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 "tgsi/tgsi_scan.h"
  50 #include "lp_bld_type.h"
  51 #include "lp_bld_const.h"
  52 #include "lp_bld_arit.h"
  53 #include "lp_bld_logic.h"
  54 #include "lp_bld_swizzle.h"
  55 #include "lp_bld_flow.h"
  56 #include "lp_bld_tgsi.h"
  57 #include "lp_bld_debug.h"
  58
  59
  60 #define LP_MAX_TEMPS 256
  61 #define LP_MAX_IMMEDIATES 256
  62
  63
  64 #define FOR_EACH_CHANNEL( CHAN )\
  65    for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
  66
  67 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  68    ((INST)->Dst[0].Register.WriteMask & (1 << (CHAN)))
  69
  70 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  71    if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
  72
  73 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
  74    FOR_EACH_CHANNEL( CHAN )\
  75       IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
  76
  77 #define CHAN_X 0
  78 #define CHAN_Y 1
  79 #define CHAN_Z 2
  80 #define CHAN_W 3
  81
  82 #define QUAD_TOP_LEFT     0
  83 #define QUAD_TOP_RIGHT    1
  84 #define QUAD_BOTTOM_LEFT  2
  85 #define QUAD_BOTTOM_RIGHT 3
  86
  87 #define LP_TGSI_MAX_NESTING 16
  88
  89 struct lp_exec_mask {
  90    struct lp_build_context *bld;
  91
  92    boolean has_mask;
  93
  94    LLVMTypeRef int_vec_type;
  95
  96    LLVMValueRef cond_stack[LP_TGSI_MAX_NESTING];
  97    int cond_stack_size;
  98    LLVMValueRef cond_mask;
  99
 100    LLVMValueRef break_stack[LP_TGSI_MAX_NESTING];
 101    int break_stack_size;
 102    LLVMValueRef break_mask;
 103
 104    LLVMValueRef cont_stack[LP_TGSI_MAX_NESTING];
 105    int cont_stack_size;
 106    LLVMValueRef cont_mask;
 107
 108    LLVMBasicBlockRef loop_stack[LP_TGSI_MAX_NESTING];
 109    int loop_stack_size;
 110    LLVMBasicBlockRef loop_block;
 111
 112
 113    LLVMValueRef exec_mask;
 114 };
 115
 116 struct lp_build_tgsi_soa_context
 117 {
 118    struct lp_build_context base;
 119
 120    LLVMValueRef consts_ptr;
 121    const LLVMValueRef *pos;
 122    const LLVMValueRef (*inputs)[NUM_CHANNELS];
 123    LLVMValueRef (*outputs)[NUM_CHANNELS];
 124
 125    struct lp_build_sampler_soa *sampler;
 126
 127    LLVMValueRef immediates[LP_MAX_IMMEDIATES][NUM_CHANNELS];
 128    LLVMValueRef temps[LP_MAX_TEMPS][NUM_CHANNELS];
 129    LLVMValueRef addr[LP_MAX_TEMPS][NUM_CHANNELS];
 130
 131    /* we allocate an array of temps if we have indirect
 132     * addressing and then the temps above is unused */
 133    LLVMValueRef temps_array;
 134    boolean has_indirect_addressing;
 135
 136    struct lp_build_mask_context *mask;
 137    struct lp_exec_mask exec_mask;
 138 };
 139
 140 static const unsigned char
 141 swizzle_left[4] = {
 142    QUAD_TOP_LEFT,     QUAD_TOP_LEFT,
 143    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_LEFT
 144 };
 145
 146 static const unsigned char
 147 swizzle_right[4] = {
 148    QUAD_TOP_RIGHT,    QUAD_TOP_RIGHT,
 149    QUAD_BOTTOM_RIGHT, QUAD_BOTTOM_RIGHT
 150 };
 151
 152 static const unsigned char
 153 swizzle_top[4] = {
 154    QUAD_TOP_LEFT,     QUAD_TOP_RIGHT,
 155    QUAD_TOP_LEFT,     QUAD_TOP_RIGHT
 156 };
 157
 158 static const unsigned char
 159 swizzle_bottom[4] = {
 160    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT,
 161    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT
 162 };
 163
 164 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
 165 {
 166    mask->bld = bld;
 167    mask->has_mask = FALSE;
 168    mask->cond_stack_size = 0;
 169    mask->loop_stack_size = 0;
 170    mask->break_stack_size = 0;
 171    mask->cont_stack_size = 0;
 172
 173    mask->int_vec_type = lp_build_int_vec_type(mask->bld->type);
 174 }
 175
 176 static void lp_exec_mask_update(struct lp_exec_mask *mask)
 177 {
 178    if (mask->loop_stack_size) {
 179       /*for loops we need to update the entire mask at runtime */
 180       LLVMValueRef tmp;
 181       assert(mask->break_mask);
 182       tmp = LLVMBuildAnd(mask->bld->builder,
 183                          mask->cont_mask,
 184                          mask->break_mask,
 185                          "maskcb");
 186       mask->exec_mask = LLVMBuildAnd(mask->bld->builder,
 187                                      mask->cond_mask,
 188                                      tmp,
 189                                      "maskfull");
 190    } else
 191       mask->exec_mask = mask->cond_mask;
 192
 193
 194    mask->has_mask = (mask->cond_stack_size > 0 ||
 195                      mask->loop_stack_size > 0);
 196 }
 197
 198 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 199                                    LLVMValueRef val)
 200 {
 201    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 202    mask->cond_mask = LLVMBuildBitCast(mask->bld->builder, val,
 203                                       mask->int_vec_type, "");
 204
 205    lp_exec_mask_update(mask);
 206 }
 207
 208 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 209 {
 210    LLVMValueRef prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 211    LLVMValueRef inv_mask = LLVMBuildNot(mask->bld->builder,
 212                                         mask->cond_mask, "");
 213
 214    /* means that we didn't have any mask before and that
 215     * we were fully enabled */
 216    if (mask->cond_stack_size <= 1) {
 217       prev_mask = LLVMConstAllOnes(mask->int_vec_type);
 218    }
 219
 220    mask->cond_mask = LLVMBuildAnd(mask->bld->builder,
 221                                   inv_mask,
 222                                   prev_mask, "");
 223    lp_exec_mask_update(mask);
 224 }
 225
 226 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 227 {
 228    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 229    lp_exec_mask_update(mask);
 230 }
 231
 232 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 233 {
 234
 235    if (mask->cont_stack_size == 0)
 236       mask->cont_mask = LLVMConstAllOnes(mask->int_vec_type);
 237    if (mask->break_stack_size == 0)
 238       mask->break_mask = LLVMConstAllOnes(mask->int_vec_type);
 239    if (mask->cond_stack_size == 0)
 240       mask->cond_mask = LLVMConstAllOnes(mask->int_vec_type);
 241
 242    mask->break_stack[mask->break_stack_size++] = mask->break_mask;
 243    mask->cont_stack[mask->cont_stack_size++] = mask->cont_mask;
 244    mask->loop_stack[mask->loop_stack_size++] = mask->loop_block;
 245    mask->loop_block = lp_build_insert_new_block(mask->bld->builder, "bgnloop");
 246    LLVMBuildBr(mask->bld->builder, mask->loop_block);
 247    LLVMPositionBuilderAtEnd(mask->bld->builder, mask->loop_block);
 248
 249    lp_exec_mask_update(mask);
 250 }
 251
 252 static void lp_exec_break(struct lp_exec_mask *mask)
 253 {
 254    LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
 255                                          mask->exec_mask,
 256                                          "break");
 257
 258    mask->break_mask = LLVMBuildAnd(mask->bld->builder,
 259                                    mask->break_mask,
 260                                    exec_mask, "break_full");
 261
 262    lp_exec_mask_update(mask);
 263 }
 264
 265 static void lp_exec_continue(struct lp_exec_mask *mask)
 266 {
 267    LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
 268                                          mask->exec_mask,
 269                                          "");
 270
 271    mask->cont_mask = LLVMBuildAnd(mask->bld->builder,
 272                                   mask->cont_mask,
 273                                   exec_mask, "");
 274
 275    lp_exec_mask_update(mask);
 276 }
 277
 278
 279 static void lp_exec_endloop(struct lp_exec_mask *mask)
 280 {
 281    LLVMBasicBlockRef endloop;
 282    LLVMTypeRef reg_type = LLVMIntType(mask->bld->type.width*
 283                                       mask->bld->type.length);
 284    LLVMValueRef i1cond;
 285
 286    assert(mask->break_mask);
 287
 288    /* i1cond = (mask == 0) */
 289    i1cond = LLVMBuildICmp(
 290       mask->bld->builder,
 291       LLVMIntNE,
 292       LLVMBuildBitCast(mask->bld->builder, mask->break_mask, reg_type, ""),
 293       LLVMConstNull(reg_type), "");
 294
 295    endloop = lp_build_insert_new_block(mask->bld->builder, "endloop");
 296
 297    LLVMBuildCondBr(mask->bld->builder,
 298                    i1cond, mask->loop_block, endloop);
 299
 300    LLVMPositionBuilderAtEnd(mask->bld->builder, endloop);
 301
 302    mask->loop_block = mask->loop_stack[--mask->loop_stack_size];
 303    /* pop the cont mask */
 304    if (mask->cont_stack_size) {
 305       mask->cont_mask = mask->cont_stack[--mask->cont_stack_size];
 306    }
 307    /* pop the break mask */
 308    if (mask->break_stack_size) {
 309       mask->break_mask = mask->break_stack[--mask->break_stack_size];
 310    }
 311
 312    lp_exec_mask_update(mask);
 313 }
 314
 315 /* stores val into an address pointed to by dst.
 316  * mask->exec_mask is used to figure out which bits of val
 317  * should be stored into the address
 318  * (0 means don't store this bit, 1 means do store).
 319  */
 320 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 321                                LLVMValueRef val,
 322                                LLVMValueRef dst)
 323 {
 324    if (mask->has_mask) {
 325       LLVMValueRef real_val, dst_val;
 326
 327       dst_val = LLVMBuildLoad(mask->bld->builder, dst, "");
 328       real_val = lp_build_select(mask->bld,
 329                                  mask->exec_mask,
 330                                  val, dst_val);
 331
 332       LLVMBuildStore(mask->bld->builder, real_val, dst);
 333    } else
 334       LLVMBuildStore(mask->bld->builder, val, dst);
 335 }
 336
 337
 338 static LLVMValueRef
 339 emit_ddx(struct lp_build_tgsi_soa_context *bld,
 340          LLVMValueRef src)
 341 {
 342    LLVMValueRef src_left  = lp_build_swizzle1_aos(&bld->base, src, swizzle_left);
 343    LLVMValueRef src_right = lp_build_swizzle1_aos(&bld->base, src, swizzle_right);
 344    return lp_build_sub(&bld->base, src_right, src_left);
 345 }
 346
 347
 348 static LLVMValueRef
 349 emit_ddy(struct lp_build_tgsi_soa_context *bld,
 350          LLVMValueRef src)
 351 {
 352    LLVMValueRef src_top    = lp_build_swizzle1_aos(&bld->base, src, swizzle_top);
 353    LLVMValueRef src_bottom = lp_build_swizzle1_aos(&bld->base, src, swizzle_bottom);
 354    return lp_build_sub(&bld->base, src_top, src_bottom);
 355 }
 356
 357 static LLVMValueRef
 358 get_temp_ptr(struct lp_build_tgsi_soa_context *bld,
 359              unsigned index,
 360              unsigned swizzle,
 361              boolean is_indirect,
 362              LLVMValueRef addr)
 363 {
 364    if (!bld->has_indirect_addressing) {
 365       return bld->temps[index][swizzle];
 366    } else {
 367       LLVMValueRef lindex =
 368          LLVMConstInt(LLVMInt32Type(), index*4 + swizzle, 0);
 369       if (is_indirect)
 370          lindex = lp_build_add(&bld->base, lindex, addr);
 371       return LLVMBuildGEP(bld->base.builder, bld->temps_array, &lindex, 1, "");
 372    }
 373 }
 374
 375 /**
 376  * Register fetch.
 377  */
 378 static LLVMValueRef
 379 emit_fetch(
 380    struct lp_build_tgsi_soa_context *bld,
 381    const struct tgsi_full_instruction *inst,
 382    unsigned index,
 383    const unsigned chan_index )
 384 {
 385    const struct tgsi_full_src_register *reg = &inst->Src[index];
 386    unsigned swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 387    LLVMValueRef res;
 388    LLVMValueRef addr;
 389
 390    switch (swizzle) {
 391    case TGSI_SWIZZLE_X:
 392    case TGSI_SWIZZLE_Y:
 393    case TGSI_SWIZZLE_Z:
 394    case TGSI_SWIZZLE_W:
 395
 396       if (reg->Register.Indirect) {
 397          LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->base.type);
 398          unsigned swizzle = tgsi_util_get_src_register_swizzle( &reg->Indirect, chan_index );
 399          addr = LLVMBuildLoad(bld->base.builder,
 400                               bld->addr[reg->Indirect.Index][swizzle],
 401                               "");
 402          /* for indexing we want integers */
 403          addr = LLVMBuildFPToSI(bld->base.builder, addr,
 404                                 int_vec_type, "");
 405          addr = LLVMBuildExtractElement(bld->base.builder,
 406                                         addr, LLVMConstInt(LLVMInt32Type(), 0, 0),
 407                                         "");
 408          addr = lp_build_mul(&bld->base, addr, LLVMConstInt(LLVMInt32Type(), 4, 0));
 409       }
 410
 411       switch (reg->Register.File) {
 412       case TGSI_FILE_CONSTANT: {
 413          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), reg->Register.Index*4 + swizzle, 0);
 414          LLVMValueRef scalar, scalar_ptr;
 415
 416          if (reg->Register.Indirect) {
 417             /*lp_build_printf(bld->base.builder,
 418               "\taddr = %d\n", addr);*/
 419             index = lp_build_add(&bld->base, index, addr);
 420          }
 421          scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr, &index, 1, "");
 422          scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
 423
 424          res = lp_build_broadcast_scalar(&bld->base, scalar);
 425          break;
 426       }
 427
 428       case TGSI_FILE_IMMEDIATE:
 429          res = bld->immediates[reg->Register.Index][swizzle];
 430          assert(res);
 431          break;
 432
 433       case TGSI_FILE_INPUT:
 434          res = bld->inputs[reg->Register.Index][swizzle];
 435          assert(res);
 436          break;
 437
 438       case TGSI_FILE_TEMPORARY: {
 439          LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
 440                                               swizzle,
 441                                               reg->Register.Indirect,
 442                                               addr);
 443          res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
 444          if(!res)
 445             return bld->base.undef;
 446          break;
 447       }
 448
 449       default:
 450          assert( 0 );
 451          return bld->base.undef;
 452       }
 453       break;
 454
 455    default:
 456       assert( 0 );
 457       return bld->base.undef;
 458    }
 459
 460    switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
 461    case TGSI_UTIL_SIGN_CLEAR:
 462       res = lp_build_abs( &bld->base, res );
 463       break;
 464
 465    case TGSI_UTIL_SIGN_SET:
 466       /* TODO: Use bitwese OR for floating point */
 467       res = lp_build_abs( &bld->base, res );
 468       res = LLVMBuildNeg( bld->base.builder, res, "" );
 469       break;
 470
 471    case TGSI_UTIL_SIGN_TOGGLE:
 472       res = LLVMBuildNeg( bld->base.builder, res, "" );
 473       break;
 474
 475    case TGSI_UTIL_SIGN_KEEP:
 476       break;
 477    }
 478
 479    return res;
 480 }
 481
 482
 483 /**
 484  * Register fetch with derivatives.
 485  */
 486 static void
 487 emit_fetch_deriv(
 488    struct lp_build_tgsi_soa_context *bld,
 489    const struct tgsi_full_instruction *inst,
 490    unsigned index,
 491    const unsigned chan_index,
 492    LLVMValueRef *res,
 493    LLVMValueRef *ddx,
 494    LLVMValueRef *ddy)
 495 {
 496    LLVMValueRef src;
 497
 498    src = emit_fetch(bld, inst, index, chan_index);
 499
 500    if(res)
 501       *res = src;
 502
 503    /* TODO: use interpolation coeffs for inputs */
 504
 505    if(ddx)
 506       *ddx = emit_ddx(bld, src);
 507
 508    if(ddy)
 509       *ddy = emit_ddy(bld, src);
 510 }
 511
 512
 513 /**
 514  * Register store.
 515  */
 516 static void
 517 emit_store(
 518    struct lp_build_tgsi_soa_context *bld,
 519    const struct tgsi_full_instruction *inst,
 520    unsigned index,
 521    unsigned chan_index,
 522    LLVMValueRef value)
 523 {
 524    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 525    LLVMValueRef addr;
 526
 527    switch( inst->Instruction.Saturate ) {
 528    case TGSI_SAT_NONE:
 529       break;
 530
 531    case TGSI_SAT_ZERO_ONE:
 532       value = lp_build_max(&bld->base, value, bld->base.zero);
 533       value = lp_build_min(&bld->base, value, bld->base.one);
 534       break;
 535
 536    case TGSI_SAT_MINUS_PLUS_ONE:
 537       value = lp_build_max(&bld->base, value, lp_build_const_vec(bld->base.type, -1.0));
 538       value = lp_build_min(&bld->base, value, bld->base.one);
 539       break;
 540
 541    default:
 542       assert(0);
 543    }
 544
 545    if (reg->Register.Indirect) {
 546       LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->base.type);
 547       unsigned swizzle = tgsi_util_get_src_register_swizzle( &reg->Indirect, chan_index );
 548       addr = LLVMBuildLoad(bld->base.builder,
 549                            bld->addr[reg->Indirect.Index][swizzle],
 550                            "");
 551       /* for indexing we want integers */
 552       addr = LLVMBuildFPToSI(bld->base.builder, addr,
 553                              int_vec_type, "");
 554       addr = LLVMBuildExtractElement(bld->base.builder,
 555                                      addr, LLVMConstInt(LLVMInt32Type(), 0, 0),
 556                                      "");
 557       addr = lp_build_mul(&bld->base, addr, LLVMConstInt(LLVMInt32Type(), 4, 0));
 558    }
 559
 560    switch( reg->Register.File ) {
 561    case TGSI_FILE_OUTPUT:
 562       lp_exec_mask_store(&bld->exec_mask, value,
 563                          bld->outputs[reg->Register.Index][chan_index]);
 564       break;
 565
 566    case TGSI_FILE_TEMPORARY: {
 567       LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
 568                                            chan_index,
 569                                            reg->Register.Indirect,
 570                                            addr);
 571       lp_exec_mask_store(&bld->exec_mask, value, temp_ptr);
 572       break;
 573    }
 574
 575    case TGSI_FILE_ADDRESS:
 576       lp_exec_mask_store(&bld->exec_mask, value,
 577                          bld->addr[reg->Indirect.Index][chan_index]);
 578       break;
 579
 580    case TGSI_FILE_PREDICATE:
 581       /* FIXME */
 582       assert(0);
 583       break;
 584
 585    default:
 586       assert( 0 );
 587    }
 588 }
 589
 590
 591 /**
 592  * High-level instruction translators.
 593  */
 594
 595
 596 static void
 597 emit_tex( struct lp_build_tgsi_soa_context *bld,
 598           const struct tgsi_full_instruction *inst,
 599           boolean apply_lodbias,
 600           boolean projected,
 601           LLVMValueRef *texel)
 602 {
 603    const uint unit = inst->Src[1].Register.Index;
 604    LLVMValueRef lodbias;
 605    LLVMValueRef oow = NULL;
 606    LLVMValueRef coords[3];
 607    unsigned num_coords;
 608    unsigned i;
 609
 610    switch (inst->Texture.Texture) {
 611    case TGSI_TEXTURE_1D:
 612       num_coords = 1;
 613       break;
 614    case TGSI_TEXTURE_2D:
 615    case TGSI_TEXTURE_RECT:
 616       num_coords = 2;
 617       break;
 618    case TGSI_TEXTURE_SHADOW1D:
 619    case TGSI_TEXTURE_SHADOW2D:
 620    case TGSI_TEXTURE_SHADOWRECT:
 621    case TGSI_TEXTURE_3D:
 622    case TGSI_TEXTURE_CUBE:
 623       num_coords = 3;
 624       break;
 625    default:
 626       assert(0);
 627       return;
 628    }
 629
 630    if(apply_lodbias)
 631       lodbias = emit_fetch( bld, inst, 0, 3 );
 632    else
 633       lodbias = bld->base.zero;
 634
 635    if (projected) {
 636       oow = emit_fetch( bld, inst, 0, 3 );
 637       oow = lp_build_rcp(&bld->base, oow);
 638    }
 639
 640    for (i = 0; i < num_coords; i++) {
 641       coords[i] = emit_fetch( bld, inst, 0, i );
 642       if (projected)
 643          coords[i] = lp_build_mul(&bld->base, coords[i], oow);
 644    }
 645    for (i = num_coords; i < 3; i++) {
 646       coords[i] = bld->base.undef;
 647    }
 648
 649    bld->sampler->emit_fetch_texel(bld->sampler,
 650                                   bld->base.builder,
 651                                   bld->base.type,
 652                                   unit, num_coords, coords, lodbias,
 653                                   texel);
 654 }
 655
 656
 657 /**
 658  * Kill fragment if any of the src register values are negative.
 659  */
 660 static void
 661 emit_kil(
 662    struct lp_build_tgsi_soa_context *bld,
 663    const struct tgsi_full_instruction *inst )
 664 {
 665    const struct tgsi_full_src_register *reg = &inst->Src[0];
 666    LLVMValueRef terms[NUM_CHANNELS];
 667    LLVMValueRef mask;
 668    unsigned chan_index;
 669
 670    memset(&terms, 0, sizeof terms);
 671
 672    FOR_EACH_CHANNEL( chan_index ) {
 673       unsigned swizzle;
 674
 675       /* Unswizzle channel */
 676       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 677
 678       /* Check if the component has not been already tested. */
 679       assert(swizzle < NUM_CHANNELS);
 680       if( !terms[swizzle] )
 681          /* TODO: change the comparison operator instead of setting the sign */
 682          terms[swizzle] =  emit_fetch(bld, inst, 0, chan_index );
 683    }
 684
 685    mask = NULL;
 686    FOR_EACH_CHANNEL( chan_index ) {
 687       if(terms[chan_index]) {
 688          LLVMValueRef chan_mask;
 689
 690          /*
 691           * If term < 0 then mask = 0 else mask = ~0.
 692           */
 693          chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero);
 694
 695          if(mask)
 696             mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, "");
 697          else
 698             mask = chan_mask;
 699       }
 700    }
 701
 702    if(mask)
 703       lp_build_mask_update(bld->mask, mask);
 704 }
 705
 706
 707 /**
 708  * Predicated fragment kill.
 709  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
 710  * The only predication is the execution mask which will apply if
 711  * we're inside a loop or conditional.
 712  */
 713 static void
 714 emit_kilp(struct lp_build_tgsi_soa_context *bld,
 715           const struct tgsi_full_instruction *inst)
 716 {
 717    LLVMValueRef mask;
 718
 719    /* For those channels which are "alive", disable fragment shader
 720     * execution.
 721     */
 722    if (bld->exec_mask.has_mask) {
 723       mask = LLVMBuildNot(bld->base.builder, bld->exec_mask.exec_mask, "kilp");
 724    }
 725    else {
 726       mask = bld->base.zero;
 727    }
 728
 729    lp_build_mask_update(bld->mask, mask);
 730 }
 731
 732 static int
 733 emit_declaration(
 734    struct lp_build_tgsi_soa_context *bld,
 735    const struct tgsi_full_declaration *decl)
 736 {
 737    LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
 738
 739    unsigned first = decl->Range.First;
 740    unsigned last = decl->Range.Last;
 741    unsigned idx, i;
 742
 743    for (idx = first; idx <= last; ++idx) {
 744       switch (decl->Declaration.File) {
 745       case TGSI_FILE_TEMPORARY:
 746          if (bld->has_indirect_addressing) {
 747             LLVMValueRef val = LLVMConstInt(LLVMInt32Type(),
 748                                             last*4 + 4, 0);
 749             bld->temps_array = lp_build_array_alloca(bld->base.builder,
 750                                                      vec_type, val, "");
 751          } else {
 752             for (i = 0; i < NUM_CHANNELS; i++)
 753                bld->temps[idx][i] = lp_build_alloca(bld->base.builder,
 754                                                     vec_type, "");
 755          }
 756          break;
 757
 758       case TGSI_FILE_OUTPUT:
 759          for (i = 0; i < NUM_CHANNELS; i++)
 760             bld->outputs[idx][i] = lp_build_alloca(bld->base.builder,
 761                                                    vec_type, "");
 762          break;
 763
 764       case TGSI_FILE_ADDRESS:
 765          for (i = 0; i < NUM_CHANNELS; i++)
 766             bld->addr[idx][i] = lp_build_alloca(bld->base.builder,
 767                                                 vec_type, "");
 768          break;
 769
 770       default:
 771          /* don't need to declare other vars */
 772          break;
 773       }
 774    }
 775
 776    return TRUE;
 777 }
 778
 779
 780 /**
 781  * Emit LLVM for one TGSI instruction.
 782  * \param return TRUE for success, FALSE otherwise
 783  */
 784 static boolean
 785 emit_instruction(
 786    struct lp_build_tgsi_soa_context *bld,
 787    const struct tgsi_full_instruction *inst,
 788    const struct tgsi_opcode_info *info)
 789 {
 790    unsigned chan_index;
 791    LLVMValueRef src0, src1, src2;
 792    LLVMValueRef tmp0, tmp1, tmp2;
 793    LLVMValueRef tmp3 = NULL;
 794    LLVMValueRef tmp4 = NULL;
 795    LLVMValueRef tmp5 = NULL;
 796    LLVMValueRef tmp6 = NULL;
 797    LLVMValueRef tmp7 = NULL;
 798    LLVMValueRef res;
 799    LLVMValueRef dst0[NUM_CHANNELS];
 800
 801    /*
 802     * Stores and write masks are handled in a general fashion after the long
 803     * instruction opcode switch statement.
 804     *
 805     * Although not stricitly necessary, we avoid generating instructions for
 806     * channels which won't be stored, in cases where's that easy. For some
 807     * complex instructions, like texture sampling, it is more convenient to
 808     * assume a full writemask and then let LLVM optimization passes eliminate
 809     * redundant code.
 810     */
 811
 812    assert(info->num_dst <= 1);
 813    if(info->num_dst) {
 814       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 815          dst0[chan_index] = bld->base.undef;
 816       }
 817    }
 818
 819    switch (inst->Instruction.Opcode) {
 820    case TGSI_OPCODE_ARL:
 821       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 822          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 823          tmp0 = lp_build_floor(&bld->base, tmp0);
 824          dst0[chan_index] = tmp0;
 825       }
 826       break;
 827
 828    case TGSI_OPCODE_MOV:
 829       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 830          dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
 831       }
 832       break;
 833
 834    case TGSI_OPCODE_LIT:
 835       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 836          dst0[CHAN_X] = bld->base.one;
 837       }
 838       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 839          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 840          dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
 841       }
 842       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 843          /* XMM[1] = SrcReg[0].yyyy */
 844          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 845          /* XMM[1] = max(XMM[1], 0) */
 846          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 847          /* XMM[2] = SrcReg[0].wwww */
 848          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 849          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 850          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 851          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 852          dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 853       }
 854       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 855          dst0[CHAN_W] = bld->base.one;
 856       }
 857       break;
 858
 859    case TGSI_OPCODE_RCP:
 860    /* TGSI_OPCODE_RECIP */
 861       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 862       res = lp_build_rcp(&bld->base, src0);
 863       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 864          dst0[chan_index] = res;
 865       }
 866       break;
 867
 868    case TGSI_OPCODE_RSQ:
 869    /* TGSI_OPCODE_RECIPSQRT */
 870       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 871       src0 = lp_build_abs(&bld->base, src0);
 872       res = lp_build_rsqrt(&bld->base, src0);
 873       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 874          dst0[chan_index] = res;
 875       }
 876       break;
 877
 878    case TGSI_OPCODE_EXP:
 879       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 880           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 881           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 882          LLVMValueRef *p_exp2_int_part = NULL;
 883          LLVMValueRef *p_frac_part = NULL;
 884          LLVMValueRef *p_exp2 = NULL;
 885
 886          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 887
 888          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 889             p_exp2_int_part = &tmp0;
 890          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 891             p_frac_part = &tmp1;
 892          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 893             p_exp2 = &tmp2;
 894
 895          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 896
 897          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 898             dst0[CHAN_X] = tmp0;
 899          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 900             dst0[CHAN_Y] = tmp1;
 901          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 902             dst0[CHAN_Z] = tmp2;
 903       }
 904       /* dst.w = 1.0 */
 905       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 906          dst0[CHAN_W] = bld->base.one;
 907       }
 908       break;
 909
 910    case TGSI_OPCODE_LOG:
 911       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 912           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 913           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 914          LLVMValueRef *p_floor_log2 = NULL;
 915          LLVMValueRef *p_exp = NULL;
 916          LLVMValueRef *p_log2 = NULL;
 917
 918          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 919          src0 = lp_build_abs( &bld->base, src0 );
 920
 921          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 922             p_floor_log2 = &tmp0;
 923          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 924             p_exp = &tmp1;
 925          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 926             p_log2 = &tmp2;
 927
 928          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 929
 930          /* dst.x = floor(lg2(abs(src.x))) */
 931          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 932             dst0[CHAN_X] = tmp0;
 933          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 934          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 935             dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
 936          }
 937          /* dst.z = lg2(abs(src.x)) */
 938          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 939             dst0[CHAN_Z] = tmp2;
 940       }
 941       /* dst.w = 1.0 */
 942       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 943          dst0[CHAN_W] = bld->base.one;
 944       }
 945       break;
 946
 947    case TGSI_OPCODE_MUL:
 948       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 949          src0 = emit_fetch( bld, inst, 0, chan_index );
 950          src1 = emit_fetch( bld, inst, 1, chan_index );
 951          dst0[chan_index] = lp_build_mul(&bld->base, src0, src1);
 952       }
 953       break;
 954
 955    case TGSI_OPCODE_ADD:
 956       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 957          src0 = emit_fetch( bld, inst, 0, chan_index );
 958          src1 = emit_fetch( bld, inst, 1, chan_index );
 959          dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
 960       }
 961       break;
 962
 963    case TGSI_OPCODE_DP3:
 964    /* TGSI_OPCODE_DOT3 */
 965       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 966       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 967       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 968       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 969       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 970       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 971       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 972       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 973       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 974       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 975       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 976       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 977          dst0[chan_index] = tmp0;
 978       }
 979       break;
 980
 981    case TGSI_OPCODE_DP4:
 982    /* TGSI_OPCODE_DOT4 */
 983       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 984       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 985       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 986       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 987       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 988       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 989       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 990       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 991       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 992       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 993       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 994       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
 995       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
 996       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 997       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 998       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 999          dst0[chan_index] = tmp0;
1000       }
1001       break;
1002
1003    case TGSI_OPCODE_DST:
1004       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1005          dst0[CHAN_X] = bld->base.one;
1006       }
1007       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1008          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1009          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
1010          dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
1011       }
1012       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1013          dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
1014       }
1015       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1016          dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
1017       }
1018       break;
1019
1020    case TGSI_OPCODE_MIN:
1021       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1022          src0 = emit_fetch( bld, inst, 0, chan_index );
1023          src1 = emit_fetch( bld, inst, 1, chan_index );
1024          dst0[chan_index] = lp_build_min( &bld->base, src0, src1 );
1025       }
1026       break;
1027
1028    case TGSI_OPCODE_MAX:
1029       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1030          src0 = emit_fetch( bld, inst, 0, chan_index );
1031          src1 = emit_fetch( bld, inst, 1, chan_index );
1032          dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
1033       }
1034       break;
1035
1036    case TGSI_OPCODE_SLT:
1037    /* TGSI_OPCODE_SETLT */
1038       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1039          src0 = emit_fetch( bld, inst, 0, chan_index );
1040          src1 = emit_fetch( bld, inst, 1, chan_index );
1041          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
1042          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1043       }
1044       break;
1045
1046    case TGSI_OPCODE_SGE:
1047    /* TGSI_OPCODE_SETGE */
1048       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1049          src0 = emit_fetch( bld, inst, 0, chan_index );
1050          src1 = emit_fetch( bld, inst, 1, chan_index );
1051          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
1052          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1053       }
1054       break;
1055
1056    case TGSI_OPCODE_MAD:
1057    /* TGSI_OPCODE_MADD */
1058       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1059          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1060          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1061          tmp2 = emit_fetch( bld, inst, 2, chan_index );
1062          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1063          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
1064          dst0[chan_index] = tmp0;
1065       }
1066       break;
1067
1068    case TGSI_OPCODE_SUB:
1069       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1070          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1071          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1072          dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
1073       }
1074       break;
1075
1076    case TGSI_OPCODE_LRP:
1077       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1078          src0 = emit_fetch( bld, inst, 0, chan_index );
1079          src1 = emit_fetch( bld, inst, 1, chan_index );
1080          src2 = emit_fetch( bld, inst, 2, chan_index );
1081          tmp0 = lp_build_sub( &bld->base, src1, src2 );
1082          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
1083          dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
1084       }
1085       break;
1086
1087    case TGSI_OPCODE_CND:
1088       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1089          src0 = emit_fetch( bld, inst, 0, chan_index );
1090          src1 = emit_fetch( bld, inst, 1, chan_index );
1091          src2 = emit_fetch( bld, inst, 2, chan_index );
1092          tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1093          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1094          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1095       }
1096       break;
1097
1098    case TGSI_OPCODE_DP2A:
1099       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1100       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1101       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1102       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1103       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1104       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1105       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1106       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
1107       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1108       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1109          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1110       }
1111       break;
1112
1113    case TGSI_OPCODE_FRC:
1114       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1115          src0 = emit_fetch( bld, inst, 0, chan_index );
1116          tmp0 = lp_build_floor(&bld->base, src0);
1117          tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1118          dst0[chan_index] = tmp0;
1119       }
1120       break;
1121
1122    case TGSI_OPCODE_CLAMP:
1123       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1124          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1125          src1 = emit_fetch( bld, inst, 1, chan_index );
1126          src2 = emit_fetch( bld, inst, 2, chan_index );
1127          tmp0 = lp_build_max(&bld->base, tmp0, src1);
1128          tmp0 = lp_build_min(&bld->base, tmp0, src2);
1129          dst0[chan_index] = tmp0;
1130       }
1131       break;
1132
1133    case TGSI_OPCODE_FLR:
1134       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1135          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1136          dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1137       }
1138       break;
1139
1140    case TGSI_OPCODE_ROUND:
1141       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1142          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1143          dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1144       }
1145       break;
1146
1147    case TGSI_OPCODE_EX2: {
1148       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1149       tmp0 = lp_build_exp2( &bld->base, tmp0);
1150       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1151          dst0[chan_index] = tmp0;
1152       }
1153       break;
1154    }
1155
1156    case TGSI_OPCODE_LG2:
1157       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1158       tmp0 = lp_build_log2( &bld->base, tmp0);
1159       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1160          dst0[chan_index] = tmp0;
1161       }
1162       break;
1163
1164    case TGSI_OPCODE_POW:
1165       src0 = emit_fetch( bld, inst, 0, CHAN_X );
1166       src1 = emit_fetch( bld, inst, 1, CHAN_X );
1167       res = lp_build_pow( &bld->base, src0, src1 );
1168       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1169          dst0[chan_index] = res;
1170       }
1171       break;
1172
1173    case TGSI_OPCODE_XPD:
1174       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1175           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1176          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1177          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1178       }
1179       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1180           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1181          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1182          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1183       }
1184       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1185          tmp2 = tmp0;
1186          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1187          tmp5 = tmp3;
1188          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1189          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1190          dst0[CHAN_X] = tmp2;
1191       }
1192       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1193           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1194          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1195          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1196       }
1197       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1198          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1199          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1200          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1201          dst0[CHAN_Y] = tmp3;
1202       }
1203       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1204          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1205          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1206          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1207          dst0[CHAN_Z] = tmp5;
1208       }
1209       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1210          dst0[CHAN_W] = bld->base.one;
1211       }
1212       break;
1213
1214    case TGSI_OPCODE_ABS:
1215       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1216          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1217          dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1218       }
1219       break;
1220
1221    case TGSI_OPCODE_RCC:
1222       /* deprecated? */
1223       assert(0);
1224       return FALSE;
1225
1226    case TGSI_OPCODE_DPH:
1227       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1228       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1229       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1230       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1231       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1232       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1233       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1234       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1235       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1236       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1237       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1238       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1239       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1240       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1241          dst0[chan_index] = tmp0;
1242       }
1243       break;
1244
1245    case TGSI_OPCODE_COS:
1246       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1247       tmp0 = lp_build_cos( &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_DDX:
1254       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1255          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1256       }
1257       break;
1258
1259    case TGSI_OPCODE_DDY:
1260       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1261          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1262       }
1263       break;
1264
1265    case TGSI_OPCODE_KILP:
1266       /* predicated kill */
1267       emit_kilp( bld, inst );
1268       break;
1269
1270    case TGSI_OPCODE_KIL:
1271       /* conditional kill */
1272       emit_kil( bld, inst );
1273       break;
1274
1275    case TGSI_OPCODE_PK2H:
1276       return FALSE;
1277       break;
1278
1279    case TGSI_OPCODE_PK2US:
1280       return FALSE;
1281       break;
1282
1283    case TGSI_OPCODE_PK4B:
1284       return FALSE;
1285       break;
1286
1287    case TGSI_OPCODE_PK4UB:
1288       return FALSE;
1289       break;
1290
1291    case TGSI_OPCODE_RFL:
1292       return FALSE;
1293       break;
1294
1295    case TGSI_OPCODE_SEQ:
1296       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1297          src0 = emit_fetch( bld, inst, 0, chan_index );
1298          src1 = emit_fetch( bld, inst, 1, chan_index );
1299          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1300          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1301       }
1302       break;
1303
1304    case TGSI_OPCODE_SFL:
1305       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1306          dst0[chan_index] = bld->base.zero;
1307       }
1308       break;
1309
1310    case TGSI_OPCODE_SGT:
1311       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1312          src0 = emit_fetch( bld, inst, 0, chan_index );
1313          src1 = emit_fetch( bld, inst, 1, chan_index );
1314          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1315          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1316       }
1317       break;
1318
1319    case TGSI_OPCODE_SIN:
1320       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1321       tmp0 = lp_build_sin( &bld->base, tmp0 );
1322       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1323          dst0[chan_index] = tmp0;
1324       }
1325       break;
1326
1327    case TGSI_OPCODE_SLE:
1328       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1329          src0 = emit_fetch( bld, inst, 0, chan_index );
1330          src1 = emit_fetch( bld, inst, 1, chan_index );
1331          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
1332          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1333       }
1334       break;
1335
1336    case TGSI_OPCODE_SNE:
1337       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1338          src0 = emit_fetch( bld, inst, 0, chan_index );
1339          src1 = emit_fetch( bld, inst, 1, chan_index );
1340          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1341          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1342       }
1343       break;
1344
1345    case TGSI_OPCODE_STR:
1346       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1347          dst0[chan_index] = bld->base.one;
1348       }
1349       break;
1350
1351    case TGSI_OPCODE_TEX:
1352       emit_tex( bld, inst, FALSE, FALSE, dst0 );
1353       break;
1354
1355    case TGSI_OPCODE_TXD:
1356       /* FIXME */
1357       return FALSE;
1358       break;
1359
1360    case TGSI_OPCODE_UP2H:
1361       /* deprecated */
1362       assert (0);
1363       return FALSE;
1364       break;
1365
1366    case TGSI_OPCODE_UP2US:
1367       /* deprecated */
1368       assert(0);
1369       return FALSE;
1370       break;
1371
1372    case TGSI_OPCODE_UP4B:
1373       /* deprecated */
1374       assert(0);
1375       return FALSE;
1376       break;
1377
1378    case TGSI_OPCODE_UP4UB:
1379       /* deprecated */
1380       assert(0);
1381       return FALSE;
1382       break;
1383
1384    case TGSI_OPCODE_X2D:
1385       /* deprecated? */
1386       assert(0);
1387       return FALSE;
1388       break;
1389
1390    case TGSI_OPCODE_ARA:
1391       /* deprecated */
1392       assert(0);
1393       return FALSE;
1394       break;
1395
1396    case TGSI_OPCODE_ARR:
1397       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1398          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1399          tmp0 = lp_build_round(&bld->base, tmp0);
1400          dst0[chan_index] = tmp0;
1401       }
1402       break;
1403
1404    case TGSI_OPCODE_BRA:
1405       /* deprecated */
1406       assert(0);
1407       return FALSE;
1408       break;
1409
1410    case TGSI_OPCODE_CAL:
1411       /* FIXME */
1412       return FALSE;
1413       break;
1414
1415    case TGSI_OPCODE_RET:
1416       /* FIXME */
1417       return FALSE;
1418       break;
1419
1420    case TGSI_OPCODE_END:
1421       break;
1422
1423    case TGSI_OPCODE_SSG:
1424    /* TGSI_OPCODE_SGN */
1425       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1426          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1427          dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1428       }
1429       break;
1430
1431    case TGSI_OPCODE_CMP:
1432       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1433          src0 = emit_fetch( bld, inst, 0, chan_index );
1434          src1 = emit_fetch( bld, inst, 1, chan_index );
1435          src2 = emit_fetch( bld, inst, 2, chan_index );
1436          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1437          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1438       }
1439       break;
1440
1441    case TGSI_OPCODE_SCS:
1442       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1443          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1444          dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1445       }
1446       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1447          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1448          dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1449       }
1450       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1451          dst0[CHAN_Z] = bld->base.zero;
1452       }
1453       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1454          dst0[CHAN_W] = bld->base.one;
1455       }
1456       break;
1457
1458    case TGSI_OPCODE_TXB:
1459       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1460       break;
1461
1462    case TGSI_OPCODE_NRM:
1463       /* fall-through */
1464    case TGSI_OPCODE_NRM4:
1465       /* 3 or 4-component normalization */
1466       {
1467          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1468
1469          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1470              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1471              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1472              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1473
1474             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1475
1476             /* xmm4 = src.x */
1477             /* xmm0 = src.x * src.x */
1478             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1479             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1480                tmp4 = tmp0;
1481             }
1482             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1483
1484             /* xmm5 = src.y */
1485             /* xmm0 = xmm0 + src.y * src.y */
1486             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1487             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1488                tmp5 = tmp1;
1489             }
1490             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1491             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1492
1493             /* xmm6 = src.z */
1494             /* xmm0 = xmm0 + src.z * src.z */
1495             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1496             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1497                tmp6 = tmp1;
1498             }
1499             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1500             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1501
1502             if (dims == 4) {
1503                /* xmm7 = src.w */
1504                /* xmm0 = xmm0 + src.w * src.w */
1505                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1506                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1507                   tmp7 = tmp1;
1508                }
1509                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1510                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1511             }
1512
1513             /* xmm1 = 1 / sqrt(xmm0) */
1514             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1515
1516             /* dst.x = xmm1 * src.x */
1517             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1518                dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1519             }
1520
1521             /* dst.y = xmm1 * src.y */
1522             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1523                dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1524             }
1525
1526             /* dst.z = xmm1 * src.z */
1527             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1528                dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1529             }
1530
1531             /* dst.w = xmm1 * src.w */
1532             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1533                dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1534             }
1535          }
1536
1537          /* dst.w = 1.0 */
1538          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1539             dst0[CHAN_W] = bld->base.one;
1540          }
1541       }
1542       break;
1543
1544    case TGSI_OPCODE_DIV:
1545       /* deprecated */
1546       assert( 0 );
1547       return FALSE;
1548       break;
1549
1550    case TGSI_OPCODE_DP2:
1551       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1552       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1553       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1554       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1555       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1556       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1557       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1558       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1559          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1560       }
1561       break;
1562
1563    case TGSI_OPCODE_TXL:
1564       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1565       break;
1566
1567    case TGSI_OPCODE_TXP:
1568       emit_tex( bld, inst, FALSE, TRUE, dst0 );
1569       break;
1570
1571    case TGSI_OPCODE_BRK:
1572       lp_exec_break(&bld->exec_mask);
1573       break;
1574
1575    case TGSI_OPCODE_IF:
1576       tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1577       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1578                           tmp0, bld->base.zero);
1579       lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1580       break;
1581
1582    case TGSI_OPCODE_BGNLOOP:
1583       lp_exec_bgnloop(&bld->exec_mask);
1584       break;
1585
1586    case TGSI_OPCODE_ELSE:
1587       lp_exec_mask_cond_invert(&bld->exec_mask);
1588       break;
1589
1590    case TGSI_OPCODE_ENDIF:
1591       lp_exec_mask_cond_pop(&bld->exec_mask);
1592       break;
1593
1594    case TGSI_OPCODE_ENDLOOP:
1595       lp_exec_endloop(&bld->exec_mask);
1596       break;
1597
1598    case TGSI_OPCODE_PUSHA:
1599       /* deprecated? */
1600       assert(0);
1601       return FALSE;
1602       break;
1603
1604    case TGSI_OPCODE_POPA:
1605       /* deprecated? */
1606       assert(0);
1607       return FALSE;
1608       break;
1609
1610    case TGSI_OPCODE_CEIL:
1611       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1612          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1613          dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1614       }
1615       break;
1616
1617    case TGSI_OPCODE_I2F:
1618       /* deprecated? */
1619       assert(0);
1620       return FALSE;
1621       break;
1622
1623    case TGSI_OPCODE_NOT:
1624       /* deprecated? */
1625       assert(0);
1626       return FALSE;
1627       break;
1628
1629    case TGSI_OPCODE_TRUNC:
1630       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1631          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1632          dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1633       }
1634       break;
1635
1636    case TGSI_OPCODE_SHL:
1637       /* deprecated? */
1638       assert(0);
1639       return FALSE;
1640       break;
1641
1642    case TGSI_OPCODE_ISHR:
1643       /* deprecated? */
1644       assert(0);
1645       return FALSE;
1646       break;
1647
1648    case TGSI_OPCODE_AND:
1649       /* deprecated? */
1650       assert(0);
1651       return FALSE;
1652       break;
1653
1654    case TGSI_OPCODE_OR:
1655       /* deprecated? */
1656       assert(0);
1657       return FALSE;
1658       break;
1659
1660    case TGSI_OPCODE_MOD:
1661       /* deprecated? */
1662       assert(0);
1663       return FALSE;
1664       break;
1665
1666    case TGSI_OPCODE_XOR:
1667       /* deprecated? */
1668       assert(0);
1669       return FALSE;
1670       break;
1671
1672    case TGSI_OPCODE_SAD:
1673       /* deprecated? */
1674       assert(0);
1675       return FALSE;
1676       break;
1677
1678    case TGSI_OPCODE_TXF:
1679       /* deprecated? */
1680       assert(0);
1681       return FALSE;
1682       break;
1683
1684    case TGSI_OPCODE_TXQ:
1685       /* deprecated? */
1686       assert(0);
1687       return FALSE;
1688       break;
1689
1690    case TGSI_OPCODE_CONT:
1691       lp_exec_continue(&bld->exec_mask);
1692       break;
1693
1694    case TGSI_OPCODE_EMIT:
1695       return FALSE;
1696       break;
1697
1698    case TGSI_OPCODE_ENDPRIM:
1699       return FALSE;
1700       break;
1701
1702    case TGSI_OPCODE_NOP:
1703       break;
1704
1705    default:
1706       return FALSE;
1707    }
1708
1709    if(info->num_dst) {
1710       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1711          emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
1712       }
1713    }
1714
1715    return TRUE;
1716 }
1717
1718
1719 void
1720 lp_build_tgsi_soa(LLVMBuilderRef builder,
1721                   const struct tgsi_token *tokens,
1722                   struct lp_type type,
1723                   struct lp_build_mask_context *mask,
1724                   LLVMValueRef consts_ptr,
1725                   const LLVMValueRef *pos,
1726                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1727                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1728                   struct lp_build_sampler_soa *sampler,
1729                   struct tgsi_shader_info *info)
1730 {
1731    struct lp_build_tgsi_soa_context bld;
1732    struct tgsi_parse_context parse;
1733    uint num_immediates = 0;
1734    unsigned i;
1735
1736    /* Setup build context */
1737    memset(&bld, 0, sizeof bld);
1738    lp_build_context_init(&bld.base, builder, type);
1739    bld.mask = mask;
1740    bld.pos = pos;
1741    bld.inputs = inputs;
1742    bld.outputs = outputs;
1743    bld.consts_ptr = consts_ptr;
1744    bld.sampler = sampler;
1745    bld.has_indirect_addressing = info->opcode_count[TGSI_OPCODE_ARR] > 0 ||
1746                                  info->opcode_count[TGSI_OPCODE_ARL] > 0;
1747
1748    lp_exec_mask_init(&bld.exec_mask, &bld.base);
1749
1750    tgsi_parse_init( &parse, tokens );
1751
1752    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1753       tgsi_parse_token( &parse );
1754
1755       switch( parse.FullToken.Token.Type ) {
1756       case TGSI_TOKEN_TYPE_DECLARATION:
1757          /* Inputs already interpolated */
1758          {
1759             if (!emit_declaration( &bld, &parse.FullToken.FullDeclaration ))
1760                _debug_printf("warning: failed to define LLVM variable\n");
1761          }
1762          break;
1763
1764       case TGSI_TOKEN_TYPE_INSTRUCTION:
1765          {
1766             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1767             const struct tgsi_opcode_info *opcode_info = tgsi_get_opcode_info(opcode);
1768             if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, opcode_info ))
1769                _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1770                              opcode_info->mnemonic);
1771          }
1772
1773          break;
1774
1775       case TGSI_TOKEN_TYPE_IMMEDIATE:
1776          /* simply copy the immediate values into the next immediates[] slot */
1777          {
1778             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1779             assert(size <= 4);
1780             assert(num_immediates < LP_MAX_IMMEDIATES);
1781             for( i = 0; i < size; ++i )
1782                bld.immediates[num_immediates][i] =
1783                   lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
1784             for( i = size; i < 4; ++i )
1785                bld.immediates[num_immediates][i] = bld.base.undef;
1786             num_immediates++;
1787          }
1788          break;
1789
1790       case TGSI_TOKEN_TYPE_PROPERTY:
1791          break;
1792
1793       default:
1794          assert( 0 );
1795       }
1796    }
1797    if (0) {
1798       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1799       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1800       debug_printf("11111111111111111111111111111 \n");
1801       tgsi_dump(tokens, 0);
1802       LLVMDumpValue(function);
1803       debug_printf("2222222222222222222222222222 \n");
1804    }
1805    tgsi_parse_free( &parse );
1806 }
1807