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       break;
 583
 584    default:
 585       assert( 0 );
 586    }
 587 }
 588
 589
 590 /**
 591  * High-level instruction translators.
 592  */
 593
 594
 595 static void
 596 emit_tex( struct lp_build_tgsi_soa_context *bld,
 597           const struct tgsi_full_instruction *inst,
 598           boolean apply_lodbias,
 599           boolean projected,
 600           LLVMValueRef *texel)
 601 {
 602    const uint unit = inst->Src[1].Register.Index;
 603    LLVMValueRef lodbias;
 604    LLVMValueRef oow = NULL;
 605    LLVMValueRef coords[3];
 606    unsigned num_coords;
 607    unsigned i;
 608
 609    if (!bld->sampler) {
 610       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 611       for (i = 0; i < 4; i++) {
 612          texel[i] = bld->base.undef;
 613       }
 614       return;
 615    }
 616
 617    switch (inst->Texture.Texture) {
 618    case TGSI_TEXTURE_1D:
 619       num_coords = 1;
 620       break;
 621    case TGSI_TEXTURE_2D:
 622    case TGSI_TEXTURE_RECT:
 623       num_coords = 2;
 624       break;
 625    case TGSI_TEXTURE_SHADOW1D:
 626    case TGSI_TEXTURE_SHADOW2D:
 627    case TGSI_TEXTURE_SHADOWRECT:
 628    case TGSI_TEXTURE_3D:
 629    case TGSI_TEXTURE_CUBE:
 630       num_coords = 3;
 631       break;
 632    default:
 633       assert(0);
 634       return;
 635    }
 636
 637    if(apply_lodbias)
 638       lodbias = emit_fetch( bld, inst, 0, 3 );
 639    else
 640       lodbias = bld->base.zero;
 641
 642    if (projected) {
 643       oow = emit_fetch( bld, inst, 0, 3 );
 644       oow = lp_build_rcp(&bld->base, oow);
 645    }
 646
 647    for (i = 0; i < num_coords; i++) {
 648       coords[i] = emit_fetch( bld, inst, 0, i );
 649       if (projected)
 650          coords[i] = lp_build_mul(&bld->base, coords[i], oow);
 651    }
 652    for (i = num_coords; i < 3; i++) {
 653       coords[i] = bld->base.undef;
 654    }
 655
 656    bld->sampler->emit_fetch_texel(bld->sampler,
 657                                   bld->base.builder,
 658                                   bld->base.type,
 659                                   unit, num_coords, coords, lodbias,
 660                                   texel);
 661 }
 662
 663
 664 /**
 665  * Kill fragment if any of the src register values are negative.
 666  */
 667 static void
 668 emit_kil(
 669    struct lp_build_tgsi_soa_context *bld,
 670    const struct tgsi_full_instruction *inst )
 671 {
 672    const struct tgsi_full_src_register *reg = &inst->Src[0];
 673    LLVMValueRef terms[NUM_CHANNELS];
 674    LLVMValueRef mask;
 675    unsigned chan_index;
 676
 677    memset(&terms, 0, sizeof terms);
 678
 679    FOR_EACH_CHANNEL( chan_index ) {
 680       unsigned swizzle;
 681
 682       /* Unswizzle channel */
 683       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 684
 685       /* Check if the component has not been already tested. */
 686       assert(swizzle < NUM_CHANNELS);
 687       if( !terms[swizzle] )
 688          /* TODO: change the comparison operator instead of setting the sign */
 689          terms[swizzle] =  emit_fetch(bld, inst, 0, chan_index );
 690    }
 691
 692    mask = NULL;
 693    FOR_EACH_CHANNEL( chan_index ) {
 694       if(terms[chan_index]) {
 695          LLVMValueRef chan_mask;
 696
 697          /*
 698           * If term < 0 then mask = 0 else mask = ~0.
 699           */
 700          chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero);
 701
 702          if(mask)
 703             mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, "");
 704          else
 705             mask = chan_mask;
 706       }
 707    }
 708
 709    if(mask)
 710       lp_build_mask_update(bld->mask, mask);
 711 }
 712
 713
 714 /**
 715  * Predicated fragment kill.
 716  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
 717  * The only predication is the execution mask which will apply if
 718  * we're inside a loop or conditional.
 719  */
 720 static void
 721 emit_kilp(struct lp_build_tgsi_soa_context *bld,
 722           const struct tgsi_full_instruction *inst)
 723 {
 724    LLVMValueRef mask;
 725
 726    /* For those channels which are "alive", disable fragment shader
 727     * execution.
 728     */
 729    if (bld->exec_mask.has_mask) {
 730       mask = LLVMBuildNot(bld->base.builder, bld->exec_mask.exec_mask, "kilp");
 731    }
 732    else {
 733       mask = bld->base.zero;
 734    }
 735
 736    lp_build_mask_update(bld->mask, mask);
 737 }
 738
 739 static void
 740 emit_declaration(
 741    struct lp_build_tgsi_soa_context *bld,
 742    const struct tgsi_full_declaration *decl)
 743 {
 744    LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
 745
 746    unsigned first = decl->Range.First;
 747    unsigned last = decl->Range.Last;
 748    unsigned idx, i;
 749
 750    for (idx = first; idx <= last; ++idx) {
 751       switch (decl->Declaration.File) {
 752       case TGSI_FILE_TEMPORARY:
 753          if (bld->has_indirect_addressing) {
 754             LLVMValueRef val = LLVMConstInt(LLVMInt32Type(),
 755                                             last*4 + 4, 0);
 756             bld->temps_array = lp_build_array_alloca(bld->base.builder,
 757                                                      vec_type, val, "");
 758          } else {
 759             for (i = 0; i < NUM_CHANNELS; i++)
 760                bld->temps[idx][i] = lp_build_alloca(bld->base.builder,
 761                                                     vec_type, "");
 762          }
 763          break;
 764
 765       case TGSI_FILE_OUTPUT:
 766          for (i = 0; i < NUM_CHANNELS; i++)
 767             bld->outputs[idx][i] = lp_build_alloca(bld->base.builder,
 768                                                    vec_type, "");
 769          break;
 770
 771       case TGSI_FILE_ADDRESS:
 772          for (i = 0; i < NUM_CHANNELS; i++)
 773             bld->addr[idx][i] = lp_build_alloca(bld->base.builder,
 774                                                 vec_type, "");
 775          break;
 776
 777       case TGSI_FILE_PREDICATE:
 778          _debug_printf("warning: predicate registers not yet implemented\n");
 779          break;
 780
 781       default:
 782          /* don't need to declare other vars */
 783          break;
 784       }
 785    }
 786 }
 787
 788
 789 /**
 790  * Emit LLVM for one TGSI instruction.
 791  * \param return TRUE for success, FALSE otherwise
 792  */
 793 static boolean
 794 emit_instruction(
 795    struct lp_build_tgsi_soa_context *bld,
 796    const struct tgsi_full_instruction *inst,
 797    const struct tgsi_opcode_info *info)
 798 {
 799    unsigned chan_index;
 800    LLVMValueRef src0, src1, src2;
 801    LLVMValueRef tmp0, tmp1, tmp2;
 802    LLVMValueRef tmp3 = NULL;
 803    LLVMValueRef tmp4 = NULL;
 804    LLVMValueRef tmp5 = NULL;
 805    LLVMValueRef tmp6 = NULL;
 806    LLVMValueRef tmp7 = NULL;
 807    LLVMValueRef res;
 808    LLVMValueRef dst0[NUM_CHANNELS];
 809
 810    /*
 811     * Stores and write masks are handled in a general fashion after the long
 812     * instruction opcode switch statement.
 813     *
 814     * Although not stricitly necessary, we avoid generating instructions for
 815     * channels which won't be stored, in cases where's that easy. For some
 816     * complex instructions, like texture sampling, it is more convenient to
 817     * assume a full writemask and then let LLVM optimization passes eliminate
 818     * redundant code.
 819     */
 820
 821    assert(info->num_dst <= 1);
 822    if(info->num_dst) {
 823       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 824          dst0[chan_index] = bld->base.undef;
 825       }
 826    }
 827
 828    switch (inst->Instruction.Opcode) {
 829    case TGSI_OPCODE_ARL:
 830       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 831          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 832          tmp0 = lp_build_floor(&bld->base, tmp0);
 833          dst0[chan_index] = tmp0;
 834       }
 835       break;
 836
 837    case TGSI_OPCODE_MOV:
 838       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 839          dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
 840       }
 841       break;
 842
 843    case TGSI_OPCODE_LIT:
 844       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 845          dst0[CHAN_X] = bld->base.one;
 846       }
 847       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 848          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 849          dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
 850       }
 851       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 852          /* XMM[1] = SrcReg[0].yyyy */
 853          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 854          /* XMM[1] = max(XMM[1], 0) */
 855          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 856          /* XMM[2] = SrcReg[0].wwww */
 857          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 858          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 859          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 860          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 861          dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 862       }
 863       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 864          dst0[CHAN_W] = bld->base.one;
 865       }
 866       break;
 867
 868    case TGSI_OPCODE_RCP:
 869    /* TGSI_OPCODE_RECIP */
 870       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 871       res = lp_build_rcp(&bld->base, src0);
 872       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 873          dst0[chan_index] = res;
 874       }
 875       break;
 876
 877    case TGSI_OPCODE_RSQ:
 878    /* TGSI_OPCODE_RECIPSQRT */
 879       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 880       src0 = lp_build_abs(&bld->base, src0);
 881       res = lp_build_rsqrt(&bld->base, src0);
 882       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 883          dst0[chan_index] = res;
 884       }
 885       break;
 886
 887    case TGSI_OPCODE_EXP:
 888       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 889           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 890           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 891          LLVMValueRef *p_exp2_int_part = NULL;
 892          LLVMValueRef *p_frac_part = NULL;
 893          LLVMValueRef *p_exp2 = NULL;
 894
 895          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 896
 897          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 898             p_exp2_int_part = &tmp0;
 899          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 900             p_frac_part = &tmp1;
 901          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 902             p_exp2 = &tmp2;
 903
 904          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 905
 906          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 907             dst0[CHAN_X] = tmp0;
 908          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 909             dst0[CHAN_Y] = tmp1;
 910          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 911             dst0[CHAN_Z] = tmp2;
 912       }
 913       /* dst.w = 1.0 */
 914       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 915          dst0[CHAN_W] = bld->base.one;
 916       }
 917       break;
 918
 919    case TGSI_OPCODE_LOG:
 920       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 921           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 922           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 923          LLVMValueRef *p_floor_log2 = NULL;
 924          LLVMValueRef *p_exp = NULL;
 925          LLVMValueRef *p_log2 = NULL;
 926
 927          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 928          src0 = lp_build_abs( &bld->base, src0 );
 929
 930          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 931             p_floor_log2 = &tmp0;
 932          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 933             p_exp = &tmp1;
 934          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 935             p_log2 = &tmp2;
 936
 937          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 938
 939          /* dst.x = floor(lg2(abs(src.x))) */
 940          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 941             dst0[CHAN_X] = tmp0;
 942          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 943          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 944             dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
 945          }
 946          /* dst.z = lg2(abs(src.x)) */
 947          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 948             dst0[CHAN_Z] = tmp2;
 949       }
 950       /* dst.w = 1.0 */
 951       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 952          dst0[CHAN_W] = bld->base.one;
 953       }
 954       break;
 955
 956    case TGSI_OPCODE_MUL:
 957       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 958          src0 = emit_fetch( bld, inst, 0, chan_index );
 959          src1 = emit_fetch( bld, inst, 1, chan_index );
 960          dst0[chan_index] = lp_build_mul(&bld->base, src0, src1);
 961       }
 962       break;
 963
 964    case TGSI_OPCODE_ADD:
 965       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 966          src0 = emit_fetch( bld, inst, 0, chan_index );
 967          src1 = emit_fetch( bld, inst, 1, chan_index );
 968          dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
 969       }
 970       break;
 971
 972    case TGSI_OPCODE_DP3:
 973    /* TGSI_OPCODE_DOT3 */
 974       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 975       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 976       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 977       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 978       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 979       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 980       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 981       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 982       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 983       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 984       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 985       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 986          dst0[chan_index] = tmp0;
 987       }
 988       break;
 989
 990    case TGSI_OPCODE_DP4:
 991    /* TGSI_OPCODE_DOT4 */
 992       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 993       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 994       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 995       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 996       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 997       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 998       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 999       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1000       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1001       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1002       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1003       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
1004       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
1005       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1006       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1007       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1008          dst0[chan_index] = tmp0;
1009       }
1010       break;
1011
1012    case TGSI_OPCODE_DST:
1013       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1014          dst0[CHAN_X] = bld->base.one;
1015       }
1016       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1017          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1018          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
1019          dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
1020       }
1021       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1022          dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
1023       }
1024       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1025          dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
1026       }
1027       break;
1028
1029    case TGSI_OPCODE_MIN:
1030       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1031          src0 = emit_fetch( bld, inst, 0, chan_index );
1032          src1 = emit_fetch( bld, inst, 1, chan_index );
1033          dst0[chan_index] = lp_build_min( &bld->base, src0, src1 );
1034       }
1035       break;
1036
1037    case TGSI_OPCODE_MAX:
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          dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
1042       }
1043       break;
1044
1045    case TGSI_OPCODE_SLT:
1046    /* TGSI_OPCODE_SETLT */
1047       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1048          src0 = emit_fetch( bld, inst, 0, chan_index );
1049          src1 = emit_fetch( bld, inst, 1, chan_index );
1050          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
1051          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1052       }
1053       break;
1054
1055    case TGSI_OPCODE_SGE:
1056    /* TGSI_OPCODE_SETGE */
1057       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1058          src0 = emit_fetch( bld, inst, 0, chan_index );
1059          src1 = emit_fetch( bld, inst, 1, chan_index );
1060          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
1061          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1062       }
1063       break;
1064
1065    case TGSI_OPCODE_MAD:
1066    /* TGSI_OPCODE_MADD */
1067       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1068          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1069          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1070          tmp2 = emit_fetch( bld, inst, 2, chan_index );
1071          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1072          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
1073          dst0[chan_index] = tmp0;
1074       }
1075       break;
1076
1077    case TGSI_OPCODE_SUB:
1078       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1079          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1080          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1081          dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
1082       }
1083       break;
1084
1085    case TGSI_OPCODE_LRP:
1086       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1087          src0 = emit_fetch( bld, inst, 0, chan_index );
1088          src1 = emit_fetch( bld, inst, 1, chan_index );
1089          src2 = emit_fetch( bld, inst, 2, chan_index );
1090          tmp0 = lp_build_sub( &bld->base, src1, src2 );
1091          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
1092          dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
1093       }
1094       break;
1095
1096    case TGSI_OPCODE_CND:
1097       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1098          src0 = emit_fetch( bld, inst, 0, chan_index );
1099          src1 = emit_fetch( bld, inst, 1, chan_index );
1100          src2 = emit_fetch( bld, inst, 2, chan_index );
1101          tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1102          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1103          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1104       }
1105       break;
1106
1107    case TGSI_OPCODE_DP2A:
1108       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1109       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1110       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1111       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1112       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1113       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1114       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1115       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
1116       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1117       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1118          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1119       }
1120       break;
1121
1122    case TGSI_OPCODE_FRC:
1123       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1124          src0 = emit_fetch( bld, inst, 0, chan_index );
1125          tmp0 = lp_build_floor(&bld->base, src0);
1126          tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1127          dst0[chan_index] = tmp0;
1128       }
1129       break;
1130
1131    case TGSI_OPCODE_CLAMP:
1132       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1133          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1134          src1 = emit_fetch( bld, inst, 1, chan_index );
1135          src2 = emit_fetch( bld, inst, 2, chan_index );
1136          tmp0 = lp_build_max(&bld->base, tmp0, src1);
1137          tmp0 = lp_build_min(&bld->base, tmp0, src2);
1138          dst0[chan_index] = tmp0;
1139       }
1140       break;
1141
1142    case TGSI_OPCODE_FLR:
1143       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1144          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1145          dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1146       }
1147       break;
1148
1149    case TGSI_OPCODE_ROUND:
1150       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1151          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1152          dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1153       }
1154       break;
1155
1156    case TGSI_OPCODE_EX2: {
1157       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1158       tmp0 = lp_build_exp2( &bld->base, tmp0);
1159       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1160          dst0[chan_index] = tmp0;
1161       }
1162       break;
1163    }
1164
1165    case TGSI_OPCODE_LG2:
1166       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1167       tmp0 = lp_build_log2( &bld->base, tmp0);
1168       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1169          dst0[chan_index] = tmp0;
1170       }
1171       break;
1172
1173    case TGSI_OPCODE_POW:
1174       src0 = emit_fetch( bld, inst, 0, CHAN_X );
1175       src1 = emit_fetch( bld, inst, 1, CHAN_X );
1176       res = lp_build_pow( &bld->base, src0, src1 );
1177       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1178          dst0[chan_index] = res;
1179       }
1180       break;
1181
1182    case TGSI_OPCODE_XPD:
1183       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1184           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1185          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1186          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1187       }
1188       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1189           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1190          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1191          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1192       }
1193       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1194          tmp2 = tmp0;
1195          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1196          tmp5 = tmp3;
1197          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1198          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1199          dst0[CHAN_X] = tmp2;
1200       }
1201       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1202           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1203          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1204          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1205       }
1206       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1207          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1208          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1209          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1210          dst0[CHAN_Y] = tmp3;
1211       }
1212       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1213          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1214          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1215          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1216          dst0[CHAN_Z] = tmp5;
1217       }
1218       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1219          dst0[CHAN_W] = bld->base.one;
1220       }
1221       break;
1222
1223    case TGSI_OPCODE_ABS:
1224       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1225          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1226          dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1227       }
1228       break;
1229
1230    case TGSI_OPCODE_RCC:
1231       /* deprecated? */
1232       assert(0);
1233       return FALSE;
1234
1235    case TGSI_OPCODE_DPH:
1236       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1237       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1238       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1239       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1240       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1241       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1242       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1243       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1244       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1245       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1246       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1247       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1248       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1249       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1250          dst0[chan_index] = tmp0;
1251       }
1252       break;
1253
1254    case TGSI_OPCODE_COS:
1255       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1256       tmp0 = lp_build_cos( &bld->base, tmp0 );
1257       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1258          dst0[chan_index] = tmp0;
1259       }
1260       break;
1261
1262    case TGSI_OPCODE_DDX:
1263       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1264          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1265       }
1266       break;
1267
1268    case TGSI_OPCODE_DDY:
1269       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1270          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1271       }
1272       break;
1273
1274    case TGSI_OPCODE_KILP:
1275       /* predicated kill */
1276       emit_kilp( bld, inst );
1277       break;
1278
1279    case TGSI_OPCODE_KIL:
1280       /* conditional kill */
1281       emit_kil( bld, inst );
1282       break;
1283
1284    case TGSI_OPCODE_PK2H:
1285       return FALSE;
1286       break;
1287
1288    case TGSI_OPCODE_PK2US:
1289       return FALSE;
1290       break;
1291
1292    case TGSI_OPCODE_PK4B:
1293       return FALSE;
1294       break;
1295
1296    case TGSI_OPCODE_PK4UB:
1297       return FALSE;
1298       break;
1299
1300    case TGSI_OPCODE_RFL:
1301       return FALSE;
1302       break;
1303
1304    case TGSI_OPCODE_SEQ:
1305       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1306          src0 = emit_fetch( bld, inst, 0, chan_index );
1307          src1 = emit_fetch( bld, inst, 1, chan_index );
1308          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1309          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1310       }
1311       break;
1312
1313    case TGSI_OPCODE_SFL:
1314       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1315          dst0[chan_index] = bld->base.zero;
1316       }
1317       break;
1318
1319    case TGSI_OPCODE_SGT:
1320       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1321          src0 = emit_fetch( bld, inst, 0, chan_index );
1322          src1 = emit_fetch( bld, inst, 1, chan_index );
1323          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1324          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1325       }
1326       break;
1327
1328    case TGSI_OPCODE_SIN:
1329       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1330       tmp0 = lp_build_sin( &bld->base, tmp0 );
1331       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1332          dst0[chan_index] = tmp0;
1333       }
1334       break;
1335
1336    case TGSI_OPCODE_SLE:
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_LEQUAL, 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_SNE:
1346       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1347          src0 = emit_fetch( bld, inst, 0, chan_index );
1348          src1 = emit_fetch( bld, inst, 1, chan_index );
1349          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1350          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1351       }
1352       break;
1353
1354    case TGSI_OPCODE_STR:
1355       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1356          dst0[chan_index] = bld->base.one;
1357       }
1358       break;
1359
1360    case TGSI_OPCODE_TEX:
1361       emit_tex( bld, inst, FALSE, FALSE, dst0 );
1362       break;
1363
1364    case TGSI_OPCODE_TXD:
1365       /* FIXME */
1366       return FALSE;
1367       break;
1368
1369    case TGSI_OPCODE_UP2H:
1370       /* deprecated */
1371       assert (0);
1372       return FALSE;
1373       break;
1374
1375    case TGSI_OPCODE_UP2US:
1376       /* deprecated */
1377       assert(0);
1378       return FALSE;
1379       break;
1380
1381    case TGSI_OPCODE_UP4B:
1382       /* deprecated */
1383       assert(0);
1384       return FALSE;
1385       break;
1386
1387    case TGSI_OPCODE_UP4UB:
1388       /* deprecated */
1389       assert(0);
1390       return FALSE;
1391       break;
1392
1393    case TGSI_OPCODE_X2D:
1394       /* deprecated? */
1395       assert(0);
1396       return FALSE;
1397       break;
1398
1399    case TGSI_OPCODE_ARA:
1400       /* deprecated */
1401       assert(0);
1402       return FALSE;
1403       break;
1404
1405    case TGSI_OPCODE_ARR:
1406       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1407          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1408          tmp0 = lp_build_round(&bld->base, tmp0);
1409          dst0[chan_index] = tmp0;
1410       }
1411       break;
1412
1413    case TGSI_OPCODE_BRA:
1414       /* deprecated */
1415       assert(0);
1416       return FALSE;
1417       break;
1418
1419    case TGSI_OPCODE_CAL:
1420       /* FIXME */
1421       return FALSE;
1422       break;
1423
1424    case TGSI_OPCODE_RET:
1425       /* FIXME */
1426       return FALSE;
1427       break;
1428
1429    case TGSI_OPCODE_END:
1430       break;
1431
1432    case TGSI_OPCODE_SSG:
1433    /* TGSI_OPCODE_SGN */
1434       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1435          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1436          dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1437       }
1438       break;
1439
1440    case TGSI_OPCODE_CMP:
1441       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1442          src0 = emit_fetch( bld, inst, 0, chan_index );
1443          src1 = emit_fetch( bld, inst, 1, chan_index );
1444          src2 = emit_fetch( bld, inst, 2, chan_index );
1445          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1446          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1447       }
1448       break;
1449
1450    case TGSI_OPCODE_SCS:
1451       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1452          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1453          dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1454       }
1455       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1456          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1457          dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1458       }
1459       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1460          dst0[CHAN_Z] = bld->base.zero;
1461       }
1462       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1463          dst0[CHAN_W] = bld->base.one;
1464       }
1465       break;
1466
1467    case TGSI_OPCODE_TXB:
1468       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1469       break;
1470
1471    case TGSI_OPCODE_NRM:
1472       /* fall-through */
1473    case TGSI_OPCODE_NRM4:
1474       /* 3 or 4-component normalization */
1475       {
1476          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1477
1478          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1479              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1480              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1481              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1482
1483             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1484
1485             /* xmm4 = src.x */
1486             /* xmm0 = src.x * src.x */
1487             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1488             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1489                tmp4 = tmp0;
1490             }
1491             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1492
1493             /* xmm5 = src.y */
1494             /* xmm0 = xmm0 + src.y * src.y */
1495             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1496             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1497                tmp5 = tmp1;
1498             }
1499             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1500             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1501
1502             /* xmm6 = src.z */
1503             /* xmm0 = xmm0 + src.z * src.z */
1504             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1505             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1506                tmp6 = tmp1;
1507             }
1508             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1509             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1510
1511             if (dims == 4) {
1512                /* xmm7 = src.w */
1513                /* xmm0 = xmm0 + src.w * src.w */
1514                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1515                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1516                   tmp7 = tmp1;
1517                }
1518                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1519                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1520             }
1521
1522             /* xmm1 = 1 / sqrt(xmm0) */
1523             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1524
1525             /* dst.x = xmm1 * src.x */
1526             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1527                dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1528             }
1529
1530             /* dst.y = xmm1 * src.y */
1531             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1532                dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1533             }
1534
1535             /* dst.z = xmm1 * src.z */
1536             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1537                dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1538             }
1539
1540             /* dst.w = xmm1 * src.w */
1541             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1542                dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1543             }
1544          }
1545
1546          /* dst.w = 1.0 */
1547          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1548             dst0[CHAN_W] = bld->base.one;
1549          }
1550       }
1551       break;
1552
1553    case TGSI_OPCODE_DIV:
1554       /* deprecated */
1555       assert( 0 );
1556       return FALSE;
1557       break;
1558
1559    case TGSI_OPCODE_DP2:
1560       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1561       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1562       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1563       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1564       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1565       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1566       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1567       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1568          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1569       }
1570       break;
1571
1572    case TGSI_OPCODE_TXL:
1573       emit_tex( bld, inst, TRUE, FALSE, dst0 );
1574       break;
1575
1576    case TGSI_OPCODE_TXP:
1577       emit_tex( bld, inst, FALSE, TRUE, dst0 );
1578       break;
1579
1580    case TGSI_OPCODE_BRK:
1581       lp_exec_break(&bld->exec_mask);
1582       break;
1583
1584    case TGSI_OPCODE_IF:
1585       tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1586       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1587                           tmp0, bld->base.zero);
1588       lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1589       break;
1590
1591    case TGSI_OPCODE_BGNLOOP:
1592       lp_exec_bgnloop(&bld->exec_mask);
1593       break;
1594
1595    case TGSI_OPCODE_ELSE:
1596       lp_exec_mask_cond_invert(&bld->exec_mask);
1597       break;
1598
1599    case TGSI_OPCODE_ENDIF:
1600       lp_exec_mask_cond_pop(&bld->exec_mask);
1601       break;
1602
1603    case TGSI_OPCODE_ENDLOOP:
1604       lp_exec_endloop(&bld->exec_mask);
1605       break;
1606
1607    case TGSI_OPCODE_PUSHA:
1608       /* deprecated? */
1609       assert(0);
1610       return FALSE;
1611       break;
1612
1613    case TGSI_OPCODE_POPA:
1614       /* deprecated? */
1615       assert(0);
1616       return FALSE;
1617       break;
1618
1619    case TGSI_OPCODE_CEIL:
1620       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1621          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1622          dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1623       }
1624       break;
1625
1626    case TGSI_OPCODE_I2F:
1627       /* deprecated? */
1628       assert(0);
1629       return FALSE;
1630       break;
1631
1632    case TGSI_OPCODE_NOT:
1633       /* deprecated? */
1634       assert(0);
1635       return FALSE;
1636       break;
1637
1638    case TGSI_OPCODE_TRUNC:
1639       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1640          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1641          dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1642       }
1643       break;
1644
1645    case TGSI_OPCODE_SHL:
1646       /* deprecated? */
1647       assert(0);
1648       return FALSE;
1649       break;
1650
1651    case TGSI_OPCODE_ISHR:
1652       /* deprecated? */
1653       assert(0);
1654       return FALSE;
1655       break;
1656
1657    case TGSI_OPCODE_AND:
1658       /* deprecated? */
1659       assert(0);
1660       return FALSE;
1661       break;
1662
1663    case TGSI_OPCODE_OR:
1664       /* deprecated? */
1665       assert(0);
1666       return FALSE;
1667       break;
1668
1669    case TGSI_OPCODE_MOD:
1670       /* deprecated? */
1671       assert(0);
1672       return FALSE;
1673       break;
1674
1675    case TGSI_OPCODE_XOR:
1676       /* deprecated? */
1677       assert(0);
1678       return FALSE;
1679       break;
1680
1681    case TGSI_OPCODE_SAD:
1682       /* deprecated? */
1683       assert(0);
1684       return FALSE;
1685       break;
1686
1687    case TGSI_OPCODE_TXF:
1688       /* deprecated? */
1689       assert(0);
1690       return FALSE;
1691       break;
1692
1693    case TGSI_OPCODE_TXQ:
1694       /* deprecated? */
1695       assert(0);
1696       return FALSE;
1697       break;
1698
1699    case TGSI_OPCODE_CONT:
1700       lp_exec_continue(&bld->exec_mask);
1701       break;
1702
1703    case TGSI_OPCODE_EMIT:
1704       return FALSE;
1705       break;
1706
1707    case TGSI_OPCODE_ENDPRIM:
1708       return FALSE;
1709       break;
1710
1711    case TGSI_OPCODE_NOP:
1712       break;
1713
1714    default:
1715       return FALSE;
1716    }
1717
1718    if(info->num_dst) {
1719       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1720          emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
1721       }
1722    }
1723
1724    return TRUE;
1725 }
1726
1727
1728 void
1729 lp_build_tgsi_soa(LLVMBuilderRef builder,
1730                   const struct tgsi_token *tokens,
1731                   struct lp_type type,
1732                   struct lp_build_mask_context *mask,
1733                   LLVMValueRef consts_ptr,
1734                   const LLVMValueRef *pos,
1735                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1736                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1737                   struct lp_build_sampler_soa *sampler,
1738                   struct tgsi_shader_info *info)
1739 {
1740    struct lp_build_tgsi_soa_context bld;
1741    struct tgsi_parse_context parse;
1742    uint num_immediates = 0;
1743    unsigned i;
1744
1745    /* Setup build context */
1746    memset(&bld, 0, sizeof bld);
1747    lp_build_context_init(&bld.base, builder, type);
1748    bld.mask = mask;
1749    bld.pos = pos;
1750    bld.inputs = inputs;
1751    bld.outputs = outputs;
1752    bld.consts_ptr = consts_ptr;
1753    bld.sampler = sampler;
1754    bld.has_indirect_addressing = info->opcode_count[TGSI_OPCODE_ARR] > 0 ||
1755                                  info->opcode_count[TGSI_OPCODE_ARL] > 0;
1756
1757    lp_exec_mask_init(&bld.exec_mask, &bld.base);
1758
1759    tgsi_parse_init( &parse, tokens );
1760
1761    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1762       tgsi_parse_token( &parse );
1763
1764       switch( parse.FullToken.Token.Type ) {
1765       case TGSI_TOKEN_TYPE_DECLARATION:
1766          /* Inputs already interpolated */
1767          emit_declaration( &bld, &parse.FullToken.FullDeclaration );
1768          break;
1769
1770       case TGSI_TOKEN_TYPE_INSTRUCTION:
1771          {
1772             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1773             const struct tgsi_opcode_info *opcode_info = tgsi_get_opcode_info(opcode);
1774             if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, opcode_info ))
1775                _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1776                              opcode_info->mnemonic);
1777          }
1778
1779          break;
1780
1781       case TGSI_TOKEN_TYPE_IMMEDIATE:
1782          /* simply copy the immediate values into the next immediates[] slot */
1783          {
1784             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1785             assert(size <= 4);
1786             assert(num_immediates < LP_MAX_IMMEDIATES);
1787             for( i = 0; i < size; ++i )
1788                bld.immediates[num_immediates][i] =
1789                   lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
1790             for( i = size; i < 4; ++i )
1791                bld.immediates[num_immediates][i] = bld.base.undef;
1792             num_immediates++;
1793          }
1794          break;
1795
1796       case TGSI_TOKEN_TYPE_PROPERTY:
1797          break;
1798
1799       default:
1800          assert( 0 );
1801       }
1802    }
1803    if (0) {
1804       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1805       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1806       debug_printf("11111111111111111111111111111 \n");
1807       tgsi_dump(tokens, 0);
1808       LLVMDumpValue(function);
1809       debug_printf("2222222222222222222222222222 \n");
1810    }
1811    tgsi_parse_free( &parse );
1812 }
1813