
 667
 668    for (idx = first; idx <= last; ++idx) {
 669       switch (decl->Declaration.File) {
 670       case TGSI_FILE_TEMPORARY:
 671          for (i = 0; i < NUM_CHANNELS; i++)
 672             bld->temps[idx][i] = lp_build_alloca(&bld->base);
 673          break;
 674
 675       case TGSI_FILE_OUTPUT:
 676          for (i = 0; i < NUM_CHANNELS; i++)
 677             bld->outputs[idx][i] = lp_build_alloca(&bld->base);
 678          break;
 679
 680       default:
 681          /* don't need to declare other vars */
 682          break;
 683       }
 684    }
 685
 686    LLVMPositionBuilderAtEnd(bld->base.builder,
 687                             current_block);
 688    return TRUE;
 689 }
 690
 691 static int
 692 emit_instruction(
 693    struct lp_build_tgsi_soa_context *bld,
 694    const struct tgsi_full_instruction *inst,
 695    const struct tgsi_opcode_info *info)
 696 {
 697    unsigned chan_index;
 698    LLVMValueRef src0, src1, src2;
 699    LLVMValueRef tmp0, tmp1, tmp2;
 700    LLVMValueRef tmp3 = NULL;
 701    LLVMValueRef tmp4 = NULL;
 702    LLVMValueRef tmp5 = NULL;
 703    LLVMValueRef tmp6 = NULL;
 704    LLVMValueRef tmp7 = NULL;
 705    LLVMValueRef res;
 706    LLVMValueRef dst0[NUM_CHANNELS];
 707
 708    /* we can't handle indirect addressing into temp register file yet */
 709    if (indirect_temp_reference(inst))
 710       return FALSE;
 711
 712    /*
 713     * Stores and write masks are handled in a general fashion after the long
 714     * instruction opcode switch statement.
 715     *
 716     * Although not stricitly necessary, we avoid generating instructions for
 717     * channels which won't be stored, in cases where's that easy. For some
 718     * complex instructions, like texture sampling, it is more convenient to
 719     * assume a full writemask and then let LLVM optimization passes eliminate
 720     * redundant code.
 721     */
 722
 723    assert(info->num_dst <= 1);
 724    if(info->num_dst) {
 725       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 726          dst0[chan_index] = bld->base.undef;
 727       }
 728    }
 729
 730    switch (inst->Instruction.Opcode) {
 731 #if 0
 732    case TGSI_OPCODE_ARL:
 733       /* FIXME */
 734       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 735          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 736          emit_flr(bld, 0, 0);
 737          emit_f2it( bld, 0 );
 738          dst0[chan_index] = tmp0;
 739       }
 740       break;
 741 #endif
 742
 743    case TGSI_OPCODE_MOV:
 744       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 745          dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
 746       }
 747       break;
 748
 749    case TGSI_OPCODE_LIT:
 750       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 751          dst0[CHAN_X] = bld->base.one;
 752       }
 753       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 754          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 755          dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
 756       }
 757       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 758          /* XMM[1] = SrcReg[0].yyyy */
 759          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 760          /* XMM[1] = max(XMM[1], 0) */
 761          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 762          /* XMM[2] = SrcReg[0].wwww */
 763          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 764          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 765          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 766          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 767          dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 768       }
 769       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 770          dst0[CHAN_W] = bld->base.one;
 771       }
 772       break;
 773
 774    case TGSI_OPCODE_RCP:
 775    /* TGSI_OPCODE_RECIP */
 776       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 777       res = lp_build_rcp(&bld->base, src0);
 778       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 779          dst0[chan_index] = res;
 780       }
 781       break;
 782
 783    case TGSI_OPCODE_RSQ:
 784    /* TGSI_OPCODE_RECIPSQRT */
 785       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 786       src0 = lp_build_abs(&bld->base, src0);
 787       res = lp_build_rsqrt(&bld->base, src0);
 788       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 789          dst0[chan_index] = res;
 790       }
 791       break;
 792
 793    case TGSI_OPCODE_EXP:
 794       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 795           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 796           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 797          LLVMValueRef *p_exp2_int_part = NULL;
 798          LLVMValueRef *p_frac_part = NULL;
 799          LLVMValueRef *p_exp2 = NULL;
 800
 801          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 802
 803          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 804             p_exp2_int_part = &tmp0;
 805          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 806             p_frac_part = &tmp1;
 807          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 808             p_exp2 = &tmp2;
 809
 810          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 811
 812          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 813             dst0[CHAN_X] = tmp0;
 814          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 815             dst0[CHAN_Y] = tmp1;
 816          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 817             dst0[CHAN_Z] = tmp2;
 818       }
 819       /* dst.w = 1.0 */
 820       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 821          dst0[CHAN_W] = bld->base.one;
 822       }
 823       break;
 824
 825    case TGSI_OPCODE_LOG:
 826       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 827           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 828           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 829          LLVMValueRef *p_floor_log2 = NULL;
 830          LLVMValueRef *p_exp = NULL;
 831          LLVMValueRef *p_log2 = NULL;
 832
 833          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 834          src0 = lp_build_abs( &bld->base, src0 );
 835
 836          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 837             p_floor_log2 = &tmp0;
 838          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 839             p_exp = &tmp1;
 840          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 841             p_log2 = &tmp2;
 842
 843          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 844
 845          /* dst.x = floor(lg2(abs(src.x))) */
 846          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 847             dst0[CHAN_X] = tmp0;
 848          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 849          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 850             dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
 851          }
 852          /* dst.z = lg2(abs(src.x)) */
 853          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 854             dst0[CHAN_Z] = tmp2;
 855       }
 856       /* dst.w = 1.0 */
 857       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 858          dst0[CHAN_W] = bld->base.one;
 859       }
 860       break;
 861
 862    case TGSI_OPCODE_MUL:
 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_mul(&bld->base, src0, src1);
 867       }
 868       break;
 869
 870    case TGSI_OPCODE_ADD:
 871       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 872          src0 = emit_fetch( bld, inst, 0, chan_index );
 873          src1 = emit_fetch( bld, inst, 1, chan_index );
 874          dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
 875       }
 876       break;
 877
 878    case TGSI_OPCODE_DP3:
 879    /* TGSI_OPCODE_DOT3 */
 880       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 881       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 882       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 883       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 884       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 885       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 886       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 887       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 888       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 889       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 890       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 891       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 892          dst0[chan_index] = tmp0;
 893       }
 894       break;
 895
 896    case TGSI_OPCODE_DP4:
 897    /* TGSI_OPCODE_DOT4 */
 898       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 899       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 900       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 901       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 902       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 903       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 904       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 905       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 906       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 907       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 908       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 909       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
 910       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
 911       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 912       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 913       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 914          dst0[chan_index] = tmp0;
 915       }
 916       break;
 917
 918    case TGSI_OPCODE_DST:
 919       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
 920          dst0[CHAN_X] = bld->base.one;
 921       }
 922       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
 923          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
 924          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
 925          dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
 926       }
 927       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
 928          dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
 929       }
 930       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
 931          dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
 932       }
 933       break;
 934
 935    case TGSI_OPCODE_MIN:
 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_min( &bld->base, src0, src1 );
 940       }
 941       break;
 942
 943    case TGSI_OPCODE_MAX:
 944       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 945          src0 = emit_fetch( bld, inst, 0, chan_index );
 946          src1 = emit_fetch( bld, inst, 1, chan_index );
 947          dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
 948       }
 949       break;
 950
 951    case TGSI_OPCODE_SLT:
 952    /* TGSI_OPCODE_SETLT */
 953       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 954          src0 = emit_fetch( bld, inst, 0, chan_index );
 955          src1 = emit_fetch( bld, inst, 1, chan_index );
 956          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
 957          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 958       }
 959       break;
 960
 961    case TGSI_OPCODE_SGE:
 962    /* TGSI_OPCODE_SETGE */
 963       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 964          src0 = emit_fetch( bld, inst, 0, chan_index );
 965          src1 = emit_fetch( bld, inst, 1, chan_index );
 966          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
 967          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 968       }
 969       break;
 970
 971    case TGSI_OPCODE_MAD:
 972    /* TGSI_OPCODE_MADD */
 973       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 974          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 975          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 976          tmp2 = emit_fetch( bld, inst, 2, chan_index );
 977          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 978          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
 979          dst0[chan_index] = tmp0;
 980       }
 981       break;
 982
 983    case TGSI_OPCODE_SUB:
 984       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 985          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 986          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 987          dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
 988       }
 989       break;
 990
 991    case TGSI_OPCODE_LRP:
 992       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 993          src0 = emit_fetch( bld, inst, 0, chan_index );
 994          src1 = emit_fetch( bld, inst, 1, chan_index );
 995          src2 = emit_fetch( bld, inst, 2, chan_index );
 996          tmp0 = lp_build_sub( &bld->base, src1, src2 );
 997          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
 998          dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
 999       }
1000       break;
1001
1002    case TGSI_OPCODE_CND:
1003       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1004          src0 = emit_fetch( bld, inst, 0, chan_index );
1005          src1 = emit_fetch( bld, inst, 1, chan_index );
1006          src2 = emit_fetch( bld, inst, 2, chan_index );
1007          tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1008          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1009          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1010       }
1011       break;
1012
1013    case TGSI_OPCODE_DP2A:
1014       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1015       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1016       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1017       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1018       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1019       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1020       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1021       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
1022       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1023       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1024          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1025       }
1026       break;
1027
1028    case TGSI_OPCODE_FRC:
1029       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1030          src0 = emit_fetch( bld, inst, 0, chan_index );
1031          tmp0 = lp_build_floor(&bld->base, src0);
1032          tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1033          dst0[chan_index] = tmp0;
1034       }
1035       break;
1036
1037    case TGSI_OPCODE_CLAMP:
1038       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1039          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1040          src1 = emit_fetch( bld, inst, 1, chan_index );
1041          src2 = emit_fetch( bld, inst, 2, chan_index );
1042          tmp0 = lp_build_max(&bld->base, tmp0, src1);
1043          tmp0 = lp_build_min(&bld->base, tmp0, src2);
1044          dst0[chan_index] = tmp0;
1045       }
1046       break;
1047
1048    case TGSI_OPCODE_FLR:
1049       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1050          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1051          dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1052       }
1053       break;
1054
1055    case TGSI_OPCODE_ROUND:
1056       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1057          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1058          dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1059       }
1060       break;
1061
1062    case TGSI_OPCODE_EX2: {
1063       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1064       tmp0 = lp_build_exp2( &bld->base, tmp0);
1065       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1066          dst0[chan_index] = tmp0;
1067       }
1068       break;
1069    }
1070
1071    case TGSI_OPCODE_LG2:
1072       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1073       tmp0 = lp_build_log2( &bld->base, tmp0);
1074       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1075          dst0[chan_index] = tmp0;
1076       }
1077       break;
1078
1079    case TGSI_OPCODE_POW:
1080       src0 = emit_fetch( bld, inst, 0, CHAN_X );
1081       src1 = emit_fetch( bld, inst, 1, CHAN_X );
1082       res = lp_build_pow( &bld->base, src0, src1 );
1083       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1084          dst0[chan_index] = res;
1085       }
1086       break;
1087
1088    case TGSI_OPCODE_XPD:
1089       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1090           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1091          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1092          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1093       }
1094       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1095           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1096          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1097          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1098       }
1099       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1100          tmp2 = tmp0;
1101          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1102          tmp5 = tmp3;
1103          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1104          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1105          dst0[CHAN_X] = tmp2;
1106       }
1107       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1108           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1109          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1110          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1111       }
1112       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1113          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1114          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1115          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1116          dst0[CHAN_Y] = tmp3;
1117       }
1118       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1119          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1120          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1121          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1122          dst0[CHAN_Z] = tmp5;
1123       }
1124       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1125          dst0[CHAN_W] = bld->base.one;
1126       }
1127       break;
1128
1129    case TGSI_OPCODE_ABS:
1130       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1131          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1132          dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1133       }
1134       break;
1135
1136    case TGSI_OPCODE_RCC:
1137       /* deprecated? */
1138       assert(0);
1139       return 0;
1140
1141    case TGSI_OPCODE_DPH:
1142       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1143       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1144       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1145       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1146       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1147       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1148       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1149       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1150       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1151       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1152       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1153       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1154       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1155       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1156          dst0[chan_index] = tmp0;
1157       }
1158       break;
1159
1160    case TGSI_OPCODE_COS:
1161       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1162       tmp0 = lp_build_cos( &bld->base, tmp0 );
1163       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1164          dst0[chan_index] = tmp0;
1165       }
1166       break;
1167
1168    case TGSI_OPCODE_DDX:
1169       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1170          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1171       }
1172       break;
1173
1174    case TGSI_OPCODE_DDY:
1175       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1176          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1177       }
1178       break;
1179
1180    case TGSI_OPCODE_KILP:
1181       /* predicated kill */
1182       /* FIXME */
1183       return 0;
1184       break;
1185
1186    case TGSI_OPCODE_KIL:
1187       /* conditional kill */
1188       emit_kil( bld, inst );
1189       break;
1190
1191    case TGSI_OPCODE_PK2H:
1192       return 0;
1193       break;
1194
1195    case TGSI_OPCODE_PK2US:
1196       return 0;
1197       break;
1198
1199    case TGSI_OPCODE_PK4B:
1200       return 0;
1201       break;
1202
1203    case TGSI_OPCODE_PK4UB:
1204       return 0;
1205       break;
1206
1207    case TGSI_OPCODE_RFL:
1208       return 0;
1209       break;
1210
1211    case TGSI_OPCODE_SEQ:
1212       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1213          src0 = emit_fetch( bld, inst, 0, chan_index );
1214          src1 = emit_fetch( bld, inst, 1, chan_index );
1215          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1216          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1217       }
1218       break;
1219
1220    case TGSI_OPCODE_SFL:
1221       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1222          dst0[chan_index] = bld->base.zero;
1223       }
1224       break;
1225
1226    case TGSI_OPCODE_SGT:
1227       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1228          src0 = emit_fetch( bld, inst, 0, chan_index );
1229          src1 = emit_fetch( bld, inst, 1, chan_index );
1230          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1231          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1232       }
1233       break;
1234
1235    case TGSI_OPCODE_SIN:
1236       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1237       tmp0 = lp_build_sin( &bld->base, tmp0 );
1238       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1239          dst0[chan_index] = tmp0;
1240       }
1241       break;
1242
1243    case TGSI_OPCODE_SLE:
1244       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1245          src0 = emit_fetch( bld, inst, 0, chan_index );
1246          src1 = emit_fetch( bld, inst, 1, chan_index );
1247          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
1248          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1249       }
1250       break;
1251
1252    case TGSI_OPCODE_SNE:
1253       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1254          src0 = emit_fetch( bld, inst, 0, chan_index );
1255          src1 = emit_fetch( bld, inst, 1, chan_index );
1256          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1257          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1258       }
1259       break;
1260
1261    case TGSI_OPCODE_STR:
1262       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1263          dst0[chan_index] = bld->base.one;
1264       }
1265       break;
1266
1267    case TGSI_OPCODE_TEX:
1268       emit_tex( bld, inst, FALSE, FALSE, dst0 );
1269       break;
1270
1271    case TGSI_OPCODE_TXD:
1272       /* FIXME */
1273       return 0;
1274       break;
1275
1276    case TGSI_OPCODE_UP2H:
1277       /* deprecated */
1278       assert (0);
1279       return 0;
1280       break;
1281
1282    case TGSI_OPCODE_UP2US:
1283       /* deprecated */
1284       assert(0);
1285       return 0;
1286       break;
1287
1288    case TGSI_OPCODE_UP4B:
1289       /* deprecated */
1290       assert(0);
1291       return 0;
1292       break;
1293
1294    case TGSI_OPCODE_UP4UB:
1295       /* deprecated */
1296       assert(0);
1297       return 0;
1298       break;
1299
1300    case TGSI_OPCODE_X2D:
1301       /* deprecated? */
1302       assert(0);
1303       return 0;
1304       break;
1305
1306    case TGSI_OPCODE_ARA:
1307       /* deprecated */
1308       assert(0);
1309       return 0;
1310       break;
1311
1312 #if 0
1313    case TGSI_OPCODE_ARR:
1314       /* FIXME */
1315       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1316          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1317          emit_rnd( bld, 0, 0 );
1318          emit_f2it( bld, 0 );
1319          dst0[chan_index] = tmp0;
1320       }
1321       break;
1322 #endif
1323
1324    case TGSI_OPCODE_BRA:
1325       /* deprecated */
1326       assert(0);
1327       return 0;
1328       break;
1329
1330    case TGSI_OPCODE_CAL:
1331       /* FIXME */
1332       return 0;
1333       break;
1334
1335    case TGSI_OPCODE_RET:
1336       /* FIXME */
1337       return 0;
1338       break;
1339
1340    case TGSI_OPCODE_END:
1341       break;
1342
1343    case TGSI_OPCODE_SSG:
1344    /* TGSI_OPCODE_SGN */
1345       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1346          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1347          dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1348       }
1349       break;
1350
1351    case TGSI_OPCODE_CMP:
1352       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1353          src0 = emit_fetch( bld, inst, 0, chan_index );
1354          src1 = emit_fetch( bld, inst, 1, chan_index );
1355          src2 = emit_fetch( bld, inst, 2, chan_index );
1356          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1357          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1358       }
1359       break;
1360
1361    case TGSI_OPCODE_SCS:
1362       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1363          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1364          dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1365       }
1366       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1367          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1368          dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1369       }
1370       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1371          dst0[CHAN_Z] = bld->base.zero;
1372       }
1373       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1374          dst0[CHAN_W] = bld->base.one;
1375       }
1376       break;
1377
1378    case TGSI_OPCODE_TXB:
1379       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1380       break;
1381
1382    case TGSI_OPCODE_NRM:
1383       /* fall-through */
1384    case TGSI_OPCODE_NRM4:
1385       /* 3 or 4-component normalization */
1386       {
1387          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1388
1389          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1390              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1391              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1392              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1393
1394             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1395
1396             /* xmm4 = src.x */
1397             /* xmm0 = src.x * src.x */
1398             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1399             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1400                tmp4 = tmp0;
1401             }
1402             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1403
1404             /* xmm5 = src.y */
1405             /* xmm0 = xmm0 + src.y * src.y */
1406             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1407             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1408                tmp5 = tmp1;
1409             }
1410             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1411             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1412
1413             /* xmm6 = src.z */
1414             /* xmm0 = xmm0 + src.z * src.z */
1415             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1416             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1417                tmp6 = tmp1;
1418             }
1419             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1420             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1421
1422             if (dims == 4) {
1423                /* xmm7 = src.w */
1424                /* xmm0 = xmm0 + src.w * src.w */
1425                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1426                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1427                   tmp7 = tmp1;
1428                }
1429                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1430                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1431             }
1432
1433             /* xmm1 = 1 / sqrt(xmm0) */
1434             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1435
1436             /* dst.x = xmm1 * src.x */
1437             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1438                dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1439             }
1440
1441             /* dst.y = xmm1 * src.y */
1442             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1443                dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1444             }
1445
1446             /* dst.z = xmm1 * src.z */
1447             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1448                dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1449             }
1450
1451             /* dst.w = xmm1 * src.w */
1452             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1453                dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1454             }
1455          }
1456
1457          /* dst.w = 1.0 */
1458          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1459             dst0[CHAN_W] = bld->base.one;
1460          }
1461       }
1462       break;
1463
1464    case TGSI_OPCODE_DIV:
1465       /* deprecated */
1466       assert( 0 );
1467       return 0;
1468       break;
1469
1470    case TGSI_OPCODE_DP2:
1471       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1472       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1473       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1474       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1475       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1476       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1477       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1478       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1479          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1480       }
1481       break;
1482
1483    case TGSI_OPCODE_TXL:
1484       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1485       break;
1486
1487    case TGSI_OPCODE_TXP:
1488       emit_tex( bld, inst, FALSE, TRUE, dst0 );
1489       break;
1490
1491    case TGSI_OPCODE_BRK:
1492       lp_exec_break(&bld->exec_mask);
1493       break;
1494
1495    case TGSI_OPCODE_IF:
1496       tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1497       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1498                           tmp0, bld->base.zero);
1499       lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1500       break;
1501
1502    case TGSI_OPCODE_BGNFOR:
1503       /* deprecated */
1504       assert(0);
1505       return 0;
1506       break;
1507
1508    case TGSI_OPCODE_BGNLOOP:
1509       lp_exec_bgnloop(&bld->exec_mask);
1510       break;
1511
1512    case TGSI_OPCODE_REP:
1513       /* deprecated */
1514       assert(0);
1515       return 0;
1516       break;
1517
1518    case TGSI_OPCODE_ELSE:
1519       lp_exec_mask_cond_invert(&bld->exec_mask);
1520       break;
1521
1522    case TGSI_OPCODE_ENDIF:
1523       lp_exec_mask_cond_pop(&bld->exec_mask);
1524       break;
1525
1526    case TGSI_OPCODE_ENDFOR:
1527       /* deprecated */
1528       assert(0);
1529       return 0;
1530       break;
1531
1532    case TGSI_OPCODE_ENDLOOP:
1533       lp_exec_endloop(&bld->exec_mask);
1534       break;
1535
1536    case TGSI_OPCODE_ENDREP:
1537       /* deprecated */
1538       assert(0);
1539       return 0;
1540       break;
1541
1542    case TGSI_OPCODE_PUSHA:
1543       /* deprecated? */
1544       assert(0);
1545       return 0;
1546       break;
1547
1548    case TGSI_OPCODE_POPA:
1549       /* deprecated? */
1550       assert(0);
1551       return 0;
1552       break;
1553
1554    case TGSI_OPCODE_CEIL:
1555       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1556          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1557          dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1558       }
1559       break;
1560
1561    case TGSI_OPCODE_I2F:
1562       /* deprecated? */
1563       assert(0);
1564       return 0;
1565       break;
1566
1567    case TGSI_OPCODE_NOT:
1568       /* deprecated? */
1569       assert(0);
1570       return 0;
1571       break;
1572
1573    case TGSI_OPCODE_TRUNC:
1574       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1575          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1576          dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1577       }
1578       break;
1579
1580    case TGSI_OPCODE_SHL:
1581       /* deprecated? */
1582       assert(0);
1583       return 0;
1584       break;
1585
1586    case TGSI_OPCODE_ISHR:
1587       /* deprecated? */
1588       assert(0);
1589       return 0;
1590       break;
1591
1592    case TGSI_OPCODE_AND:
1593       /* deprecated? */
1594       assert(0);
1595       return 0;
1596       break;
1597
1598    case TGSI_OPCODE_OR:
1599       /* deprecated? */
1600       assert(0);
1601       return 0;
1602       break;
1603
1604    case TGSI_OPCODE_MOD:
1605       /* deprecated? */
1606       assert(0);
1607       return 0;
1608       break;
1609
1610    case TGSI_OPCODE_XOR:
1611       /* deprecated? */
1612       assert(0);
1613       return 0;
1614       break;
1615
1616    case TGSI_OPCODE_SAD:
1617       /* deprecated? */
1618       assert(0);
1619       return 0;
1620       break;
1621
1622    case TGSI_OPCODE_TXF:
1623       /* deprecated? */
1624       assert(0);
1625       return 0;
1626       break;
1627
1628    case TGSI_OPCODE_TXQ:
1629       /* deprecated? */
1630       assert(0);
1631       return 0;
1632       break;
1633
1634    case TGSI_OPCODE_CONT:
1635       lp_exec_continue(&bld->exec_mask);
1636       break;
1637
1638    case TGSI_OPCODE_EMIT:
1639       return 0;
1640       break;
1641
1642    case TGSI_OPCODE_ENDPRIM:
1643       return 0;
1644       break;
1645
1646    case TGSI_OPCODE_NOP:
1647       break;
1648
1649    default:
1650       return 0;
1651    }
1652
1653    if(info->num_dst) {
1654       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1655          emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
1656       }
1657    }
1658
1659    return 1;
1660 }
1661
1662
1663 void
1664 lp_build_tgsi_soa(LLVMBuilderRef builder,
1665                   const struct tgsi_token *tokens,
1666                   struct lp_type type,
1667                   struct lp_build_mask_context *mask,
1668                   LLVMValueRef consts_ptr,
1669                   const LLVMValueRef *pos,
1670                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1671                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1672                   struct lp_build_sampler_soa *sampler)
1673 {
1674    struct lp_build_tgsi_soa_context bld;
1675    struct tgsi_parse_context parse;
1676    uint num_immediates = 0;
1677    unsigned i;
1678
1679    /* Setup build context */
1680    memset(&bld, 0, sizeof bld);
1681    lp_build_context_init(&bld.base, builder, type);
1682    bld.mask = mask;
1683    bld.pos = pos;
1684    bld.inputs = inputs;
1685    bld.outputs = outputs;
1686    bld.consts_ptr = consts_ptr;
1687    bld.sampler = sampler;
1688
1689    lp_exec_mask_init(&bld.exec_mask, &bld.base);
1690
1691    tgsi_parse_init( &parse, tokens );
1692
1693    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1694       tgsi_parse_token( &parse );
1695
1696       switch( parse.FullToken.Token.Type ) {
1697       case TGSI_TOKEN_TYPE_DECLARATION:
1698          /* Inputs already interpolated */
1699          {
1700             if (!emit_declaration( &bld, &parse.FullToken.FullDeclaration ))
1701                _debug_printf("warning: failed to define LLVM variable\n");
1702          }
1703          break;
1704
1705       case TGSI_TOKEN_TYPE_INSTRUCTION:
1706          {
1707             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1708             const struct tgsi_opcode_info *info = tgsi_get_opcode_info(opcode);
1709             if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, info ))
1710                _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1711                              info ? info->mnemonic : "<invalid>");
1712          }
1713
1714          break;
1715
1716       case TGSI_TOKEN_TYPE_IMMEDIATE:
1717          /* simply copy the immediate values into the next immediates[] slot */
1718          {
1719             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1720             assert(size <= 4);
1721             assert(num_immediates < LP_MAX_IMMEDIATES);
1722             for( i = 0; i < size; ++i )
1723                bld.immediates[num_immediates][i] =
1724                   lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
1725             for( i = size; i < 4; ++i )
1726                bld.immediates[num_immediates][i] = bld.base.undef;
1727             num_immediates++;
1728          }
1729          break;
1730
1731       case TGSI_TOKEN_TYPE_PROPERTY:
1732          break;
1733
1734       default:
1735          assert( 0 );
1736       }
1737    }
1738    if (0) {
1739       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1740       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1741       debug_printf("11111111111111111111111111111 \n");
1742       tgsi_dump(tokens, 0);
1743       LLVMDumpValue(function);
1744       debug_printf("2222222222222222222222222222 \n");
1745    }
1746    tgsi_parse_free( &parse );
1747 }
1748