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 lod_bias, explicit_lod;
 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    if (modifier == TEX_MODIFIER_LOD_BIAS) {
 647       lod_bias = emit_fetch( bld, inst, 0, 3 );
 648       explicit_lod = NULL;
 649    }
 650    else if (modifier == TEX_MODIFIER_EXPLICIT_LOD) {
 651       lod_bias = NULL;
 652       explicit_lod = emit_fetch( bld, inst, 0, 3 );
 653    }
 654    else {
 655       lod_bias = NULL;
 656       explicit_lod = NULL;
 657    }
 658
 659    if (modifier == TEX_MODIFIER_PROJECTED) {
 660       oow = emit_fetch( bld, inst, 0, 3 );
 661       oow = lp_build_rcp(&bld->base, oow);
 662    }
 663
 664    for (i = 0; i < num_coords; i++) {
 665       coords[i] = emit_fetch( bld, inst, 0, i );
 666       if (modifier == TEX_MODIFIER_PROJECTED)
 667          coords[i] = lp_build_mul(&bld->base, coords[i], oow);
 668    }
 669    for (i = num_coords; i < 3; i++) {
 670       coords[i] = bld->base.undef;
 671    }
 672
 673    if (modifier == TEX_MODIFIER_EXPLICIT_DERIV) {
 674       for (i = 0; i < num_coords; i++) {
 675          ddx[i] = emit_fetch( bld, inst, 1, i );
 676          ddy[i] = emit_fetch( bld, inst, 2, i );
 677       }
 678       unit = inst->Src[3].Register.Index;
 679    }  else {
 680       for (i = 0; i < num_coords; i++) {
 681          ddx[i] = emit_ddx( bld, coords[i] );
 682          ddy[i] = emit_ddy( bld, coords[i] );
 683       }
 684       unit = inst->Src[1].Register.Index;
 685    }
 686    for (i = num_coords; i < 3; i++) {
 687       ddx[i] = bld->base.undef;
 688       ddy[i] = bld->base.undef;
 689    }
 690
 691    bld->sampler->emit_fetch_texel(bld->sampler,
 692                                   bld->base.builder,
 693                                   bld->base.type,
 694                                   unit, num_coords, coords,
 695                                   ddx, ddy,
 696                                   lod_bias, explicit_lod,
 697                                   texel);
 698 }
 699
 700
 701 /**
 702  * Kill fragment if any of the src register values are negative.
 703  */
 704 static void
 705 emit_kil(
 706    struct lp_build_tgsi_soa_context *bld,
 707    const struct tgsi_full_instruction *inst )
 708 {
 709    const struct tgsi_full_src_register *reg = &inst->Src[0];
 710    LLVMValueRef terms[NUM_CHANNELS];
 711    LLVMValueRef mask;
 712    unsigned chan_index;
 713
 714    memset(&terms, 0, sizeof terms);
 715
 716    FOR_EACH_CHANNEL( chan_index ) {
 717       unsigned swizzle;
 718
 719       /* Unswizzle channel */
 720       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
 721
 722       /* Check if the component has not been already tested. */
 723       assert(swizzle < NUM_CHANNELS);
 724       if( !terms[swizzle] )
 725          /* TODO: change the comparison operator instead of setting the sign */
 726          terms[swizzle] =  emit_fetch(bld, inst, 0, chan_index );
 727    }
 728
 729    mask = NULL;
 730    FOR_EACH_CHANNEL( chan_index ) {
 731       if(terms[chan_index]) {
 732          LLVMValueRef chan_mask;
 733
 734          /*
 735           * If term < 0 then mask = 0 else mask = ~0.
 736           */
 737          chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero);
 738
 739          if(mask)
 740             mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, "");
 741          else
 742             mask = chan_mask;
 743       }
 744    }
 745
 746    if(mask)
 747       lp_build_mask_update(bld->mask, mask);
 748 }
 749
 750
 751 /**
 752  * Predicated fragment kill.
 753  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
 754  * The only predication is the execution mask which will apply if
 755  * we're inside a loop or conditional.
 756  */
 757 static void
 758 emit_kilp(struct lp_build_tgsi_soa_context *bld,
 759           const struct tgsi_full_instruction *inst)
 760 {
 761    LLVMValueRef mask;
 762
 763    /* For those channels which are "alive", disable fragment shader
 764     * execution.
 765     */
 766    if (bld->exec_mask.has_mask) {
 767       mask = LLVMBuildNot(bld->base.builder, bld->exec_mask.exec_mask, "kilp");
 768    }
 769    else {
 770       mask = bld->base.zero;
 771    }
 772
 773    lp_build_mask_update(bld->mask, mask);
 774 }
 775
 776 static void
 777 emit_declaration(
 778    struct lp_build_tgsi_soa_context *bld,
 779    const struct tgsi_full_declaration *decl)
 780 {
 781    LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
 782
 783    unsigned first = decl->Range.First;
 784    unsigned last = decl->Range.Last;
 785    unsigned idx, i;
 786
 787    for (idx = first; idx <= last; ++idx) {
 788       switch (decl->Declaration.File) {
 789       case TGSI_FILE_TEMPORARY:
 790          assert(idx < LP_MAX_TGSI_TEMPS);
 791          if (bld->has_indirect_addressing) {
 792             LLVMValueRef val = LLVMConstInt(LLVMInt32Type(),
 793                                             last*4 + 4, 0);
 794             bld->temps_array = lp_build_array_alloca(bld->base.builder,
 795                                                      vec_type, val, "");
 796          } else {
 797             for (i = 0; i < NUM_CHANNELS; i++)
 798                bld->temps[idx][i] = lp_build_alloca(bld->base.builder,
 799                                                     vec_type, "");
 800          }
 801          break;
 802
 803       case TGSI_FILE_OUTPUT:
 804          for (i = 0; i < NUM_CHANNELS; i++)
 805             bld->outputs[idx][i] = lp_build_alloca(bld->base.builder,
 806                                                    vec_type, "");
 807          break;
 808
 809       case TGSI_FILE_ADDRESS:
 810          assert(idx < LP_MAX_TGSI_ADDRS);
 811          for (i = 0; i < NUM_CHANNELS; i++)
 812             bld->addr[idx][i] = lp_build_alloca(bld->base.builder,
 813                                                 vec_type, "");
 814          break;
 815
 816       case TGSI_FILE_PREDICATE:
 817          _debug_printf("warning: predicate registers not yet implemented\n");
 818          break;
 819
 820       default:
 821          /* don't need to declare other vars */
 822          break;
 823       }
 824    }
 825 }
 826
 827
 828 /**
 829  * Emit LLVM for one TGSI instruction.
 830  * \param return TRUE for success, FALSE otherwise
 831  */
 832 static boolean
 833 emit_instruction(
 834    struct lp_build_tgsi_soa_context *bld,
 835    const struct tgsi_full_instruction *inst,
 836    const struct tgsi_opcode_info *info)
 837 {
 838    unsigned chan_index;
 839    LLVMValueRef src0, src1, src2;
 840    LLVMValueRef tmp0, tmp1, tmp2;
 841    LLVMValueRef tmp3 = NULL;
 842    LLVMValueRef tmp4 = NULL;
 843    LLVMValueRef tmp5 = NULL;
 844    LLVMValueRef tmp6 = NULL;
 845    LLVMValueRef tmp7 = NULL;
 846    LLVMValueRef res;
 847    LLVMValueRef dst0[NUM_CHANNELS];
 848
 849    /*
 850     * Stores and write masks are handled in a general fashion after the long
 851     * instruction opcode switch statement.
 852     *
 853     * Although not stricitly necessary, we avoid generating instructions for
 854     * channels which won't be stored, in cases where's that easy. For some
 855     * complex instructions, like texture sampling, it is more convenient to
 856     * assume a full writemask and then let LLVM optimization passes eliminate
 857     * redundant code.
 858     */
 859
 860    assert(info->num_dst <= 1);
 861    if(info->num_dst) {
 862       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 863          dst0[chan_index] = bld->base.undef;
 864       }
 865    }
 866
 867    switch (inst->Instruction.Opcode) {
 868    case TGSI_OPCODE_ARL:
 869       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 870          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 871          tmp0 = lp_build_floor(&bld->base, tmp0);
 872          dst0[chan_index] = tmp0;
 873       }
 874       break;
 875
 876    case TGSI_OPCODE_MOV:
 877       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 878          dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
 879       }
 880       break;
 881
 882    case TGSI_OPCODE_LIT:
 883       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 884          dst0[CHAN_X] = bld->base.one;
 885       }
 886       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 887          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 888          dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
 889       }
 890       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 891          /* XMM[1] = SrcReg[0].yyyy */
 892          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 893          /* XMM[1] = max(XMM[1], 0) */
 894          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 895          /* XMM[2] = SrcReg[0].wwww */
 896          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 897          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 898          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 899          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 900          dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 901       }
 902       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 903          dst0[CHAN_W] = bld->base.one;
 904       }
 905       break;
 906
 907    case TGSI_OPCODE_RCP:
 908    /* TGSI_OPCODE_RECIP */
 909       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 910       res = lp_build_rcp(&bld->base, src0);
 911       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 912          dst0[chan_index] = res;
 913       }
 914       break;
 915
 916    case TGSI_OPCODE_RSQ:
 917    /* TGSI_OPCODE_RECIPSQRT */
 918       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 919       src0 = lp_build_abs(&bld->base, src0);
 920       res = lp_build_rsqrt(&bld->base, src0);
 921       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 922          dst0[chan_index] = res;
 923       }
 924       break;
 925
 926    case TGSI_OPCODE_EXP:
 927       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 928           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 929           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 930          LLVMValueRef *p_exp2_int_part = NULL;
 931          LLVMValueRef *p_frac_part = NULL;
 932          LLVMValueRef *p_exp2 = NULL;
 933
 934          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 935
 936          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 937             p_exp2_int_part = &tmp0;
 938          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 939             p_frac_part = &tmp1;
 940          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 941             p_exp2 = &tmp2;
 942
 943          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 944
 945          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 946             dst0[CHAN_X] = tmp0;
 947          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 948             dst0[CHAN_Y] = tmp1;
 949          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 950             dst0[CHAN_Z] = tmp2;
 951       }
 952       /* dst.w = 1.0 */
 953       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 954          dst0[CHAN_W] = bld->base.one;
 955       }
 956       break;
 957
 958    case TGSI_OPCODE_LOG:
 959       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 960           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 961           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 962          LLVMValueRef *p_floor_log2 = NULL;
 963          LLVMValueRef *p_exp = NULL;
 964          LLVMValueRef *p_log2 = NULL;
 965
 966          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 967          src0 = lp_build_abs( &bld->base, src0 );
 968
 969          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 970             p_floor_log2 = &tmp0;
 971          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 972             p_exp = &tmp1;
 973          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 974             p_log2 = &tmp2;
 975
 976          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 977
 978          /* dst.x = floor(lg2(abs(src.x))) */
 979          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 980             dst0[CHAN_X] = tmp0;
 981          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 982          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 983             dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
 984          }
 985          /* dst.z = lg2(abs(src.x)) */
 986          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 987             dst0[CHAN_Z] = tmp2;
 988       }
 989       /* dst.w = 1.0 */
 990       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 991          dst0[CHAN_W] = bld->base.one;
 992       }
 993       break;
 994
 995    case TGSI_OPCODE_MUL:
 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_mul(&bld->base, src0, src1);
1000       }
1001       break;
1002
1003    case TGSI_OPCODE_ADD:
1004       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1005          src0 = emit_fetch( bld, inst, 0, chan_index );
1006          src1 = emit_fetch( bld, inst, 1, chan_index );
1007          dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
1008       }
1009       break;
1010
1011    case TGSI_OPCODE_DP3:
1012    /* TGSI_OPCODE_DOT3 */
1013       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1014       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1015       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1016       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1017       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1018       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1019       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1020       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1021       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1022       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1023       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1024       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1025          dst0[chan_index] = tmp0;
1026       }
1027       break;
1028
1029    case TGSI_OPCODE_DP4:
1030    /* TGSI_OPCODE_DOT4 */
1031       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1032       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1033       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1034       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1035       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1036       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1037       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1038       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1039       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1040       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1041       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1042       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
1043       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
1044       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1045       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1046       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1047          dst0[chan_index] = tmp0;
1048       }
1049       break;
1050
1051    case TGSI_OPCODE_DST:
1052       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1053          dst0[CHAN_X] = bld->base.one;
1054       }
1055       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1056          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1057          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
1058          dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
1059       }
1060       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1061          dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
1062       }
1063       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1064          dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
1065       }
1066       break;
1067
1068    case TGSI_OPCODE_MIN:
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_min( &bld->base, src0, src1 );
1073       }
1074       break;
1075
1076    case TGSI_OPCODE_MAX:
1077       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1078          src0 = emit_fetch( bld, inst, 0, chan_index );
1079          src1 = emit_fetch( bld, inst, 1, chan_index );
1080          dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
1081       }
1082       break;
1083
1084    case TGSI_OPCODE_SLT:
1085    /* TGSI_OPCODE_SETLT */
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          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
1090          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1091       }
1092       break;
1093
1094    case TGSI_OPCODE_SGE:
1095    /* TGSI_OPCODE_SETGE */
1096       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1097          src0 = emit_fetch( bld, inst, 0, chan_index );
1098          src1 = emit_fetch( bld, inst, 1, chan_index );
1099          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
1100          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1101       }
1102       break;
1103
1104    case TGSI_OPCODE_MAD:
1105    /* TGSI_OPCODE_MADD */
1106       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1107          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1108          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1109          tmp2 = emit_fetch( bld, inst, 2, chan_index );
1110          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1111          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
1112          dst0[chan_index] = tmp0;
1113       }
1114       break;
1115
1116    case TGSI_OPCODE_SUB:
1117       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1118          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1119          tmp1 = emit_fetch( bld, inst, 1, chan_index );
1120          dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
1121       }
1122       break;
1123
1124    case TGSI_OPCODE_LRP:
1125       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1126          src0 = emit_fetch( bld, inst, 0, chan_index );
1127          src1 = emit_fetch( bld, inst, 1, chan_index );
1128          src2 = emit_fetch( bld, inst, 2, chan_index );
1129          tmp0 = lp_build_sub( &bld->base, src1, src2 );
1130          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
1131          dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
1132       }
1133       break;
1134
1135    case TGSI_OPCODE_CND:
1136       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1137          src0 = emit_fetch( bld, inst, 0, chan_index );
1138          src1 = emit_fetch( bld, inst, 1, chan_index );
1139          src2 = emit_fetch( bld, inst, 2, chan_index );
1140          tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1141          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1142          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1143       }
1144       break;
1145
1146    case TGSI_OPCODE_DP2A:
1147       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1148       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1149       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1150       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1151       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1152       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1153       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1154       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
1155       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1156       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1157          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1158       }
1159       break;
1160
1161    case TGSI_OPCODE_FRC:
1162       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1163          src0 = emit_fetch( bld, inst, 0, chan_index );
1164          tmp0 = lp_build_floor(&bld->base, src0);
1165          tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1166          dst0[chan_index] = tmp0;
1167       }
1168       break;
1169
1170    case TGSI_OPCODE_CLAMP:
1171       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1172          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1173          src1 = emit_fetch( bld, inst, 1, chan_index );
1174          src2 = emit_fetch( bld, inst, 2, chan_index );
1175          tmp0 = lp_build_max(&bld->base, tmp0, src1);
1176          tmp0 = lp_build_min(&bld->base, tmp0, src2);
1177          dst0[chan_index] = tmp0;
1178       }
1179       break;
1180
1181    case TGSI_OPCODE_FLR:
1182       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1183          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1184          dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1185       }
1186       break;
1187
1188    case TGSI_OPCODE_ROUND:
1189       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1190          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1191          dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1192       }
1193       break;
1194
1195    case TGSI_OPCODE_EX2: {
1196       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1197       tmp0 = lp_build_exp2( &bld->base, tmp0);
1198       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1199          dst0[chan_index] = tmp0;
1200       }
1201       break;
1202    }
1203
1204    case TGSI_OPCODE_LG2:
1205       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1206       tmp0 = lp_build_log2( &bld->base, tmp0);
1207       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1208          dst0[chan_index] = tmp0;
1209       }
1210       break;
1211
1212    case TGSI_OPCODE_POW:
1213       src0 = emit_fetch( bld, inst, 0, CHAN_X );
1214       src1 = emit_fetch( bld, inst, 1, CHAN_X );
1215       res = lp_build_pow( &bld->base, src0, src1 );
1216       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1217          dst0[chan_index] = res;
1218       }
1219       break;
1220
1221    case TGSI_OPCODE_XPD:
1222       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1223           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1224          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1225          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1226       }
1227       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1228           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1229          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1230          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1231       }
1232       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1233          tmp2 = tmp0;
1234          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1235          tmp5 = tmp3;
1236          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1237          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1238          dst0[CHAN_X] = tmp2;
1239       }
1240       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1241           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1242          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1243          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1244       }
1245       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1246          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1247          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1248          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1249          dst0[CHAN_Y] = tmp3;
1250       }
1251       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1252          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1253          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1254          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1255          dst0[CHAN_Z] = tmp5;
1256       }
1257       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1258          dst0[CHAN_W] = bld->base.one;
1259       }
1260       break;
1261
1262    case TGSI_OPCODE_ABS:
1263       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1264          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1265          dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1266       }
1267       break;
1268
1269    case TGSI_OPCODE_RCC:
1270       /* deprecated? */
1271       assert(0);
1272       return FALSE;
1273
1274    case TGSI_OPCODE_DPH:
1275       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1276       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1277       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1278       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1279       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1280       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1281       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1282       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1283       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1284       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1285       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1286       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1287       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1288       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1289          dst0[chan_index] = tmp0;
1290       }
1291       break;
1292
1293    case TGSI_OPCODE_COS:
1294       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1295       tmp0 = lp_build_cos( &bld->base, tmp0 );
1296       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1297          dst0[chan_index] = tmp0;
1298       }
1299       break;
1300
1301    case TGSI_OPCODE_DDX:
1302       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1303          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1304       }
1305       break;
1306
1307    case TGSI_OPCODE_DDY:
1308       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1309          emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1310       }
1311       break;
1312
1313    case TGSI_OPCODE_KILP:
1314       /* predicated kill */
1315       emit_kilp( bld, inst );
1316       break;
1317
1318    case TGSI_OPCODE_KIL:
1319       /* conditional kill */
1320       emit_kil( bld, inst );
1321       break;
1322
1323    case TGSI_OPCODE_PK2H:
1324       return FALSE;
1325       break;
1326
1327    case TGSI_OPCODE_PK2US:
1328       return FALSE;
1329       break;
1330
1331    case TGSI_OPCODE_PK4B:
1332       return FALSE;
1333       break;
1334
1335    case TGSI_OPCODE_PK4UB:
1336       return FALSE;
1337       break;
1338
1339    case TGSI_OPCODE_RFL:
1340       return FALSE;
1341       break;
1342
1343    case TGSI_OPCODE_SEQ:
1344       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1345          src0 = emit_fetch( bld, inst, 0, chan_index );
1346          src1 = emit_fetch( bld, inst, 1, chan_index );
1347          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1348          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1349       }
1350       break;
1351
1352    case TGSI_OPCODE_SFL:
1353       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1354          dst0[chan_index] = bld->base.zero;
1355       }
1356       break;
1357
1358    case TGSI_OPCODE_SGT:
1359       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1360          src0 = emit_fetch( bld, inst, 0, chan_index );
1361          src1 = emit_fetch( bld, inst, 1, chan_index );
1362          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1363          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1364       }
1365       break;
1366
1367    case TGSI_OPCODE_SIN:
1368       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1369       tmp0 = lp_build_sin( &bld->base, tmp0 );
1370       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1371          dst0[chan_index] = tmp0;
1372       }
1373       break;
1374
1375    case TGSI_OPCODE_SLE:
1376       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1377          src0 = emit_fetch( bld, inst, 0, chan_index );
1378          src1 = emit_fetch( bld, inst, 1, chan_index );
1379          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
1380          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1381       }
1382       break;
1383
1384    case TGSI_OPCODE_SNE:
1385       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1386          src0 = emit_fetch( bld, inst, 0, chan_index );
1387          src1 = emit_fetch( bld, inst, 1, chan_index );
1388          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1389          dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1390       }
1391       break;
1392
1393    case TGSI_OPCODE_STR:
1394       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1395          dst0[chan_index] = bld->base.one;
1396       }
1397       break;
1398
1399    case TGSI_OPCODE_TEX:
1400       emit_tex( bld, inst, TEX_MODIFIER_NONE, dst0 );
1401       break;
1402
1403    case TGSI_OPCODE_TXD:
1404       emit_tex( bld, inst, TEX_MODIFIER_EXPLICIT_DERIV, dst0 );
1405       break;
1406
1407    case TGSI_OPCODE_UP2H:
1408       /* deprecated */
1409       assert (0);
1410       return FALSE;
1411       break;
1412
1413    case TGSI_OPCODE_UP2US:
1414       /* deprecated */
1415       assert(0);
1416       return FALSE;
1417       break;
1418
1419    case TGSI_OPCODE_UP4B:
1420       /* deprecated */
1421       assert(0);
1422       return FALSE;
1423       break;
1424
1425    case TGSI_OPCODE_UP4UB:
1426       /* deprecated */
1427       assert(0);
1428       return FALSE;
1429       break;
1430
1431    case TGSI_OPCODE_X2D:
1432       /* deprecated? */
1433       assert(0);
1434       return FALSE;
1435       break;
1436
1437    case TGSI_OPCODE_ARA:
1438       /* deprecated */
1439       assert(0);
1440       return FALSE;
1441       break;
1442
1443    case TGSI_OPCODE_ARR:
1444       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1445          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1446          tmp0 = lp_build_round(&bld->base, tmp0);
1447          dst0[chan_index] = tmp0;
1448       }
1449       break;
1450
1451    case TGSI_OPCODE_BRA:
1452       /* deprecated */
1453       assert(0);
1454       return FALSE;
1455       break;
1456
1457    case TGSI_OPCODE_CAL:
1458       /* FIXME */
1459       return FALSE;
1460       break;
1461
1462    case TGSI_OPCODE_RET:
1463       /* FIXME */
1464       return FALSE;
1465       break;
1466
1467    case TGSI_OPCODE_END:
1468       break;
1469
1470    case TGSI_OPCODE_SSG:
1471    /* TGSI_OPCODE_SGN */
1472       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1473          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1474          dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1475       }
1476       break;
1477
1478    case TGSI_OPCODE_CMP:
1479       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1480          src0 = emit_fetch( bld, inst, 0, chan_index );
1481          src1 = emit_fetch( bld, inst, 1, chan_index );
1482          src2 = emit_fetch( bld, inst, 2, chan_index );
1483          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1484          dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1485       }
1486       break;
1487
1488    case TGSI_OPCODE_SCS:
1489       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1490          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1491          dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1492       }
1493       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1494          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1495          dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1496       }
1497       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1498          dst0[CHAN_Z] = bld->base.zero;
1499       }
1500       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1501          dst0[CHAN_W] = bld->base.one;
1502       }
1503       break;
1504
1505    case TGSI_OPCODE_TXB:
1506       emit_tex( bld, inst, TEX_MODIFIER_LOD_BIAS, dst0 );
1507       break;
1508
1509    case TGSI_OPCODE_NRM:
1510       /* fall-through */
1511    case TGSI_OPCODE_NRM4:
1512       /* 3 or 4-component normalization */
1513       {
1514          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1515
1516          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1517              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1518              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1519              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1520
1521             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1522
1523             /* xmm4 = src.x */
1524             /* xmm0 = src.x * src.x */
1525             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1526             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1527                tmp4 = tmp0;
1528             }
1529             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1530
1531             /* xmm5 = src.y */
1532             /* xmm0 = xmm0 + src.y * src.y */
1533             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1534             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1535                tmp5 = tmp1;
1536             }
1537             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1538             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1539
1540             /* xmm6 = src.z */
1541             /* xmm0 = xmm0 + src.z * src.z */
1542             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1543             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1544                tmp6 = tmp1;
1545             }
1546             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1547             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1548
1549             if (dims == 4) {
1550                /* xmm7 = src.w */
1551                /* xmm0 = xmm0 + src.w * src.w */
1552                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1553                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1554                   tmp7 = tmp1;
1555                }
1556                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1557                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1558             }
1559
1560             /* xmm1 = 1 / sqrt(xmm0) */
1561             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1562
1563             /* dst.x = xmm1 * src.x */
1564             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1565                dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1566             }
1567
1568             /* dst.y = xmm1 * src.y */
1569             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1570                dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1571             }
1572
1573             /* dst.z = xmm1 * src.z */
1574             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1575                dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1576             }
1577
1578             /* dst.w = xmm1 * src.w */
1579             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1580                dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1581             }
1582          }
1583
1584          /* dst.w = 1.0 */
1585          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1586             dst0[CHAN_W] = bld->base.one;
1587          }
1588       }
1589       break;
1590
1591    case TGSI_OPCODE_DIV:
1592       /* deprecated */
1593       assert( 0 );
1594       return FALSE;
1595       break;
1596
1597    case TGSI_OPCODE_DP2:
1598       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1599       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1600       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1601       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1602       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1603       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1604       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1605       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1606          dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
1607       }
1608       break;
1609
1610    case TGSI_OPCODE_TXL:
1611       emit_tex( bld, inst, TEX_MODIFIER_EXPLICIT_LOD, dst0 );
1612       break;
1613
1614    case TGSI_OPCODE_TXP:
1615       emit_tex( bld, inst, TEX_MODIFIER_PROJECTED, dst0 );
1616       break;
1617
1618    case TGSI_OPCODE_BRK:
1619       lp_exec_break(&bld->exec_mask);
1620       break;
1621
1622    case TGSI_OPCODE_IF:
1623       tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1624       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1625                           tmp0, bld->base.zero);
1626       lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1627       break;
1628
1629    case TGSI_OPCODE_BGNLOOP:
1630       lp_exec_bgnloop(&bld->exec_mask);
1631       break;
1632
1633    case TGSI_OPCODE_ELSE:
1634       lp_exec_mask_cond_invert(&bld->exec_mask);
1635       break;
1636
1637    case TGSI_OPCODE_ENDIF:
1638       lp_exec_mask_cond_pop(&bld->exec_mask);
1639       break;
1640
1641    case TGSI_OPCODE_ENDLOOP:
1642       lp_exec_endloop(&bld->exec_mask);
1643       break;
1644
1645    case TGSI_OPCODE_PUSHA:
1646       /* deprecated? */
1647       assert(0);
1648       return FALSE;
1649       break;
1650
1651    case TGSI_OPCODE_POPA:
1652       /* deprecated? */
1653       assert(0);
1654       return FALSE;
1655       break;
1656
1657    case TGSI_OPCODE_CEIL:
1658       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1659          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1660          dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1661       }
1662       break;
1663
1664    case TGSI_OPCODE_I2F:
1665       /* deprecated? */
1666       assert(0);
1667       return FALSE;
1668       break;
1669
1670    case TGSI_OPCODE_NOT:
1671       /* deprecated? */
1672       assert(0);
1673       return FALSE;
1674       break;
1675
1676    case TGSI_OPCODE_TRUNC:
1677       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1678          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1679          dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1680       }
1681       break;
1682
1683    case TGSI_OPCODE_SHL:
1684       /* deprecated? */
1685       assert(0);
1686       return FALSE;
1687       break;
1688
1689    case TGSI_OPCODE_ISHR:
1690       /* deprecated? */
1691       assert(0);
1692       return FALSE;
1693       break;
1694
1695    case TGSI_OPCODE_AND:
1696       /* deprecated? */
1697       assert(0);
1698       return FALSE;
1699       break;
1700
1701    case TGSI_OPCODE_OR:
1702       /* deprecated? */
1703       assert(0);
1704       return FALSE;
1705       break;
1706
1707    case TGSI_OPCODE_MOD:
1708       /* deprecated? */
1709       assert(0);
1710       return FALSE;
1711       break;
1712
1713    case TGSI_OPCODE_XOR:
1714       /* deprecated? */
1715       assert(0);
1716       return FALSE;
1717       break;
1718
1719    case TGSI_OPCODE_SAD:
1720       /* deprecated? */
1721       assert(0);
1722       return FALSE;
1723       break;
1724
1725    case TGSI_OPCODE_TXF:
1726       /* deprecated? */
1727       assert(0);
1728       return FALSE;
1729       break;
1730
1731    case TGSI_OPCODE_TXQ:
1732       /* deprecated? */
1733       assert(0);
1734       return FALSE;
1735       break;
1736
1737    case TGSI_OPCODE_CONT:
1738       lp_exec_continue(&bld->exec_mask);
1739       break;
1740
1741    case TGSI_OPCODE_EMIT:
1742       return FALSE;
1743       break;
1744
1745    case TGSI_OPCODE_ENDPRIM:
1746       return FALSE;
1747       break;
1748
1749    case TGSI_OPCODE_NOP:
1750       break;
1751
1752    default:
1753       return FALSE;
1754    }
1755
1756    if(info->num_dst) {
1757       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1758          emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
1759       }
1760    }
1761
1762    return TRUE;
1763 }
1764
1765
1766 void
1767 lp_build_tgsi_soa(LLVMBuilderRef builder,
1768                   const struct tgsi_token *tokens,
1769                   struct lp_type type,
1770                   struct lp_build_mask_context *mask,
1771                   LLVMValueRef consts_ptr,
1772                   const LLVMValueRef *pos,
1773                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1774                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1775                   struct lp_build_sampler_soa *sampler,
1776                   struct tgsi_shader_info *info)
1777 {
1778    struct lp_build_tgsi_soa_context bld;
1779    struct tgsi_parse_context parse;
1780    uint num_immediates = 0;
1781    unsigned i;
1782
1783    /* Setup build context */
1784    memset(&bld, 0, sizeof bld);
1785    lp_build_context_init(&bld.base, builder, type);
1786    bld.mask = mask;
1787    bld.pos = pos;
1788    bld.inputs = inputs;
1789    bld.outputs = outputs;
1790    bld.consts_ptr = consts_ptr;
1791    bld.sampler = sampler;
1792    bld.has_indirect_addressing = info->opcode_count[TGSI_OPCODE_ARR] > 0 ||
1793                                  info->opcode_count[TGSI_OPCODE_ARL] > 0;
1794
1795    lp_exec_mask_init(&bld.exec_mask, &bld.base);
1796
1797    tgsi_parse_init( &parse, tokens );
1798
1799    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1800       tgsi_parse_token( &parse );
1801
1802       switch( parse.FullToken.Token.Type ) {
1803       case TGSI_TOKEN_TYPE_DECLARATION:
1804          /* Inputs already interpolated */
1805          emit_declaration( &bld, &parse.FullToken.FullDeclaration );
1806          break;
1807
1808       case TGSI_TOKEN_TYPE_INSTRUCTION:
1809          {
1810             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1811             const struct tgsi_opcode_info *opcode_info = tgsi_get_opcode_info(opcode);
1812             if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, opcode_info ))
1813                _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1814                              opcode_info->mnemonic);
1815          }
1816
1817          break;
1818
1819       case TGSI_TOKEN_TYPE_IMMEDIATE:
1820          /* simply copy the immediate values into the next immediates[] slot */
1821          {
1822             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1823             assert(size <= 4);
1824             assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
1825             for( i = 0; i < size; ++i )
1826                bld.immediates[num_immediates][i] =
1827                   lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
1828             for( i = size; i < 4; ++i )
1829                bld.immediates[num_immediates][i] = bld.base.undef;
1830             num_immediates++;
1831          }
1832          break;
1833
1834       case TGSI_TOKEN_TYPE_PROPERTY:
1835          break;
1836
1837       default:
1838          assert( 0 );
1839       }
1840    }
1841    if (0) {
1842       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1843       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1844       debug_printf("11111111111111111111111111111 \n");
1845       tgsi_dump(tokens, 0);
1846       LLVMDumpValue(function);
1847       debug_printf("2222222222222222222222222222 \n");
1848    }
1849    tgsi_parse_free( &parse );
1850 }
1851