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       boolean ok;
 670
 671       switch (decl->Declaration.File) {
 672       case TGSI_FILE_TEMPORARY:
 673          for (i = 0; i < NUM_CHANNELS; i++)
 674             bld->temps[idx][i] = lp_build_alloca(&bld->base);
 675          ok = TRUE;
 676          break;
 677
 678       case TGSI_FILE_OUTPUT:
 679          for (i = 0; i < NUM_CHANNELS; i++)
 680             bld->outputs[idx][i] = lp_build_alloca(&bld->base);
 681          ok = TRUE;
 682          break;
 683
 684       default:
 685          /* don't need to declare other vars */
 686          ok = TRUE;
 687       }
 688
 689       if (!ok) {
 690          LLVMPositionBuilderAtEnd(bld->base.builder,
 691                                   current_block);
 692          return FALSE;
 693       }
 694    }
 695
 696    LLVMPositionBuilderAtEnd(bld->base.builder,
 697                             current_block);
 698    return TRUE;
 699 }
 700
 701 static int
 702 emit_instruction(
 703    struct lp_build_tgsi_soa_context *bld,
 704    const struct tgsi_full_instruction *inst,
 705    const struct tgsi_opcode_info *info)
 706 {
 707    unsigned chan_index;
 708    LLVMValueRef src0, src1, src2;
 709    LLVMValueRef tmp0, tmp1, tmp2;
 710    LLVMValueRef tmp3 = NULL;
 711    LLVMValueRef tmp4 = NULL;
 712    LLVMValueRef tmp5 = NULL;
 713    LLVMValueRef tmp6 = NULL;
 714    LLVMValueRef tmp7 = NULL;
 715    LLVMValueRef res;
 716    LLVMValueRef dst0[NUM_CHANNELS];
 717
 718    /* we can't handle indirect addressing into temp register file yet */
 719    if (indirect_temp_reference(inst))
 720       return FALSE;
 721
 722    /*
 723     * Stores and write masks are handled in a general fashion after the long
 724     * instruction opcode switch statement.
 725     *
 726     * Although not stricitly necessary, we avoid generating instructions for
 727     * channels which won't be stored, in cases where's that easy. For some
 728     * complex instructions, like texture sampling, it is more convenient to
 729     * assume a full writemask and then let LLVM optimization passes eliminate
 730     * redundant code.
 731     */
 732
 733    assert(info->num_dst <= 1);
 734    if(info->num_dst) {
 735       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 736          dst0[chan_index] = bld->base.undef;
 737       }
 738    }
 739
 740    switch (inst->Instruction.Opcode) {
 741 #if 0
 742    case TGSI_OPCODE_ARL:
 743       /* FIXME */
 744       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 745          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 746          emit_flr(bld, 0, 0);
 747          emit_f2it( bld, 0 );
 748          dst0[chan_index] = tmp0;
 749       }
 750       break;
 751 #endif
 752
 753    case TGSI_OPCODE_MOV:
 754       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 755          dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
 756       }
 757       break;
 758
 759    case TGSI_OPCODE_LIT:
 760       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 761          dst0[CHAN_X] = bld->base.one;
 762       }
 763       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 764          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 765          dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
 766       }
 767       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 768          /* XMM[1] = SrcReg[0].yyyy */
 769          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 770          /* XMM[1] = max(XMM[1], 0) */
 771          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 772          /* XMM[2] = SrcReg[0].wwww */
 773          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 774          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 775          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 776          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 777          dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 778       }
 779       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 780          dst0[CHAN_W] = bld->base.one;
 781       }
 782       break;
 783
 784    case TGSI_OPCODE_RCP:
 785    /* TGSI_OPCODE_RECIP */
 786       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 787       res = lp_build_rcp(&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_RSQ:
 794    /* TGSI_OPCODE_RECIPSQRT */
 795       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 796       src0 = lp_build_abs(&bld->base, src0);
 797       res = lp_build_rsqrt(&bld->base, src0);
 798       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 799          dst0[chan_index] = res;
 800       }
 801       break;
 802
 803    case TGSI_OPCODE_EXP:
 804       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 805           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 806           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 807          LLVMValueRef *p_exp2_int_part = NULL;
 808          LLVMValueRef *p_frac_part = NULL;
 809          LLVMValueRef *p_exp2 = NULL;
 810
 811          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 812
 813          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 814             p_exp2_int_part = &tmp0;
 815          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 816             p_frac_part = &tmp1;
 817          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 818             p_exp2 = &tmp2;
 819
 820          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 821
 822          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 823             dst0[CHAN_X] = tmp0;
 824          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 825             dst0[CHAN_Y] = tmp1;
 826          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 827             dst0[CHAN_Z] = tmp2;
 828       }
 829       /* dst.w = 1.0 */
 830       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 831          dst0[CHAN_W] = bld->base.one;
 832       }
 833       break;
 834
 835    case TGSI_OPCODE_LOG:
 836       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 837           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 838           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 839          LLVMValueRef *p_floor_log2 = NULL;
 840          LLVMValueRef *p_exp = NULL;
 841          LLVMValueRef *p_log2 = NULL;
 842
 843          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 844          src0 = lp_build_abs( &bld->base, src0 );
 845
 846          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 847             p_floor_log2 = &tmp0;
 848          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 849             p_exp = &tmp1;
 850          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 851             p_log2 = &tmp2;
 852
 853          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 854
 855          /* dst.x = floor(lg2(abs(src.x))) */
 856          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 857             dst0[CHAN_X] = tmp0;
 858          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 859          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 860             dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
 861          }
 862          /* dst.z = lg2(abs(src.x)) */
 863          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 864             dst0[CHAN_Z] = tmp2;
 865       }
 866       /* dst.w = 1.0 */
 867       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 868          dst0[CHAN_W] = bld->base.one;
 869       }
 870       break;
 871
 872    case TGSI_OPCODE_MUL:
 873       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 874          src0 = emit_fetch( bld, inst, 0, chan_index );
 875          src1 = emit_fetch( bld, inst, 1, chan_index );
 876          dst0[chan_index] = lp_build_mul(&bld->base, src0, src1);
 877       }
 878       break;
 879
 880    case TGSI_OPCODE_ADD:
 881       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 882          src0 = emit_fetch( bld, inst, 0, chan_index );
 883          src1 = emit_fetch( bld, inst, 1, chan_index );
 884          dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
 885       }
 886       break;
 887
 888    case TGSI_OPCODE_DP3:
 889    /* TGSI_OPCODE_DOT3 */
 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       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 902          dst0[chan_index] = tmp0;
 903       }
 904       break;
 905
 906    case TGSI_OPCODE_DP4:
 907    /* TGSI_OPCODE_DOT4 */
 908       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 909       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 910       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 911       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 912       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 913       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 914       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 915       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 916       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 917       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 918       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 919       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
 920       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
 921       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 922       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 923       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 924          dst0[chan_index] = tmp0;
 925       }
 926       break;
 927
 928    case TGSI_OPCODE_DST:
 929       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
 930          dst0[CHAN_X] = bld->base.one;
 931       }
 932       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
 933          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
 934          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
 935          dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
 936       }
 937       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
 938          dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
 939       }
 940       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
 941          dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
 942       }
 943       break;
 944
 945    case TGSI_OPCODE_MIN:
 946       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 947          src0 = emit_fetch( bld, inst, 0, chan_index );
 948          src1 = emit_fetch( bld, inst, 1, chan_index );
 949          dst0[chan_index] = lp_build_min( &bld->base, src0, src1 );
 950       }
 951       break;
 952
 953    case TGSI_OPCODE_MAX:
 954       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 955          src0 = emit_fetch( bld, inst, 0, chan_index );
 956          src1 = emit_fetch( bld, inst, 1, chan_index );
 957          dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
 958       }
 959       break;
 960
 961    case TGSI_OPCODE_SLT:
 962    /* TGSI_OPCODE_SETLT */
 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_LESS, 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_SGE:
 972    /* TGSI_OPCODE_SETGE */
 973       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 974          src0 = emit_fetch( bld, inst, 0, chan_index );
 975          src1 = emit_fetch( bld, inst, 1, chan_index );
 976          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
 977          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 978       }
 979       break;
 980
 981    case TGSI_OPCODE_MAD:
 982    /* TGSI_OPCODE_MADD */
 983       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 984          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 985          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 986          tmp2 = emit_fetch( bld, inst, 2, chan_index );
 987          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 988          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
 989          dst0[chan_index] = tmp0;
 990       }
 991       break;
 992
 993    case TGSI_OPCODE_SUB:
 994       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 995          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 996          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 997          dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
 998       }
 999       break;
1000
1001    case TGSI_OPCODE_LRP:
1002       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1003          src0 = emit_fetch( bld, inst, 0, chan_index );
1004          src1 = emit_fetch( bld, inst, 1, chan_index );
1005          src2 = emit_fetch( bld, inst, 2, chan_index );
1006          tmp0 = lp_build_sub( &bld->base, src1, src2 );
1007          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
1008          dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
1009       }
1010       break;
1011
1012    case TGSI_OPCODE_CND:
1013       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1014          src0 = emit_fetch( bld, inst, 0, chan_index );
1015          src1 = emit_fetch( bld, inst, 1, chan_index );
1016          src2 = emit_fetch( bld, inst, 2, chan_index );
1017          tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1018          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1019          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1020       }
1021       break;
1022
1023    case TGSI_OPCODE_DP2A:
1024       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1025       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1026       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1027       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1028       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1029       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1030       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1031       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
1032       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1033       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1034          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1035       }
1036       break;
1037
1038    case TGSI_OPCODE_FRC:
1039       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1040          src0 = emit_fetch( bld, inst, 0, chan_index );
1041          tmp0 = lp_build_floor(&bld->base, src0);
1042          tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1043          dst0[chan_index] = tmp0;
1044       }
1045       break;
1046
1047    case TGSI_OPCODE_CLAMP:
1048       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1049          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1050          src1 = emit_fetch( bld, inst, 1, chan_index );
1051          src2 = emit_fetch( bld, inst, 2, chan_index );
1052          tmp0 = lp_build_max(&bld->base, tmp0, src1);
1053          tmp0 = lp_build_min(&bld->base, tmp0, src2);
1054          dst0[chan_index] = tmp0;
1055       }
1056       break;
1057
1058    case TGSI_OPCODE_FLR:
1059       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1060          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1061          dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1062       }
1063       break;
1064
1065    case TGSI_OPCODE_ROUND:
1066       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1067          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1068          dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1069       }
1070       break;
1071
1072    case TGSI_OPCODE_EX2: {
1073       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1074       tmp0 = lp_build_exp2( &bld->base, tmp0);
1075       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1076          dst0[chan_index] = tmp0;
1077       }
1078       break;
1079    }
1080
1081    case TGSI_OPCODE_LG2:
1082       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1083       tmp0 = lp_build_log2( &bld->base, tmp0);
1084       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1085          dst0[chan_index] = tmp0;
1086       }
1087       break;
1088
1089    case TGSI_OPCODE_POW:
1090       src0 = emit_fetch( bld, inst, 0, CHAN_X );
1091       src1 = emit_fetch( bld, inst, 1, CHAN_X );
1092       res = lp_build_pow( &bld->base, src0, src1 );
1093       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1094          dst0[chan_index] = res;
1095       }
1096       break;
1097
1098    case TGSI_OPCODE_XPD:
1099       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1100           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1101          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1102          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1103       }
1104       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1105           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1106          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1107          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1108       }
1109       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1110          tmp2 = tmp0;
1111          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1112          tmp5 = tmp3;
1113          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1114          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1115          dst0[CHAN_X] = tmp2;
1116       }
1117       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1118           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1119          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1120          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1121       }
1122       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1123          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1124          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1125          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1126          dst0[CHAN_Y] = tmp3;
1127       }
1128       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1129          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1130          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1131          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1132          dst0[CHAN_Z] = tmp5;
1133       }
1134       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1135          dst0[CHAN_W] = bld->base.one;
1136       }
1137       break;
1138
1139    case TGSI_OPCODE_ABS:
1140       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1141          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1142          dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1143       }
1144       break;
1145
1146    case TGSI_OPCODE_RCC:
1147       /* deprecated? */
1148       assert(0);
1149       return 0;
1150
1151    case TGSI_OPCODE_DPH:
1152       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1153       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1154       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1155       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1156       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1157       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1158       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1159       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1160       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1161       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1162       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1163       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1164       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1165       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1166          dst0[chan_index] = tmp0;
1167       }
1168       break;
1169
1170    case TGSI_OPCODE_COS:
1171       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1172       tmp0 = lp_build_cos( &bld->base, tmp0 );
1173       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1174          dst0[chan_index] = tmp0;
1175       }
1176       break;
1177
1178    case TGSI_OPCODE_DDX:
1179       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1180          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1181       }
1182       break;
1183
1184    case TGSI_OPCODE_DDY:
1185       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1186          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1187       }
1188       break;
1189
1190    case TGSI_OPCODE_KILP:
1191       /* predicated kill */
1192       /* FIXME */
1193       return 0;
1194       break;
1195
1196    case TGSI_OPCODE_KIL:
1197       /* conditional kill */
1198       emit_kil( bld, inst );
1199       break;
1200
1201    case TGSI_OPCODE_PK2H:
1202       return 0;
1203       break;
1204
1205    case TGSI_OPCODE_PK2US:
1206       return 0;
1207       break;
1208
1209    case TGSI_OPCODE_PK4B:
1210       return 0;
1211       break;
1212
1213    case TGSI_OPCODE_PK4UB:
1214       return 0;
1215       break;
1216
1217    case TGSI_OPCODE_RFL:
1218       return 0;
1219       break;
1220
1221    case TGSI_OPCODE_SEQ:
1222       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1223          src0 = emit_fetch( bld, inst, 0, chan_index );
1224          src1 = emit_fetch( bld, inst, 1, chan_index );
1225          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1226          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1227       }
1228       break;
1229
1230    case TGSI_OPCODE_SFL:
1231       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1232          dst0[chan_index] = bld->base.zero;
1233       }
1234       break;
1235
1236    case TGSI_OPCODE_SGT:
1237       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1238          src0 = emit_fetch( bld, inst, 0, chan_index );
1239          src1 = emit_fetch( bld, inst, 1, chan_index );
1240          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1241          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1242       }
1243       break;
1244
1245    case TGSI_OPCODE_SIN:
1246       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1247       tmp0 = lp_build_sin( &bld->base, tmp0 );
1248       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1249          dst0[chan_index] = tmp0;
1250       }
1251       break;
1252
1253    case TGSI_OPCODE_SLE:
1254       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1255          src0 = emit_fetch( bld, inst, 0, chan_index );
1256          src1 = emit_fetch( bld, inst, 1, chan_index );
1257          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
1258          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1259       }
1260       break;
1261
1262    case TGSI_OPCODE_SNE:
1263       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1264          src0 = emit_fetch( bld, inst, 0, chan_index );
1265          src1 = emit_fetch( bld, inst, 1, chan_index );
1266          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1267          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1268       }
1269       break;
1270
1271    case TGSI_OPCODE_STR:
1272       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1273          dst0[chan_index] = bld->base.one;
1274       }
1275       break;
1276
1277    case TGSI_OPCODE_TEX:
1278       emit_tex( bld, inst, FALSE, FALSE, dst0 );
1279       break;
1280
1281    case TGSI_OPCODE_TXD:
1282       /* FIXME */
1283       return 0;
1284       break;
1285
1286    case TGSI_OPCODE_UP2H:
1287       /* deprecated */
1288       assert (0);
1289       return 0;
1290       break;
1291
1292    case TGSI_OPCODE_UP2US:
1293       /* deprecated */
1294       assert(0);
1295       return 0;
1296       break;
1297
1298    case TGSI_OPCODE_UP4B:
1299       /* deprecated */
1300       assert(0);
1301       return 0;
1302       break;
1303
1304    case TGSI_OPCODE_UP4UB:
1305       /* deprecated */
1306       assert(0);
1307       return 0;
1308       break;
1309
1310    case TGSI_OPCODE_X2D:
1311       /* deprecated? */
1312       assert(0);
1313       return 0;
1314       break;
1315
1316    case TGSI_OPCODE_ARA:
1317       /* deprecated */
1318       assert(0);
1319       return 0;
1320       break;
1321
1322 #if 0
1323    case TGSI_OPCODE_ARR:
1324       /* FIXME */
1325       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1326          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1327          emit_rnd( bld, 0, 0 );
1328          emit_f2it( bld, 0 );
1329          dst0[chan_index] = tmp0;
1330       }
1331       break;
1332 #endif
1333
1334    case TGSI_OPCODE_BRA:
1335       /* deprecated */
1336       assert(0);
1337       return 0;
1338       break;
1339
1340    case TGSI_OPCODE_CAL:
1341       /* FIXME */
1342       return 0;
1343       break;
1344
1345    case TGSI_OPCODE_RET:
1346       /* FIXME */
1347       return 0;
1348       break;
1349
1350    case TGSI_OPCODE_END:
1351       break;
1352
1353    case TGSI_OPCODE_SSG:
1354    /* TGSI_OPCODE_SGN */
1355       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1356          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1357          dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1358       }
1359       break;
1360
1361    case TGSI_OPCODE_CMP:
1362       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1363          src0 = emit_fetch( bld, inst, 0, chan_index );
1364          src1 = emit_fetch( bld, inst, 1, chan_index );
1365          src2 = emit_fetch( bld, inst, 2, chan_index );
1366          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1367          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1368       }
1369       break;
1370
1371    case TGSI_OPCODE_SCS:
1372       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1373          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1374          dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1375       }
1376       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1377          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1378          dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1379       }
1380       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1381          dst0[CHAN_Z] = bld->base.zero;
1382       }
1383       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1384          dst0[CHAN_W] = bld->base.one;
1385       }
1386       break;
1387
1388    case TGSI_OPCODE_TXB:
1389       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1390       break;
1391
1392    case TGSI_OPCODE_NRM:
1393       /* fall-through */
1394    case TGSI_OPCODE_NRM4:
1395       /* 3 or 4-component normalization */
1396       {
1397          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1398
1399          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1400              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1401              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1402              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1403
1404             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1405
1406             /* xmm4 = src.x */
1407             /* xmm0 = src.x * src.x */
1408             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1409             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1410                tmp4 = tmp0;
1411             }
1412             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1413
1414             /* xmm5 = src.y */
1415             /* xmm0 = xmm0 + src.y * src.y */
1416             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1417             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1418                tmp5 = tmp1;
1419             }
1420             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1421             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1422
1423             /* xmm6 = src.z */
1424             /* xmm0 = xmm0 + src.z * src.z */
1425             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1426             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1427                tmp6 = tmp1;
1428             }
1429             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1430             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1431
1432             if (dims == 4) {
1433                /* xmm7 = src.w */
1434                /* xmm0 = xmm0 + src.w * src.w */
1435                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1436                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1437                   tmp7 = tmp1;
1438                }
1439                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1440                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1441             }
1442
1443             /* xmm1 = 1 / sqrt(xmm0) */
1444             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1445
1446             /* dst.x = xmm1 * src.x */
1447             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1448                dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1449             }
1450
1451             /* dst.y = xmm1 * src.y */
1452             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1453                dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1454             }
1455
1456             /* dst.z = xmm1 * src.z */
1457             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1458                dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1459             }
1460
1461             /* dst.w = xmm1 * src.w */
1462             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1463                dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1464             }
1465          }
1466
1467          /* dst.w = 1.0 */
1468          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1469             dst0[CHAN_W] = bld->base.one;
1470          }
1471       }
1472       break;
1473
1474    case TGSI_OPCODE_DIV:
1475       /* deprecated */
1476       assert( 0 );
1477       return 0;
1478       break;
1479
1480    case TGSI_OPCODE_DP2:
1481       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1482       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1483       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1484       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1485       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1486       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1487       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1488       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1489          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1490       }
1491       break;
1492
1493    case TGSI_OPCODE_TXL:
1494       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1495       break;
1496
1497    case TGSI_OPCODE_TXP:
1498       emit_tex( bld, inst, FALSE, TRUE, dst0 );
1499       break;
1500
1501    case TGSI_OPCODE_BRK:
1502       lp_exec_break(&bld->exec_mask);
1503       break;
1504
1505    case TGSI_OPCODE_IF:
1506       tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1507       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1508                           tmp0, bld->base.zero);
1509       lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1510       break;
1511
1512    case TGSI_OPCODE_BGNFOR:
1513       /* deprecated */
1514       assert(0);
1515       return 0;
1516       break;
1517
1518    case TGSI_OPCODE_BGNLOOP:
1519       lp_exec_bgnloop(&bld->exec_mask);
1520       break;
1521
1522    case TGSI_OPCODE_REP:
1523       /* deprecated */
1524       assert(0);
1525       return 0;
1526       break;
1527
1528    case TGSI_OPCODE_ELSE:
1529       lp_exec_mask_cond_invert(&bld->exec_mask);
1530       break;
1531
1532    case TGSI_OPCODE_ENDIF:
1533       lp_exec_mask_cond_pop(&bld->exec_mask);
1534       break;
1535
1536    case TGSI_OPCODE_ENDFOR:
1537       /* deprecated */
1538       assert(0);
1539       return 0;
1540       break;
1541
1542    case TGSI_OPCODE_ENDLOOP:
1543       lp_exec_endloop(&bld->exec_mask);
1544       break;
1545
1546    case TGSI_OPCODE_ENDREP:
1547       /* deprecated */
1548       assert(0);
1549       return 0;
1550       break;
1551
1552    case TGSI_OPCODE_PUSHA:
1553       /* deprecated? */
1554       assert(0);
1555       return 0;
1556       break;
1557
1558    case TGSI_OPCODE_POPA:
1559       /* deprecated? */
1560       assert(0);
1561       return 0;
1562       break;
1563
1564    case TGSI_OPCODE_CEIL:
1565       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1566          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1567          dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1568       }
1569       break;
1570
1571    case TGSI_OPCODE_I2F:
1572       /* deprecated? */
1573       assert(0);
1574       return 0;
1575       break;
1576
1577    case TGSI_OPCODE_NOT:
1578       /* deprecated? */
1579       assert(0);
1580       return 0;
1581       break;
1582
1583    case TGSI_OPCODE_TRUNC:
1584       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1585          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1586          dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1587       }
1588       break;
1589
1590    case TGSI_OPCODE_SHL:
1591       /* deprecated? */
1592       assert(0);
1593       return 0;
1594       break;
1595
1596    case TGSI_OPCODE_ISHR:
1597       /* deprecated? */
1598       assert(0);
1599       return 0;
1600       break;
1601
1602    case TGSI_OPCODE_AND:
1603       /* deprecated? */
1604       assert(0);
1605       return 0;
1606       break;
1607
1608    case TGSI_OPCODE_OR:
1609       /* deprecated? */
1610       assert(0);
1611       return 0;
1612       break;
1613
1614    case TGSI_OPCODE_MOD:
1615       /* deprecated? */
1616       assert(0);
1617       return 0;
1618       break;
1619
1620    case TGSI_OPCODE_XOR:
1621       /* deprecated? */
1622       assert(0);
1623       return 0;
1624       break;
1625
1626    case TGSI_OPCODE_SAD:
1627       /* deprecated? */
1628       assert(0);
1629       return 0;
1630       break;
1631
1632    case TGSI_OPCODE_TXF:
1633       /* deprecated? */
1634       assert(0);
1635       return 0;
1636       break;
1637
1638    case TGSI_OPCODE_TXQ:
1639       /* deprecated? */
1640       assert(0);
1641       return 0;
1642       break;
1643
1644    case TGSI_OPCODE_CONT:
1645       lp_exec_continue(&bld->exec_mask);
1646       break;
1647
1648    case TGSI_OPCODE_EMIT:
1649       return 0;
1650       break;
1651
1652    case TGSI_OPCODE_ENDPRIM:
1653       return 0;
1654       break;
1655
1656    case TGSI_OPCODE_NOP:
1657       break;
1658
1659    default:
1660       return 0;
1661    }
1662
1663    if(info->num_dst) {
1664       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1665          emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
1666       }
1667    }
1668
1669    return 1;
1670 }
1671
1672
1673 void
1674 lp_build_tgsi_soa(LLVMBuilderRef builder,
1675                   const struct tgsi_token *tokens,
1676                   struct lp_type type,
1677                   struct lp_build_mask_context *mask,
1678                   LLVMValueRef consts_ptr,
1679                   const LLVMValueRef *pos,
1680                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1681                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1682                   struct lp_build_sampler_soa *sampler)
1683 {
1684    struct lp_build_tgsi_soa_context bld;
1685    struct tgsi_parse_context parse;
1686    uint num_immediates = 0;
1687    unsigned i;
1688
1689    /* Setup build context */
1690    memset(&bld, 0, sizeof bld);
1691    lp_build_context_init(&bld.base, builder, type);
1692    bld.mask = mask;
1693    bld.pos = pos;
1694    bld.inputs = inputs;
1695    bld.outputs = outputs;
1696    bld.consts_ptr = consts_ptr;
1697    bld.sampler = sampler;
1698
1699    lp_exec_mask_init(&bld.exec_mask, &bld.base);
1700
1701    tgsi_parse_init( &parse, tokens );
1702
1703    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1704       tgsi_parse_token( &parse );
1705
1706       switch( parse.FullToken.Token.Type ) {
1707       case TGSI_TOKEN_TYPE_DECLARATION:
1708          /* Inputs already interpolated */
1709          {
1710             if (!emit_declaration( &bld, &parse.FullToken.FullDeclaration ))
1711                _debug_printf("warning: failed to define LLVM variable\n");
1712          }
1713          break;
1714
1715       case TGSI_TOKEN_TYPE_INSTRUCTION:
1716          {
1717             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1718             const struct tgsi_opcode_info *info = tgsi_get_opcode_info(opcode);
1719             if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, info ))
1720                _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1721                              info ? info->mnemonic : "<invalid>");
1722          }
1723
1724          break;
1725
1726       case TGSI_TOKEN_TYPE_IMMEDIATE:
1727          /* simply copy the immediate values into the next immediates[] slot */
1728          {
1729             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1730             assert(size <= 4);
1731             assert(num_immediates < LP_MAX_IMMEDIATES);
1732             for( i = 0; i < size; ++i )
1733                bld.immediates[num_immediates][i] =
1734                   lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
1735             for( i = size; i < 4; ++i )
1736                bld.immediates[num_immediates][i] = bld.base.undef;
1737             num_immediates++;
1738          }
1739          break;
1740
1741       case TGSI_TOKEN_TYPE_PROPERTY:
1742          break;
1743
1744       default:
1745          assert( 0 );
1746       }
1747    }
1748    if (0) {
1749       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1750       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1751       debug_printf("11111111111111111111111111111 \n");
1752       tgsi_dump(tokens, 0);
1753       LLVMDumpValue(function);
1754       debug_printf("2222222222222222222222222222 \n");
1755    }
1756    tgsi_parse_free( &parse );
1757 }
1758