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_limits.h"
  58 #include "lp_bld_debug.h"
  59
  60
  61 #define FOR_EACH_CHANNEL( CHAN )\
  62    for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
  63
  64 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  65    ((INST)->Dst[0].Register.WriteMask & (1 << (CHAN)))
  66
  67 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  68    if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
  69
  70 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
  71    FOR_EACH_CHANNEL( CHAN )\
  72       IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
  73
  74 #define CHAN_X 0
  75 #define CHAN_Y 1
  76 #define CHAN_Z 2
  77 #define CHAN_W 3
  78
  79 #define QUAD_TOP_LEFT     0
  80 #define QUAD_TOP_RIGHT    1
  81 #define QUAD_BOTTOM_LEFT  2
  82 #define QUAD_BOTTOM_RIGHT 3
  83
  84
  85 struct lp_exec_mask {
  86    struct lp_build_context *bld;
  87
  88    boolean has_mask;
  89
  90    LLVMTypeRef int_vec_type;
  91
  92    LLVMValueRef cond_stack[LP_MAX_TGSI_NESTING];
  93    int cond_stack_size;
  94    LLVMValueRef cond_mask;
  95
  96    LLVMValueRef break_stack[LP_MAX_TGSI_NESTING];
  97    int break_stack_size;
  98    LLVMValueRef break_mask;
  99
 100    LLVMValueRef cont_stack[LP_MAX_TGSI_NESTING];
 101    int cont_stack_size;
 102    LLVMValueRef cont_mask;
 103
 104    LLVMBasicBlockRef loop_stack[LP_MAX_TGSI_NESTING];
 105    int loop_stack_size;
 106    LLVMBasicBlockRef loop_block;
 107
 108
 109    LLVMValueRef exec_mask;
 110 };
 111
 112 struct lp_build_tgsi_soa_context
 113 {
 114    struct lp_build_context base;
 115
 116    LLVMValueRef consts_ptr;
 117    const LLVMValueRef *pos;
 118    const LLVMValueRef (*inputs)[NUM_CHANNELS];
 119    LLVMValueRef (*outputs)[NUM_CHANNELS];
 120
 121    struct lp_build_sampler_soa *sampler;
 122
 123    LLVMValueRef immediates[LP_MAX_TGSI_IMMEDIATES][NUM_CHANNELS];
 124    LLVMValueRef temps[LP_MAX_TGSI_TEMPS][NUM_CHANNELS];
 125    LLVMValueRef addr[LP_MAX_TGSI_ADDRS][NUM_CHANNELS];
 126
 127    /* we allocate an array of temps if we have indirect
 128     * addressing and then the temps above is unused */
 129    LLVMValueRef temps_array;
 130    boolean has_indirect_addressing;
 131
 132    struct lp_build_mask_context *mask;
 133    struct lp_exec_mask exec_mask;
 134 };
 135
 136 static const unsigned char
 137 swizzle_left[4] = {
 138    QUAD_TOP_LEFT,     QUAD_TOP_LEFT,
 139    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_LEFT
 140 };
 141
 142 static const unsigned char
 143 swizzle_right[4] = {
 144    QUAD_TOP_RIGHT,    QUAD_TOP_RIGHT,
 145    QUAD_BOTTOM_RIGHT, QUAD_BOTTOM_RIGHT
 146 };
 147
 148 static const unsigned char
 149 swizzle_top[4] = {
 150    QUAD_TOP_LEFT,     QUAD_TOP_RIGHT,
 151    QUAD_TOP_LEFT,     QUAD_TOP_RIGHT
 152 };
 153
 154 static const unsigned char
 155 swizzle_bottom[4] = {
 156    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT,
 157    QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT
 158 };
 159
 160 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
 161 {
 162    mask->bld = bld;
 163    mask->has_mask = FALSE;
 164    mask->cond_stack_size = 0;
 165    mask->loop_stack_size = 0;
 166    mask->break_stack_size = 0;
 167    mask->cont_stack_size = 0;
 168
 169    mask->int_vec_type = lp_build_int_vec_type(mask->bld->type);
 170 }
 171
 172 static void lp_exec_mask_update(struct lp_exec_mask *mask)
 173 {
 174    if (mask->loop_stack_size) {
 175       /*for loops we need to update the entire mask at runtime */
 176       LLVMValueRef tmp;
 177       assert(mask->break_mask);
 178       tmp = LLVMBuildAnd(mask->bld->builder,
 179                          mask->cont_mask,
 180                          mask->break_mask,
 181                          "maskcb");
 182       mask->exec_mask = LLVMBuildAnd(mask->bld->builder,
 183                                      mask->cond_mask,
 184                                      tmp,
 185                                      "maskfull");
 186    } else
 187       mask->exec_mask = mask->cond_mask;
 188
 189
 190    mask->has_mask = (mask->cond_stack_size > 0 ||
 191                      mask->loop_stack_size > 0);
 192 }
 193
 194 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 195                                    LLVMValueRef val)
 196 {
 197    assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
 198    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 199    mask->cond_mask = LLVMBuildBitCast(mask->bld->builder, val,
 200                                       mask->int_vec_type, "");
 201
 202    lp_exec_mask_update(mask);
 203 }
 204
 205 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 206 {
 207    LLVMValueRef prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 208    LLVMValueRef inv_mask = LLVMBuildNot(mask->bld->builder,
 209                                         mask->cond_mask, "");
 210
 211    /* means that we didn't have any mask before and that
 212     * we were fully enabled */
 213    if (mask->cond_stack_size <= 1) {
 214       prev_mask = LLVMConstAllOnes(mask->int_vec_type);
 215    }
 216
 217    mask->cond_mask = LLVMBuildAnd(mask->bld->builder,
 218                                   inv_mask,
 219                                   prev_mask, "");
 220    lp_exec_mask_update(mask);
 221 }
 222
 223 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 224 {
 225    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 226    lp_exec_mask_update(mask);
 227 }
 228
 229 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 230 {
 231
 232    if (mask->cont_stack_size == 0)
 233       mask->cont_mask = LLVMConstAllOnes(mask->int_vec_type);
 234    if (mask->break_stack_size == 0)
 235       mask->break_mask = LLVMConstAllOnes(mask->int_vec_type);
 236    if (mask->cond_stack_size == 0)
 237       mask->cond_mask = LLVMConstAllOnes(mask->int_vec_type);
 238
 239    assert(mask->break_stack_size < LP_MAX_TGSI_NESTING);
 240    assert(mask->cont_stack_size < LP_MAX_TGSI_NESTING);
 241    assert(mask->break_stack_size < LP_MAX_TGSI_NESTING);
 242
 243    mask->break_stack[mask->break_stack_size++] = mask->break_mask;
 244    mask->cont_stack[mask->cont_stack_size++] = mask->cont_mask;
 245    mask->loop_stack[mask->loop_stack_size++] = mask->loop_block;
 246    mask->loop_block = lp_build_insert_new_block(mask->bld->builder, "bgnloop");
 247    LLVMBuildBr(mask->bld->builder, mask->loop_block);
 248    LLVMPositionBuilderAtEnd(mask->bld->builder, mask->loop_block);
 249
 250    lp_exec_mask_update(mask);
 251 }
 252
 253 static void lp_exec_break(struct lp_exec_mask *mask)
 254 {
 255    LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
 256                                          mask->exec_mask,
 257                                          "break");
 258
 259    mask->break_mask = LLVMBuildAnd(mask->bld->builder,
 260                                    mask->break_mask,
 261                                    exec_mask, "break_full");
 262
 263    lp_exec_mask_update(mask);
 264 }
 265
 266 static void lp_exec_continue(struct lp_exec_mask *mask)
 267 {
 268    LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
 269                                          mask->exec_mask,
 270                                          "");
 271
 272    mask->cont_mask = LLVMBuildAnd(mask->bld->builder,
 273                                   mask->cont_mask,
 274                                   exec_mask, "");
 275
 276    lp_exec_mask_update(mask);
 277 }
 278
 279
 280 static void lp_exec_endloop(struct lp_exec_mask *mask)
 281 {
 282    LLVMBasicBlockRef endloop;
 283    LLVMTypeRef reg_type = LLVMIntType(mask->bld->type.width*
 284                                       mask->bld->type.length);
 285    LLVMValueRef i1cond;
 286
 287    assert(mask->break_mask);
 288
 289    /* i1cond = (mask == 0) */
 290    i1cond = LLVMBuildICmp(
 291       mask->bld->builder,
 292       LLVMIntNE,
 293       LLVMBuildBitCast(mask->bld->builder, mask->break_mask, reg_type, ""),
 294       LLVMConstNull(reg_type), "");
 295
 296    endloop = lp_build_insert_new_block(mask->bld->builder, "endloop");
 297
 298    LLVMBuildCondBr(mask->bld->builder,
 299                    i1cond, mask->loop_block, endloop);
 300
 301    LLVMPositionBuilderAtEnd(mask->bld->builder, endloop);
 302
 303    mask->loop_block = mask->loop_stack[--mask->loop_stack_size];
 304    /* pop the cont mask */
 305    if (mask->cont_stack_size) {
 306       mask->cont_mask = mask->cont_stack[--mask->cont_stack_size];
 307    }
 308    /* pop the break mask */
 309    if (mask->break_stack_size) {
 310       mask->break_mask = mask->break_stack[--mask->break_stack_size];
 311    }
 312
 313    lp_exec_mask_update(mask);
 314 }
 315
 316 /* stores val into an address pointed to by dst.
 317  * mask->exec_mask is used to figure out which bits of val
 318  * should be stored into the address
 319  * (0 means don't store this bit, 1 means do store).
 320  */
 321 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 322                                LLVMValueRef val,
 323                                LLVMValueRef dst)
 324 {
 325    if (mask->has_mask) {
 326       LLVMValueRef real_val, dst_val;
 327
 328       dst_val = LLVMBuildLoad(mask->bld->builder, dst, "");
 329       real_val = lp_build_select(mask->bld,
 330                                  mask->exec_mask,
 331                                  val, dst_val);
 332
 333       LLVMBuildStore(mask->bld->builder, real_val, dst);
 334    } else
 335       LLVMBuildStore(mask->bld->builder, val, dst);
 336 }
 337
 338
 339 static LLVMValueRef
 340 emit_ddx(struct lp_build_tgsi_soa_context *bld,
 341          LLVMValueRef src)
 342 {
 343    LLVMValueRef src_left  = lp_build_swizzle1_aos(&bld->base, src, swizzle_left);
 344    LLVMValueRef src_right = lp_build_swizzle1_aos(&bld->base, src, swizzle_right);
 345    return lp_build_sub(&bld->base, src_right, src_left);
 346 }
 347
 348
 349 static LLVMValueRef
 350 emit_ddy(struct lp_build_tgsi_soa_context *bld,
 351          LLVMValueRef src)
 352 {
 353    LLVMValueRef src_top    = lp_build_swizzle1_aos(&bld->base, src, swizzle_top);
 354    LLVMValueRef src_bottom = lp_build_swizzle1_aos(&bld->base, src, swizzle_bottom);
 355    return lp_build_sub(&bld->base, src_top, src_bottom);
 356 }
 357
 358 static LLVMValueRef
 359 get_temp_ptr(struct lp_build_tgsi_soa_context *bld,
 360              unsigned index,
 361              unsigned swizzle,
 362              boolean is_indirect,
 363              LLVMValueRef addr)
 364 {
 365    if (!bld->has_indirect_addressing) {
 366       return bld->temps[index][swizzle];
 367    } else {
 368       LLVMValueRef lindex =
 369          LLVMConstInt(LLVMInt32Type(), index*4 + swizzle, 0);
 370       if (is_indirect)
 371          lindex = lp_build_add(&bld->base, lindex, addr);
 372       return LLVMBuildGEP(bld->base.builder, bld->temps_array, &lindex, 1, "");
 373    }
 374 }
 375
 376 /**
 377  * Register fetch.
 378  */
 379 static LLVMValueRef
 380 emit_fetch(
 381    struct lp_build_tgsi_soa_context *bld,
 382    const struct tgsi_full_instruction *inst,
 383    unsigned index,
 384    const unsigned chan_index )
 385 {
 386    const struct tgsi_full_src_register *reg = &inst->Src[index];
 387    unsigned swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 388    LLVMValueRef res;
 389    LLVMValueRef addr;
 390
 391    switch (swizzle) {
 392    case TGSI_SWIZZLE_X:
 393    case TGSI_SWIZZLE_Y:
 394    case TGSI_SWIZZLE_Z:
 395    case TGSI_SWIZZLE_W:
 396
 397       if (reg->Register.Indirect) {
 398          LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->base.type);
 399          unsigned swizzle = tgsi_util_get_src_register_swizzle( &reg->Indirect, chan_index );
 400          addr = LLVMBuildLoad(bld->base.builder,
 401                               bld->addr[reg->Indirect.Index][swizzle],
 402                               "");
 403          /* for indexing we want integers */
 404          addr = LLVMBuildFPToSI(bld->base.builder, addr,
 405                                 int_vec_type, "");
 406          addr = LLVMBuildExtractElement(bld->base.builder,
 407                                         addr, LLVMConstInt(LLVMInt32Type(), 0, 0),
 408                                         "");
 409          addr = lp_build_mul(&bld->base, addr, LLVMConstInt(LLVMInt32Type(), 4, 0));
 410       }
 411
 412       switch (reg->Register.File) {
 413       case TGSI_FILE_CONSTANT: {
 414          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), reg->Register.Index*4 + swizzle, 0);
 415          LLVMValueRef scalar, scalar_ptr;
 416
 417          if (reg->Register.Indirect) {
 418             /*lp_build_printf(bld->base.builder,
 419               "\taddr = %d\n", addr);*/
 420             index = lp_build_add(&bld->base, index, addr);
 421          }
 422          scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr, &index, 1, "");
 423          scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
 424
 425          res = lp_build_broadcast_scalar(&bld->base, scalar);
 426          break;
 427       }
 428
 429       case TGSI_FILE_IMMEDIATE:
 430          res = bld->immediates[reg->Register.Index][swizzle];
 431          assert(res);
 432          break;
 433
 434       case TGSI_FILE_INPUT:
 435          res = bld->inputs[reg->Register.Index][swizzle];
 436          assert(res);
 437          break;
 438
 439       case TGSI_FILE_TEMPORARY: {
 440          LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
 441                                               swizzle,
 442                                               reg->Register.Indirect,
 443                                               addr);
 444          res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
 445          if(!res)
 446             return bld->base.undef;
 447          break;
 448       }
 449
 450       default:
 451          assert( 0 );
 452          return bld->base.undef;
 453       }
 454       break;
 455
 456    default:
 457       assert( 0 );
 458       return bld->base.undef;
 459    }
 460
 461    switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
 462    case TGSI_UTIL_SIGN_CLEAR:
 463       res = lp_build_abs( &bld->base, res );
 464       break;
 465
 466    case TGSI_UTIL_SIGN_SET:
 467       /* TODO: Use bitwese OR for floating point */
 468       res = lp_build_abs( &bld->base, res );
 469       res = LLVMBuildNeg( bld->base.builder, res, "" );
 470       break;
 471
 472    case TGSI_UTIL_SIGN_TOGGLE:
 473       res = LLVMBuildNeg( bld->base.builder, res, "" );
 474       break;
 475
 476    case TGSI_UTIL_SIGN_KEEP:
 477       break;
 478    }
 479
 480    return res;
 481 }
 482
 483
 484 /**
 485  * Register fetch with derivatives.
 486  */
 487 static void
 488 emit_fetch_deriv(
 489    struct lp_build_tgsi_soa_context *bld,
 490    const struct tgsi_full_instruction *inst,
 491    unsigned index,
 492    const unsigned chan_index,
 493    LLVMValueRef *res,
 494    LLVMValueRef *ddx,
 495    LLVMValueRef *ddy)
 496 {
 497    LLVMValueRef src;
 498
 499    src = emit_fetch(bld, inst, index, chan_index);
 500
 501    if(res)
 502       *res = src;
 503
 504    /* TODO: use interpolation coeffs for inputs */
 505
 506    if(ddx)
 507       *ddx = emit_ddx(bld, src);
 508
 509    if(ddy)
 510       *ddy = emit_ddy(bld, src);
 511 }
 512
 513
 514 /**
 515  * Register store.
 516  */
 517 static void
 518 emit_store(
 519    struct lp_build_tgsi_soa_context *bld,
 520    const struct tgsi_full_instruction *inst,
 521    unsigned index,
 522    unsigned chan_index,
 523    LLVMValueRef value)
 524 {
 525    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 526    LLVMValueRef addr;
 527
 528    switch( inst->Instruction.Saturate ) {
 529    case TGSI_SAT_NONE:
 530       break;
 531
 532    case TGSI_SAT_ZERO_ONE:
 533       value = lp_build_max(&bld->base, value, bld->base.zero);
 534       value = lp_build_min(&bld->base, value, bld->base.one);
 535       break;
 536
 537    case TGSI_SAT_MINUS_PLUS_ONE:
 538       value = lp_build_max(&bld->base, value, lp_build_const_vec(bld->base.type, -1.0));
 539       value = lp_build_min(&bld->base, value, bld->base.one);
 540       break;
 541
 542    default:
 543       assert(0);
 544    }
 545
 546    if (reg->Register.Indirect) {
 547       LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->base.type);
 548       unsigned swizzle = tgsi_util_get_src_register_swizzle( &reg->Indirect, chan_index );
 549       addr = LLVMBuildLoad(bld->base.builder,
 550                            bld->addr[reg->Indirect.Index][swizzle],
 551                            "");
 552       /* for indexing we want integers */
 553       addr = LLVMBuildFPToSI(bld->base.builder, addr,
 554                              int_vec_type, "");
 555       addr = LLVMBuildExtractElement(bld->base.builder,
 556                                      addr, LLVMConstInt(LLVMInt32Type(), 0, 0),
 557                                      "");
 558       addr = lp_build_mul(&bld->base, addr, LLVMConstInt(LLVMInt32Type(), 4, 0));
 559    }
 560
 561    switch( reg->Register.File ) {
 562    case TGSI_FILE_OUTPUT:
 563       lp_exec_mask_store(&bld->exec_mask, value,
 564                          bld->outputs[reg->Register.Index][chan_index]);
 565       break;
 566
 567    case TGSI_FILE_TEMPORARY: {
 568       LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
 569                                            chan_index,
 570                                            reg->Register.Indirect,
 571                                            addr);
 572       lp_exec_mask_store(&bld->exec_mask, value, temp_ptr);
 573       break;
 574    }
 575
 576    case TGSI_FILE_ADDRESS:
 577       lp_exec_mask_store(&bld->exec_mask, value,
 578                          bld->addr[reg->Indirect.Index][chan_index]);
 579       break;
 580
 581    case TGSI_FILE_PREDICATE:
 582       /* FIXME */
 583       break;
 584
 585    default:
 586       assert( 0 );
 587    }
 588 }
 589
 590
 591 /**
 592  * High-level instruction translators.
 593  */
 594
 595 enum tex_modifier {
 596    TEX_MODIFIER_NONE = 0,
 597    TEX_MODIFIER_PROJECTED,
 598    TEX_MODIFIER_LOD_BIAS,
 599    TEX_MODIFIER_EXPLICIT_LOD,
 600    TEX_MODIFIER_EXPLICIT_DERIV
 601 };
 602
 603 static void
 604 emit_tex( struct lp_build_tgsi_soa_context *bld,
 605           const struct tgsi_full_instruction *inst,
 606           enum tex_modifier modifier,
 607           LLVMValueRef *texel)
 608 {
 609    unsigned unit;
 610    LLVMValueRef lodbias;
 611    LLVMValueRef oow = NULL;
 612    LLVMValueRef coords[3];
 613    LLVMValueRef ddx[3];
 614    LLVMValueRef ddy[3];
 615    unsigned num_coords;
 616    unsigned i;
 617
 618    if (!bld->sampler) {
 619       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 620       for (i = 0; i < 4; i++) {
 621          texel[i] = bld->base.undef;
 622       }
 623       return;
 624    }
 625
 626    switch (inst->Texture.Texture) {
 627    case TGSI_TEXTURE_1D:
 628       num_coords = 1;
 629       break;
 630    case TGSI_TEXTURE_2D:
 631    case TGSI_TEXTURE_RECT:
 632       num_coords = 2;
 633       break;
 634    case TGSI_TEXTURE_SHADOW1D:
 635    case TGSI_TEXTURE_SHADOW2D:
 636    case TGSI_TEXTURE_SHADOWRECT:
 637    case TGSI_TEXTURE_3D:
 638    case TGSI_TEXTURE_CUBE:
 639       num_coords = 3;
 640       break;
 641    default:
 642       assert(0);
 643       return;
 644    }
 645
 646    /* FIXME: Treat TEX_MODIFIER_EXPLICIT_LOD correctly */
 647    if (modifier == TEX_MODIFIER_LOD_BIAS || modifier == TEX_MODIFIER_EXPLICIT_LOD)
 648       lodbias = emit_fetch( bld, inst, 0, 3 );
 649    else
 650       lodbias = bld->base.zero;
 651
 652    if (modifier == TEX_MODIFIER_PROJECTED) {
 653       oow = emit_fetch( bld, inst, 0, 3 );
 654       oow = lp_build_rcp(&bld->base, oow);
 655    }
 656
 657    for (i = 0; i < num_coords; i++) {
 658       coords[i] = emit_fetch( bld, inst, 0, i );
 659       if (modifier == TEX_MODIFIER_PROJECTED)
 660          coords[i] = lp_build_mul(&bld->base, coords[i], oow);
 661    }
 662    for (i = num_coords; i < 3; i++) {
 663       coords[i] = bld->base.undef;
 664    }
 665
 666    if (modifier == TEX_MODIFIER_EXPLICIT_DERIV) {
 667       for (i = 0; i < num_coords; i++) {
 668          ddx[i] = emit_fetch( bld, inst, 1, i );
 669          ddy[i] = emit_fetch( bld, inst, 2, i );
 670       }
 671       unit = inst->Src[3].Register.Index;
 672    }  else {
 673       for (i = 0; i < num_coords; i++) {
 674          ddx[i] = emit_ddx( bld, coords[i] );
 675          ddy[i] = emit_ddy( bld, coords[i] );
 676       }
 677       unit = inst->Src[1].Register.Index;
 678    }
 679    for (i = num_coords; i < 3; i++) {
 680       ddx[i] = bld->base.undef;
 681       ddy[i] = bld->base.undef;
 682    }
 683
 684    bld->sampler->emit_fetch_texel(bld->sampler,
 685                                   bld->base.builder,
 686                                   bld->base.type,
 687                                   unit, num_coords, coords,
 688                                   ddx, ddy, lodbias,
 689                                   texel);
 690 }
 691
 692
 693 /**
 694  * Kill fragment if any of the src register values are negative.
 695  */
 696 static void
 697 emit_kil(
 698    struct lp_build_tgsi_soa_context *bld,
 699    const struct tgsi_full_instruction *inst )
 700 {
 701    const struct tgsi_full_src_register *reg = &inst->Src[0];
 702    LLVMValueRef terms[NUM_CHANNELS];
 703    LLVMValueRef mask;
 704    unsigned chan_index;
 705
 706    memset(&terms, 0, sizeof terms);
 707
 708    FOR_EACH_CHANNEL( chan_index ) {
 709       unsigned swizzle;
 710
 711       /* Unswizzle channel */
 712       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 713
 714       /* Check if the component has not been already tested. */
 715       assert(swizzle < NUM_CHANNELS);
 716       if( !terms[swizzle] )
 717          /* TODO: change the comparison operator instead of setting the sign */
 718          terms[swizzle] =  emit_fetch(bld, inst, 0, chan_index );
 719    }
 720
 721    mask = NULL;
 722    FOR_EACH_CHANNEL( chan_index ) {
 723       if(terms[chan_index]) {
 724          LLVMValueRef chan_mask;
 725
 726          /*
 727           * If term < 0 then mask = 0 else mask = ~0.
 728           */
 729          chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero);
 730
 731          if(mask)
 732             mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, "");
 733          else
 734             mask = chan_mask;
 735       }
 736    }
 737
 738    if(mask)
 739       lp_build_mask_update(bld->mask, mask);
 740 }
 741
 742
 743 /**
 744  * Predicated fragment kill.
 745  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
 746  * The only predication is the execution mask which will apply if
 747  * we're inside a loop or conditional.
 748  */
 749 static void
 750 emit_kilp(struct lp_build_tgsi_soa_context *bld,
 751           const struct tgsi_full_instruction *inst)
 752 {
 753    LLVMValueRef mask;
 754
 755    /* For those channels which are "alive", disable fragment shader
 756     * execution.
 757     */
 758    if (bld->exec_mask.has_mask) {
 759       mask = LLVMBuildNot(bld->base.builder, bld->exec_mask.exec_mask, "kilp");
 760    }
 761    else {
 762       mask = bld->base.zero;
 763    }
 764
 765    lp_build_mask_update(bld->mask, mask);
 766 }
 767
 768 static void
 769 emit_declaration(
 770    struct lp_build_tgsi_soa_context *bld,
 771    const struct tgsi_full_declaration *decl)
 772 {
 773    LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
 774
 775    unsigned first = decl->Range.First;
 776    unsigned last = decl->Range.Last;
 777    unsigned idx, i;
 778
 779    for (idx = first; idx <= last; ++idx) {
 780       switch (decl->Declaration.File) {
 781       case TGSI_FILE_TEMPORARY:
 782          assert(idx < LP_MAX_TGSI_TEMPS);
 783          if (bld->has_indirect_addressing) {
 784             LLVMValueRef val = LLVMConstInt(LLVMInt32Type(),
 785                                             last*4 + 4, 0);
 786             bld->temps_array = lp_build_array_alloca(bld->base.builder,
 787                                                      vec_type, val, "");
 788          } else {
 789             for (i = 0; i < NUM_CHANNELS; i++)
 790                bld->temps[idx][i] = lp_build_alloca(bld->base.builder,
 791                                                     vec_type, "");
 792          }
 793          break;
 794
 795       case TGSI_FILE_OUTPUT:
 796          for (i = 0; i < NUM_CHANNELS; i++)
 797             bld->outputs[idx][i] = lp_build_alloca(bld->base.builder,
 798                                                    vec_type, "");
 799          break;
 800
 801       case TGSI_FILE_ADDRESS:
 802          assert(idx < LP_MAX_TGSI_ADDRS);
 803          for (i = 0; i < NUM_CHANNELS; i++)
 804             bld->addr[idx][i] = lp_build_alloca(bld->base.builder,
 805                                                 vec_type, "");
 806          break;
 807
 808       case TGSI_FILE_PREDICATE:
 809          _debug_printf("warning: predicate registers not yet implemented\n");
 810          break;
 811
 812       default:
 813          /* don't need to declare other vars */
 814          break;
 815       }
 816    }
 817 }
 818
 819
 820 /**
 821  * Emit LLVM for one TGSI instruction.
 822  * \param return TRUE for success, FALSE otherwise
 823  */
 824 static boolean
 825 emit_instruction(
 826    struct lp_build_tgsi_soa_context *bld,
 827    const struct tgsi_full_instruction *inst,
 828    const struct tgsi_opcode_info *info)
 829 {
 830    unsigned chan_index;
 831    LLVMValueRef src0, src1, src2;
 832    LLVMValueRef tmp0, tmp1, tmp2;
 833    LLVMValueRef tmp3 = NULL;
 834    LLVMValueRef tmp4 = NULL;
 835    LLVMValueRef tmp5 = NULL;
 836    LLVMValueRef tmp6 = NULL;
 837    LLVMValueRef tmp7 = NULL;
 838    LLVMValueRef res;
 839    LLVMValueRef dst0[NUM_CHANNELS];
 840
 841    /*
 842     * Stores and write masks are handled in a general fashion after the long
 843     * instruction opcode switch statement.
 844     *
 845     * Although not stricitly necessary, we avoid generating instructions for
 846     * channels which won't be stored, in cases where's that easy. For some
 847     * complex instructions, like texture sampling, it is more convenient to
 848     * assume a full writemask and then let LLVM optimization passes eliminate
 849     * redundant code.
 850     */
 851
 852    assert(info->num_dst <= 1);
 853    if(info->num_dst) {
 854       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 855          dst0[chan_index] = bld->base.undef;
 856       }
 857    }
 858
 859    switch (inst->Instruction.Opcode) {
 860    case TGSI_OPCODE_ARL:
 861       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 862          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 863          tmp0 = lp_build_floor(&bld->base, tmp0);
 864          dst0[chan_index] = tmp0;
 865       }
 866       break;
 867
 868    case TGSI_OPCODE_MOV:
 869       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 870          dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
 871       }
 872       break;
 873
 874    case TGSI_OPCODE_LIT:
 875       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 876          dst0[CHAN_X] = bld->base.one;
 877       }
 878       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 879          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 880          dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
 881       }
 882       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 883          /* XMM[1] = SrcReg[0].yyyy */
 884          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 885          /* XMM[1] = max(XMM[1], 0) */
 886          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 887          /* XMM[2] = SrcReg[0].wwww */
 888          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 889          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 890          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 891          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 892          dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 893       }
 894       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 895          dst0[CHAN_W] = bld->base.one;
 896       }
 897       break;
 898
 899    case TGSI_OPCODE_RCP:
 900    /* TGSI_OPCODE_RECIP */
 901       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 902       res = lp_build_rcp(&bld->base, src0);
 903       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 904          dst0[chan_index] = res;
 905       }
 906       break;
 907
 908    case TGSI_OPCODE_RSQ:
 909    /* TGSI_OPCODE_RECIPSQRT */
 910       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 911       src0 = lp_build_abs(&bld->base, src0);
 912       res = lp_build_rsqrt(&bld->base, src0);
 913       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 914          dst0[chan_index] = res;
 915       }
 916       break;
 917
 918    case TGSI_OPCODE_EXP:
 919       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 920           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 921           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 922          LLVMValueRef *p_exp2_int_part = NULL;
 923          LLVMValueRef *p_frac_part = NULL;
 924          LLVMValueRef *p_exp2 = NULL;
 925
 926          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 927
 928          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 929             p_exp2_int_part = &tmp0;
 930          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 931             p_frac_part = &tmp1;
 932          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 933             p_exp2 = &tmp2;
 934
 935          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 936
 937          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 938             dst0[CHAN_X] = tmp0;
 939          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 940             dst0[CHAN_Y] = tmp1;
 941          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 942             dst0[CHAN_Z] = tmp2;
 943       }
 944       /* dst.w = 1.0 */
 945       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 946          dst0[CHAN_W] = bld->base.one;
 947       }
 948       break;
 949
 950    case TGSI_OPCODE_LOG:
 951       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 952           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 953           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 954          LLVMValueRef *p_floor_log2 = NULL;
 955          LLVMValueRef *p_exp = NULL;
 956          LLVMValueRef *p_log2 = NULL;
 957
 958          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 959          src0 = lp_build_abs( &bld->base, src0 );
 960
 961          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 962             p_floor_log2 = &tmp0;
 963          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 964             p_exp = &tmp1;
 965          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 966             p_log2 = &tmp2;
 967
 968          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 969
 970          /* dst.x = floor(lg2(abs(src.x))) */
 971          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 972             dst0[CHAN_X] = tmp0;
 973          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 974          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 975             dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
 976          }
 977          /* dst.z = lg2(abs(src.x)) */
 978          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 979             dst0[CHAN_Z] = tmp2;
 980       }
 981       /* dst.w = 1.0 */
 982       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 983          dst0[CHAN_W] = bld->base.one;
 984       }
 985       break;
 986
 987    case TGSI_OPCODE_MUL:
 988       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 989          src0 = emit_fetch( bld, inst, 0, chan_index );
 990          src1 = emit_fetch( bld, inst, 1, chan_index );
 991          dst0[chan_index] = lp_build_mul(&bld->base, src0, src1);
 992       }
 993       break;
 994
 995    case TGSI_OPCODE_ADD:
 996       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 997          src0 = emit_fetch( bld, inst, 0, chan_index );
 998          src1 = emit_fetch( bld, inst, 1, chan_index );
 999          dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
1000       }
1001       break;
1002
1003    case TGSI_OPCODE_DP3:
1004    /* TGSI_OPCODE_DOT3 */
1005       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1006       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1007       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1008       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1009       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1010       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1011       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1012       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1013       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1014       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1015       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1016       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1017          dst0[chan_index] = tmp0;
1018       }
1019       break;
1020
1021    case TGSI_OPCODE_DP4:
1022    /* TGSI_OPCODE_DOT4 */
1023       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1024       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1025       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1026       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1027       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1028       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1029       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1030       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1031       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1032       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1033       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1034       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
1035       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
1036       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1037       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1038       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1039          dst0[chan_index] = tmp0;
1040       }
1041       break;
1042
1043    case TGSI_OPCODE_DST:
1044       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1045          dst0[CHAN_X] = bld->base.one;
1046       }
1047       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1048          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1049          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
1050          dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
1051       }
1052       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1053          dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
1054       }
1055       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1056          dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
1057       }
1058       break;
1059
1060    case TGSI_OPCODE_MIN:
1061       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1062          src0 = emit_fetch( bld, inst, 0, chan_index );
1063          src1 = emit_fetch( bld, inst, 1, chan_index );
1064          dst0[chan_index] = lp_build_min( &bld->base, src0, src1 );
1065       }
1066       break;
1067
1068    case TGSI_OPCODE_MAX:
1069       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1070          src0 = emit_fetch( bld, inst, 0, chan_index );
1071          src1 = emit_fetch( bld, inst, 1, chan_index );
1072          dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
1073       }
1074       break;
1075
1076    case TGSI_OPCODE_SLT:
1077    /* TGSI_OPCODE_SETLT */
1078       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1079          src0 = emit_fetch( bld, inst, 0, chan_index );
1080          src1 = emit_fetch( bld, inst, 1, chan_index );
1081          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
1082          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1083       }
1084       break;
1085
1086    case TGSI_OPCODE_SGE:
1087    /* TGSI_OPCODE_SETGE */
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          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
1092          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1093       }
1094       break;
1095
1096    case TGSI_OPCODE_MAD:
1097    /* TGSI_OPCODE_MADD */
1098       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1099          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1100          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1101          tmp2 = emit_fetch( bld, inst, 2, chan_index );
1102          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1103          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
1104          dst0[chan_index] = tmp0;
1105       }
1106       break;
1107
1108    case TGSI_OPCODE_SUB:
1109       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1110          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1111          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1112          dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
1113       }
1114       break;
1115
1116    case TGSI_OPCODE_LRP:
1117       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1118          src0 = emit_fetch( bld, inst, 0, chan_index );
1119          src1 = emit_fetch( bld, inst, 1, chan_index );
1120          src2 = emit_fetch( bld, inst, 2, chan_index );
1121          tmp0 = lp_build_sub( &bld->base, src1, src2 );
1122          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
1123          dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
1124       }
1125       break;
1126
1127    case TGSI_OPCODE_CND:
1128       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1129          src0 = emit_fetch( bld, inst, 0, chan_index );
1130          src1 = emit_fetch( bld, inst, 1, chan_index );
1131          src2 = emit_fetch( bld, inst, 2, chan_index );
1132          tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1133          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1134          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1135       }
1136       break;
1137
1138    case TGSI_OPCODE_DP2A:
1139       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1140       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1141       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1142       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1143       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1144       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1145       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1146       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
1147       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1148       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1149          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1150       }
1151       break;
1152
1153    case TGSI_OPCODE_FRC:
1154       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1155          src0 = emit_fetch( bld, inst, 0, chan_index );
1156          tmp0 = lp_build_floor(&bld->base, src0);
1157          tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1158          dst0[chan_index] = tmp0;
1159       }
1160       break;
1161
1162    case TGSI_OPCODE_CLAMP:
1163       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1164          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1165          src1 = emit_fetch( bld, inst, 1, chan_index );
1166          src2 = emit_fetch( bld, inst, 2, chan_index );
1167          tmp0 = lp_build_max(&bld->base, tmp0, src1);
1168          tmp0 = lp_build_min(&bld->base, tmp0, src2);
1169          dst0[chan_index] = tmp0;
1170       }
1171       break;
1172
1173    case TGSI_OPCODE_FLR:
1174       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1175          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1176          dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1177       }
1178       break;
1179
1180    case TGSI_OPCODE_ROUND:
1181       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1182          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1183          dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1184       }
1185       break;
1186
1187    case TGSI_OPCODE_EX2: {
1188       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1189       tmp0 = lp_build_exp2( &bld->base, tmp0);
1190       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1191          dst0[chan_index] = tmp0;
1192       }
1193       break;
1194    }
1195
1196    case TGSI_OPCODE_LG2:
1197       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1198       tmp0 = lp_build_log2( &bld->base, tmp0);
1199       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1200          dst0[chan_index] = tmp0;
1201       }
1202       break;
1203
1204    case TGSI_OPCODE_POW:
1205       src0 = emit_fetch( bld, inst, 0, CHAN_X );
1206       src1 = emit_fetch( bld, inst, 1, CHAN_X );
1207       res = lp_build_pow( &bld->base, src0, src1 );
1208       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1209          dst0[chan_index] = res;
1210       }
1211       break;
1212
1213    case TGSI_OPCODE_XPD:
1214       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1215           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1216          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1217          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1218       }
1219       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1220           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1221          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1222          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1223       }
1224       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1225          tmp2 = tmp0;
1226          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1227          tmp5 = tmp3;
1228          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1229          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1230          dst0[CHAN_X] = tmp2;
1231       }
1232       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1233           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1234          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1235          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1236       }
1237       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1238          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1239          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1240          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1241          dst0[CHAN_Y] = tmp3;
1242       }
1243       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1244          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1245          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1246          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1247          dst0[CHAN_Z] = tmp5;
1248       }
1249       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1250          dst0[CHAN_W] = bld->base.one;
1251       }
1252       break;
1253
1254    case TGSI_OPCODE_ABS:
1255       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1256          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1257          dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1258       }
1259       break;
1260
1261    case TGSI_OPCODE_RCC:
1262       /* deprecated? */
1263       assert(0);
1264       return FALSE;
1265
1266    case TGSI_OPCODE_DPH:
1267       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1268       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1269       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1270       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1271       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1272       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1273       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1274       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1275       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1276       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1277       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1278       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1279       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1280       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1281          dst0[chan_index] = tmp0;
1282       }
1283       break;
1284
1285    case TGSI_OPCODE_COS:
1286       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1287       tmp0 = lp_build_cos( &bld->base, tmp0 );
1288       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1289          dst0[chan_index] = tmp0;
1290       }
1291       break;
1292
1293    case TGSI_OPCODE_DDX:
1294       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1295          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1296       }
1297       break;
1298
1299    case TGSI_OPCODE_DDY:
1300       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1301          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1302       }
1303       break;
1304
1305    case TGSI_OPCODE_KILP:
1306       /* predicated kill */
1307       emit_kilp( bld, inst );
1308       break;
1309
1310    case TGSI_OPCODE_KIL:
1311       /* conditional kill */
1312       emit_kil( bld, inst );
1313       break;
1314
1315    case TGSI_OPCODE_PK2H:
1316       return FALSE;
1317       break;
1318
1319    case TGSI_OPCODE_PK2US:
1320       return FALSE;
1321       break;
1322
1323    case TGSI_OPCODE_PK4B:
1324       return FALSE;
1325       break;
1326
1327    case TGSI_OPCODE_PK4UB:
1328       return FALSE;
1329       break;
1330
1331    case TGSI_OPCODE_RFL:
1332       return FALSE;
1333       break;
1334
1335    case TGSI_OPCODE_SEQ:
1336       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1337          src0 = emit_fetch( bld, inst, 0, chan_index );
1338          src1 = emit_fetch( bld, inst, 1, chan_index );
1339          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1340          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1341       }
1342       break;
1343
1344    case TGSI_OPCODE_SFL:
1345       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1346          dst0[chan_index] = bld->base.zero;
1347       }
1348       break;
1349
1350    case TGSI_OPCODE_SGT:
1351       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1352          src0 = emit_fetch( bld, inst, 0, chan_index );
1353          src1 = emit_fetch( bld, inst, 1, chan_index );
1354          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1355          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1356       }
1357       break;
1358
1359    case TGSI_OPCODE_SIN:
1360       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1361       tmp0 = lp_build_sin( &bld->base, tmp0 );
1362       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1363          dst0[chan_index] = tmp0;
1364       }
1365       break;
1366
1367    case TGSI_OPCODE_SLE:
1368       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1369          src0 = emit_fetch( bld, inst, 0, chan_index );
1370          src1 = emit_fetch( bld, inst, 1, chan_index );
1371          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
1372          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1373       }
1374       break;
1375
1376    case TGSI_OPCODE_SNE:
1377       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1378          src0 = emit_fetch( bld, inst, 0, chan_index );
1379          src1 = emit_fetch( bld, inst, 1, chan_index );
1380          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1381          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1382       }
1383       break;
1384
1385    case TGSI_OPCODE_STR:
1386       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1387          dst0[chan_index] = bld->base.one;
1388       }
1389       break;
1390
1391    case TGSI_OPCODE_TEX:
1392       emit_tex( bld, inst, TEX_MODIFIER_NONE, dst0 );
1393       break;
1394
1395    case TGSI_OPCODE_TXD:
1396       emit_tex( bld, inst, TEX_MODIFIER_EXPLICIT_DERIV, dst0 );
1397       break;
1398
1399    case TGSI_OPCODE_UP2H:
1400       /* deprecated */
1401       assert (0);
1402       return FALSE;
1403       break;
1404
1405    case TGSI_OPCODE_UP2US:
1406       /* deprecated */
1407       assert(0);
1408       return FALSE;
1409       break;
1410
1411    case TGSI_OPCODE_UP4B:
1412       /* deprecated */
1413       assert(0);
1414       return FALSE;
1415       break;
1416
1417    case TGSI_OPCODE_UP4UB:
1418       /* deprecated */
1419       assert(0);
1420       return FALSE;
1421       break;
1422
1423    case TGSI_OPCODE_X2D:
1424       /* deprecated? */
1425       assert(0);
1426       return FALSE;
1427       break;
1428
1429    case TGSI_OPCODE_ARA:
1430       /* deprecated */
1431       assert(0);
1432       return FALSE;
1433       break;
1434
1435    case TGSI_OPCODE_ARR:
1436       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1437          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1438          tmp0 = lp_build_round(&bld->base, tmp0);
1439          dst0[chan_index] = tmp0;
1440       }
1441       break;
1442
1443    case TGSI_OPCODE_BRA:
1444       /* deprecated */
1445       assert(0);
1446       return FALSE;
1447       break;
1448
1449    case TGSI_OPCODE_CAL:
1450       /* FIXME */
1451       return FALSE;
1452       break;
1453
1454    case TGSI_OPCODE_RET:
1455       /* FIXME */
1456       return FALSE;
1457       break;
1458
1459    case TGSI_OPCODE_END:
1460       break;
1461
1462    case TGSI_OPCODE_SSG:
1463    /* TGSI_OPCODE_SGN */
1464       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1465          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1466          dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1467       }
1468       break;
1469
1470    case TGSI_OPCODE_CMP:
1471       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1472          src0 = emit_fetch( bld, inst, 0, chan_index );
1473          src1 = emit_fetch( bld, inst, 1, chan_index );
1474          src2 = emit_fetch( bld, inst, 2, chan_index );
1475          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1476          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1477       }
1478       break;
1479
1480    case TGSI_OPCODE_SCS:
1481       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1482          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1483          dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1484       }
1485       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1486          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1487          dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1488       }
1489       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1490          dst0[CHAN_Z] = bld->base.zero;
1491       }
1492       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1493          dst0[CHAN_W] = bld->base.one;
1494       }
1495       break;
1496
1497    case TGSI_OPCODE_TXB:
1498       emit_tex( bld, inst, TEX_MODIFIER_LOD_BIAS, dst0 );
1499       break;
1500
1501    case TGSI_OPCODE_NRM:
1502       /* fall-through */
1503    case TGSI_OPCODE_NRM4:
1504       /* 3 or 4-component normalization */
1505       {
1506          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1507
1508          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1509              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1510              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1511              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1512
1513             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1514
1515             /* xmm4 = src.x */
1516             /* xmm0 = src.x * src.x */
1517             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1518             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1519                tmp4 = tmp0;
1520             }
1521             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1522
1523             /* xmm5 = src.y */
1524             /* xmm0 = xmm0 + src.y * src.y */
1525             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1526             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1527                tmp5 = tmp1;
1528             }
1529             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1530             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1531
1532             /* xmm6 = src.z */
1533             /* xmm0 = xmm0 + src.z * src.z */
1534             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1535             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1536                tmp6 = tmp1;
1537             }
1538             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1539             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1540
1541             if (dims == 4) {
1542                /* xmm7 = src.w */
1543                /* xmm0 = xmm0 + src.w * src.w */
1544                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1545                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1546                   tmp7 = tmp1;
1547                }
1548                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1549                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1550             }
1551
1552             /* xmm1 = 1 / sqrt(xmm0) */
1553             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1554
1555             /* dst.x = xmm1 * src.x */
1556             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1557                dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1558             }
1559
1560             /* dst.y = xmm1 * src.y */
1561             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1562                dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1563             }
1564
1565             /* dst.z = xmm1 * src.z */
1566             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1567                dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1568             }
1569
1570             /* dst.w = xmm1 * src.w */
1571             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1572                dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1573             }
1574          }
1575
1576          /* dst.w = 1.0 */
1577          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1578             dst0[CHAN_W] = bld->base.one;
1579          }
1580       }
1581       break;
1582
1583    case TGSI_OPCODE_DIV:
1584       /* deprecated */
1585       assert( 0 );
1586       return FALSE;
1587       break;
1588
1589    case TGSI_OPCODE_DP2:
1590       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1591       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1592       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1593       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1594       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1595       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1596       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1597       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1598          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1599       }
1600       break;
1601
1602    case TGSI_OPCODE_TXL:
1603       emit_tex( bld, inst, TEX_MODIFIER_EXPLICIT_LOD, dst0 );
1604       break;
1605
1606    case TGSI_OPCODE_TXP:
1607       emit_tex( bld, inst, TEX_MODIFIER_PROJECTED, dst0 );
1608       break;
1609
1610    case TGSI_OPCODE_BRK:
1611       lp_exec_break(&bld->exec_mask);
1612       break;
1613
1614    case TGSI_OPCODE_IF:
1615       tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1616       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1617                           tmp0, bld->base.zero);
1618       lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1619       break;
1620
1621    case TGSI_OPCODE_BGNLOOP:
1622       lp_exec_bgnloop(&bld->exec_mask);
1623       break;
1624
1625    case TGSI_OPCODE_ELSE:
1626       lp_exec_mask_cond_invert(&bld->exec_mask);
1627       break;
1628
1629    case TGSI_OPCODE_ENDIF:
1630       lp_exec_mask_cond_pop(&bld->exec_mask);
1631       break;
1632
1633    case TGSI_OPCODE_ENDLOOP:
1634       lp_exec_endloop(&bld->exec_mask);
1635       break;
1636
1637    case TGSI_OPCODE_PUSHA:
1638       /* deprecated? */
1639       assert(0);
1640       return FALSE;
1641       break;
1642
1643    case TGSI_OPCODE_POPA:
1644       /* deprecated? */
1645       assert(0);
1646       return FALSE;
1647       break;
1648
1649    case TGSI_OPCODE_CEIL:
1650       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1651          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1652          dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1653       }
1654       break;
1655
1656    case TGSI_OPCODE_I2F:
1657       /* deprecated? */
1658       assert(0);
1659       return FALSE;
1660       break;
1661
1662    case TGSI_OPCODE_NOT:
1663       /* deprecated? */
1664       assert(0);
1665       return FALSE;
1666       break;
1667
1668    case TGSI_OPCODE_TRUNC:
1669       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1670          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1671          dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1672       }
1673       break;
1674
1675    case TGSI_OPCODE_SHL:
1676       /* deprecated? */
1677       assert(0);
1678       return FALSE;
1679       break;
1680
1681    case TGSI_OPCODE_ISHR:
1682       /* deprecated? */
1683       assert(0);
1684       return FALSE;
1685       break;
1686
1687    case TGSI_OPCODE_AND:
1688       /* deprecated? */
1689       assert(0);
1690       return FALSE;
1691       break;
1692
1693    case TGSI_OPCODE_OR:
1694       /* deprecated? */
1695       assert(0);
1696       return FALSE;
1697       break;
1698
1699    case TGSI_OPCODE_MOD:
1700       /* deprecated? */
1701       assert(0);
1702       return FALSE;
1703       break;
1704
1705    case TGSI_OPCODE_XOR:
1706       /* deprecated? */
1707       assert(0);
1708       return FALSE;
1709       break;
1710
1711    case TGSI_OPCODE_SAD:
1712       /* deprecated? */
1713       assert(0);
1714       return FALSE;
1715       break;
1716
1717    case TGSI_OPCODE_TXF:
1718       /* deprecated? */
1719       assert(0);
1720       return FALSE;
1721       break;
1722
1723    case TGSI_OPCODE_TXQ:
1724       /* deprecated? */
1725       assert(0);
1726       return FALSE;
1727       break;
1728
1729    case TGSI_OPCODE_CONT:
1730       lp_exec_continue(&bld->exec_mask);
1731       break;
1732
1733    case TGSI_OPCODE_EMIT:
1734       return FALSE;
1735       break;
1736
1737    case TGSI_OPCODE_ENDPRIM:
1738       return FALSE;
1739       break;
1740
1741    case TGSI_OPCODE_NOP:
1742       break;
1743
1744    default:
1745       return FALSE;
1746    }
1747
1748    if(info->num_dst) {
1749       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1750          emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
1751       }
1752    }
1753
1754    return TRUE;
1755 }
1756
1757
1758 void
1759 lp_build_tgsi_soa(LLVMBuilderRef builder,
1760                   const struct tgsi_token *tokens,
1761                   struct lp_type type,
1762                   struct lp_build_mask_context *mask,
1763                   LLVMValueRef consts_ptr,
1764                   const LLVMValueRef *pos,
1765                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1766                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1767                   struct lp_build_sampler_soa *sampler,
1768                   struct tgsi_shader_info *info)
1769 {
1770    struct lp_build_tgsi_soa_context bld;
1771    struct tgsi_parse_context parse;
1772    uint num_immediates = 0;
1773    unsigned i;
1774
1775    /* Setup build context */
1776    memset(&bld, 0, sizeof bld);
1777    lp_build_context_init(&bld.base, builder, type);
1778    bld.mask = mask;
1779    bld.pos = pos;
1780    bld.inputs = inputs;
1781    bld.outputs = outputs;
1782    bld.consts_ptr = consts_ptr;
1783    bld.sampler = sampler;
1784    bld.has_indirect_addressing = info->opcode_count[TGSI_OPCODE_ARR] > 0 ||
1785                                  info->opcode_count[TGSI_OPCODE_ARL] > 0;
1786
1787    lp_exec_mask_init(&bld.exec_mask, &bld.base);
1788
1789    tgsi_parse_init( &parse, tokens );
1790
1791    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1792       tgsi_parse_token( &parse );
1793
1794       switch( parse.FullToken.Token.Type ) {
1795       case TGSI_TOKEN_TYPE_DECLARATION:
1796          /* Inputs already interpolated */
1797          emit_declaration( &bld, &parse.FullToken.FullDeclaration );
1798          break;
1799
1800       case TGSI_TOKEN_TYPE_INSTRUCTION:
1801          {
1802             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1803             const struct tgsi_opcode_info *opcode_info = tgsi_get_opcode_info(opcode);
1804             if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, opcode_info ))
1805                _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1806                              opcode_info->mnemonic);
1807          }
1808
1809          break;
1810
1811       case TGSI_TOKEN_TYPE_IMMEDIATE:
1812          /* simply copy the immediate values into the next immediates[] slot */
1813          {
1814             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1815             assert(size <= 4);
1816             assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
1817             for( i = 0; i < size; ++i )
1818                bld.immediates[num_immediates][i] =
1819                   lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
1820             for( i = size; i < 4; ++i )
1821                bld.immediates[num_immediates][i] = bld.base.undef;
1822             num_immediates++;
1823          }
1824          break;
1825
1826       case TGSI_TOKEN_TYPE_PROPERTY:
1827          break;
1828
1829       default:
1830          assert( 0 );
1831       }
1832    }
1833    if (0) {
1834       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1835       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1836       debug_printf("11111111111111111111111111111 \n");
1837       tgsi_dump(tokens, 0);
1838       LLVMDumpValue(function);
1839       debug_printf("2222222222222222222222222222 \n");
1840    }
1841    tgsi_parse_free( &parse );
1842 }
1843