src/gallium/auxiliary/gallivm/lp_bld_tgsi_aos.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2010 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 /**
  29  * @file
  30  * TGSI to LLVM IR translation -- AoS.
  31  *
  32  * FIXME:
  33  * - No control flow support: the existing control flow code should be factored
  34  * out into from the SoA code into a common module and shared.
  35  * - No derivatives. Derivate logic should be pluggable, just like the samplers.
  36  *
  37  * @author Jose Fonseca <jfonseca@vmware.com>
  38  */
  39
  40 #include "pipe/p_config.h"
  41 #include "pipe/p_shader_tokens.h"
  42 #include "util/u_debug.h"
  43 #include "util/u_math.h"
  44 #include "util/u_memory.h"
  45 #include "tgsi/tgsi_dump.h"
  46 #include "tgsi/tgsi_info.h"
  47 #include "tgsi/tgsi_parse.h"
  48 #include "tgsi/tgsi_util.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_quad.h"
  57 #include "lp_bld_tgsi.h"
  58 #include "lp_bld_limits.h"
  59 #include "lp_bld_debug.h"
  60
  61
  62 #define LP_MAX_INSTRUCTIONS 256
  63
  64
  65 struct lp_build_tgsi_aos_context
  66 {
  67    struct lp_build_context base;
  68
  69    /* Builder for integer masks and indices */
  70    struct lp_build_context int_bld;
  71
  72    /*
  73     * AoS swizzle used:
  74     * - swizzles[0] = red index
  75     * - swizzles[1] = green index
  76     * - swizzles[2] = blue index
  77     * - swizzles[3] = alpha index
  78     */
  79    unsigned char swizzles[4];
  80    unsigned char inv_swizzles[4];
  81
  82    LLVMValueRef consts_ptr;
  83    const LLVMValueRef *inputs;
  84    LLVMValueRef *outputs;
  85
  86    struct lp_build_sampler_aos *sampler;
  87
  88    LLVMValueRef immediates[LP_MAX_TGSI_IMMEDIATES];
  89    LLVMValueRef temps[LP_MAX_TGSI_TEMPS];
  90    LLVMValueRef addr[LP_MAX_TGSI_ADDRS];
  91    LLVMValueRef preds[LP_MAX_TGSI_PREDS];
  92
  93    /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
  94     * set in the indirect_files field.
  95     * The temps[] array above is unused then.
  96     */
  97    LLVMValueRef temps_array;
  98
  99    /** bitmask indicating which register files are accessed indirectly */
 100    unsigned indirect_files;
 101
 102    struct tgsi_full_instruction *instructions;
 103    uint max_instructions;
 104 };
 105
 106
 107 /**
 108  * Wrapper around lp_build_swizzle_aos which translates swizzles to another
 109  * ordering.
 110  */
 111 static LLVMValueRef
 112 swizzle_aos(struct lp_build_tgsi_aos_context *bld,
 113             LLVMValueRef a,
 114             unsigned swizzle_x,
 115             unsigned swizzle_y,
 116             unsigned swizzle_z,
 117             unsigned swizzle_w)
 118 {
 119    unsigned char swizzles[4];
 120
 121    assert(swizzle_x < 4);
 122    assert(swizzle_y < 4);
 123    assert(swizzle_z < 4);
 124    assert(swizzle_w < 4);
 125
 126    swizzles[bld->inv_swizzles[0]] = bld->swizzles[swizzle_x];
 127    swizzles[bld->inv_swizzles[1]] = bld->swizzles[swizzle_y];
 128    swizzles[bld->inv_swizzles[2]] = bld->swizzles[swizzle_z];
 129    swizzles[bld->inv_swizzles[3]] = bld->swizzles[swizzle_w];
 130
 131    return lp_build_swizzle_aos(&bld->base, a, swizzles);
 132 }
 133
 134
 135 static LLVMValueRef
 136 swizzle_scalar_aos(struct lp_build_tgsi_aos_context *bld,
 137                    LLVMValueRef a,
 138                    unsigned chan)
 139 {
 140    chan = bld->swizzles[chan];
 141    return lp_build_swizzle_scalar_aos(&bld->base, a, chan);
 142 }
 143
 144
 145 /**
 146  * Register fetch.
 147  */
 148 static LLVMValueRef
 149 emit_fetch(
 150    struct lp_build_tgsi_aos_context *bld,
 151    const struct tgsi_full_instruction *inst,
 152    unsigned src_op)
 153 {
 154    struct lp_type type = bld->base.type;
 155    const struct tgsi_full_src_register *reg = &inst->Src[src_op];
 156    LLVMValueRef res;
 157    unsigned chan;
 158
 159    assert(!reg->Register.Indirect);
 160
 161    /*
 162     * Fetch the from the register file.
 163     */
 164
 165    switch (reg->Register.File) {
 166    case TGSI_FILE_CONSTANT:
 167       /*
 168        * Get the constants components
 169        */
 170
 171       res = bld->base.undef;
 172       for (chan = 0; chan < 4; ++chan) {
 173          LLVMValueRef index;
 174          LLVMValueRef scalar_ptr;
 175          LLVMValueRef scalar;
 176          LLVMValueRef swizzle;
 177
 178          index = LLVMConstInt(LLVMInt32Type(),
 179                               reg->Register.Index*4 + chan,
 180                               0);
 181
 182          scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr,
 183                                    &index, 1, "");
 184
 185          scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
 186
 187          lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
 188
 189          /*
 190           * NOTE: constants array is always assumed to be RGBA
 191           */
 192
 193          swizzle = LLVMConstInt(LLVMInt32Type(), chan, 0);
 194
 195          res = LLVMBuildInsertElement(bld->base.builder, res, scalar, swizzle, "");
 196       }
 197
 198       /*
 199        * Broadcast the first quaternion to all others.
 200        *
 201        * XXX: could be factored into a reusable function.
 202        */
 203
 204       if (type.length > 4) {
 205          LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 206          unsigned i;
 207
 208          for (chan = 0; chan < 4; ++chan) {
 209             shuffles[chan] = LLVMConstInt(LLVMInt32Type(), chan, 0);
 210          }
 211
 212          for (i = 4; i < type.length; ++i) {
 213             shuffles[i] = shuffles[i % 4];
 214          }
 215
 216          res = LLVMBuildShuffleVector(bld->base.builder,
 217                                       res, bld->base.undef,
 218                                       LLVMConstVector(shuffles, type.length),
 219                                       "");
 220       }
 221       break;
 222
 223    case TGSI_FILE_IMMEDIATE:
 224       res = bld->immediates[reg->Register.Index];
 225       assert(res);
 226       break;
 227
 228    case TGSI_FILE_INPUT:
 229       res = bld->inputs[reg->Register.Index];
 230       assert(res);
 231       break;
 232
 233    case TGSI_FILE_TEMPORARY:
 234       {
 235          LLVMValueRef temp_ptr;
 236          temp_ptr = bld->temps[reg->Register.Index];
 237          res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
 238          if (!res)
 239             return bld->base.undef;
 240       }
 241       break;
 242
 243    default:
 244       assert(0 && "invalid src register in emit_fetch()");
 245       return bld->base.undef;
 246    }
 247
 248    /*
 249     * Apply sign modifier.
 250     */
 251
 252    if (reg->Register.Absolute) {
 253       res = lp_build_abs(&bld->base, res);
 254    }
 255
 256    if(reg->Register.Negate) {
 257       res = lp_build_negate(&bld->base, res);
 258    }
 259
 260    /*
 261     * Swizzle the argument
 262     */
 263
 264    res = swizzle_aos(bld, res,
 265                      reg->Register.SwizzleX,
 266                      reg->Register.SwizzleY,
 267                      reg->Register.SwizzleZ,
 268                      reg->Register.SwizzleW);
 269
 270    return res;
 271 }
 272
 273
 274 /**
 275  * Register store.
 276  */
 277 static void
 278 emit_store(
 279    struct lp_build_tgsi_aos_context *bld,
 280    const struct tgsi_full_instruction *inst,
 281    unsigned index,
 282    LLVMValueRef value)
 283 {
 284    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 285    LLVMValueRef mask = NULL;
 286    LLVMValueRef ptr;
 287
 288    /*
 289     * Saturate the value
 290     */
 291
 292    switch (inst->Instruction.Saturate) {
 293    case TGSI_SAT_NONE:
 294       break;
 295
 296    case TGSI_SAT_ZERO_ONE:
 297       value = lp_build_max(&bld->base, value, bld->base.zero);
 298       value = lp_build_min(&bld->base, value, bld->base.one);
 299       break;
 300
 301    case TGSI_SAT_MINUS_PLUS_ONE:
 302       value = lp_build_max(&bld->base, value, lp_build_const_vec(bld->base.type, -1.0));
 303       value = lp_build_min(&bld->base, value, bld->base.one);
 304       break;
 305
 306    default:
 307       assert(0);
 308    }
 309
 310    /*
 311     * Translate the register file
 312     */
 313
 314    assert(!reg->Register.Indirect);
 315
 316    switch (reg->Register.File) {
 317    case TGSI_FILE_OUTPUT:
 318       ptr = bld->outputs[reg->Register.Index];
 319       break;
 320
 321    case TGSI_FILE_TEMPORARY:
 322       ptr = bld->temps[reg->Register.Index];
 323       break;
 324
 325    case TGSI_FILE_ADDRESS:
 326       ptr = bld->addr[reg->Indirect.Index];
 327       break;
 328
 329    case TGSI_FILE_PREDICATE:
 330       ptr = bld->preds[reg->Register.Index];
 331       break;
 332
 333    default:
 334       assert(0);
 335       return;
 336    }
 337
 338    /*
 339     * Predicate
 340     */
 341
 342    if (inst->Instruction.Predicate) {
 343       LLVMValueRef pred;
 344
 345       assert(inst->Predicate.Index < LP_MAX_TGSI_PREDS);
 346
 347       pred = LLVMBuildLoad(bld->base.builder,
 348                            bld->preds[inst->Predicate.Index], "");
 349
 350       /*
 351        * Convert the value to an integer mask.
 352        */
 353       pred = lp_build_compare(bld->base.builder,
 354                                bld->base.type,
 355                                PIPE_FUNC_NOTEQUAL,
 356                                pred,
 357                                bld->base.zero);
 358
 359       if (inst->Predicate.Negate) {
 360          pred = LLVMBuildNot(bld->base.builder, pred, "");
 361       }
 362
 363       pred = swizzle_aos(bld, pred,
 364                          inst->Predicate.SwizzleX,
 365                          inst->Predicate.SwizzleY,
 366                          inst->Predicate.SwizzleZ,
 367                          inst->Predicate.SwizzleW);
 368
 369       if (mask) {
 370          mask = LLVMBuildAnd(bld->base.builder, mask, pred, "");
 371       } else {
 372          mask = pred;
 373       }
 374    }
 375
 376    /*
 377     * Writemask
 378     */
 379
 380    if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
 381       LLVMValueRef writemask;
 382
 383       writemask = lp_build_const_mask_aos(bld->base.type, reg->Register.WriteMask);
 384
 385       if (mask) {
 386          mask = LLVMBuildAnd(bld->base.builder, mask, writemask, "");
 387       } else {
 388          mask = writemask;
 389       }
 390    }
 391
 392    if (mask) {
 393       LLVMValueRef orig_value;
 394
 395       orig_value = LLVMBuildLoad(bld->base.builder, ptr, "");
 396       value = lp_build_select(&bld->base,
 397                               mask, value, orig_value);
 398    }
 399
 400    LLVMBuildStore(bld->base.builder, value, ptr);
 401 }
 402
 403
 404 /**
 405  * High-level instruction translators.
 406  */
 407
 408 static LLVMValueRef
 409 emit_tex(struct lp_build_tgsi_aos_context *bld,
 410          const struct tgsi_full_instruction *inst,
 411          enum lp_build_tex_modifier modifier)
 412 {
 413    unsigned target;
 414    unsigned unit;
 415    LLVMValueRef coords;
 416    LLVMValueRef ddx;
 417    LLVMValueRef ddy;
 418
 419    if (!bld->sampler) {
 420       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 421       return bld->base.undef;
 422    }
 423
 424    target = inst->Texture.Texture;
 425
 426    coords = emit_fetch( bld, inst, 0 );
 427
 428    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 429       ddx = emit_fetch( bld, inst, 1 );
 430       ddy = emit_fetch( bld, inst, 2 );
 431       unit = inst->Src[3].Register.Index;
 432    }  else {
 433 #if 0
 434       ddx = lp_build_ddx( &bld->base, coords );
 435       ddy = lp_build_ddy( &bld->base, coords );
 436 #else
 437       /* TODO */
 438       ddx = bld->base.one;
 439       ddy = bld->base.one;
 440 #endif
 441       unit = inst->Src[1].Register.Index;
 442    }
 443
 444    return bld->sampler->emit_fetch_texel(bld->sampler,
 445                                          &bld->base,
 446                                          target, unit,
 447                                          coords, ddx, ddy,
 448                                          modifier);
 449 }
 450
 451
 452 static void
 453 emit_declaration(
 454    struct lp_build_tgsi_aos_context *bld,
 455    const struct tgsi_full_declaration *decl)
 456 {
 457    LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
 458
 459    unsigned first = decl->Range.First;
 460    unsigned last = decl->Range.Last;
 461    unsigned idx;
 462
 463    for (idx = first; idx <= last; ++idx) {
 464       switch (decl->Declaration.File) {
 465       case TGSI_FILE_TEMPORARY:
 466          assert(idx < LP_MAX_TGSI_TEMPS);
 467          if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 468             LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
 469                                                    last + 1, 0);
 470             bld->temps_array = lp_build_array_alloca(bld->base.builder,
 471                                                      vec_type, array_size, "");
 472          } else {
 473             bld->temps[idx] = lp_build_alloca(bld->base.builder,
 474                                               vec_type, "");
 475          }
 476          break;
 477
 478       case TGSI_FILE_OUTPUT:
 479          bld->outputs[idx] = lp_build_alloca(bld->base.builder,
 480                                              vec_type, "");
 481          break;
 482
 483       case TGSI_FILE_ADDRESS:
 484          assert(idx < LP_MAX_TGSI_ADDRS);
 485          bld->addr[idx] = lp_build_alloca(bld->base.builder,
 486                                           vec_type, "");
 487          break;
 488
 489       case TGSI_FILE_PREDICATE:
 490          assert(idx < LP_MAX_TGSI_PREDS);
 491          bld->preds[idx] = lp_build_alloca(bld->base.builder,
 492                                            vec_type, "");
 493          break;
 494
 495       default:
 496          /* don't need to declare other vars */
 497          break;
 498       }
 499    }
 500 }
 501
 502
 503 /**
 504  * Emit LLVM for one TGSI instruction.
 505  * \param return TRUE for success, FALSE otherwise
 506  */
 507 static boolean
 508 emit_instruction(
 509    struct lp_build_tgsi_aos_context *bld,
 510    const struct tgsi_full_instruction *inst,
 511    const struct tgsi_opcode_info *info,
 512    int *pc)
 513 {
 514    LLVMValueRef src0, src1, src2;
 515    LLVMValueRef tmp0, tmp1;
 516    LLVMValueRef dst0 = NULL;
 517
 518    /*
 519     * Stores and write masks are handled in a general fashion after the long
 520     * instruction opcode switch statement.
 521     *
 522     * Although not stricitly necessary, we avoid generating instructions for
 523     * channels which won't be stored, in cases where's that easy. For some
 524     * complex instructions, like texture sampling, it is more convenient to
 525     * assume a full writemask and then let LLVM optimization passes eliminate
 526     * redundant code.
 527     */
 528
 529    (*pc)++;
 530
 531    assert(info->num_dst <= 1);
 532    if (info->num_dst) {
 533       dst0 = bld->base.undef;
 534    }
 535
 536    switch (inst->Instruction.Opcode) {
 537    case TGSI_OPCODE_ARL:
 538       src0 = emit_fetch(bld, inst, 0);
 539       dst0 = lp_build_floor(&bld->base, src0);
 540       break;
 541
 542    case TGSI_OPCODE_MOV:
 543       dst0 = emit_fetch(bld, inst, 0);
 544       break;
 545
 546    case TGSI_OPCODE_LIT:
 547       return FALSE;
 548
 549    case TGSI_OPCODE_RCP:
 550    /* TGSI_OPCODE_RECIP */
 551       src0 = emit_fetch(bld, inst, 0);
 552       dst0 = lp_build_rcp(&bld->base, src0);
 553       break;
 554
 555    case TGSI_OPCODE_RSQ:
 556    /* TGSI_OPCODE_RECIPSQRT */
 557       src0 = emit_fetch(bld, inst, 0);
 558       tmp0 = lp_build_abs(&bld->base, src0);
 559       dst0 = lp_build_rsqrt(&bld->base, tmp0);
 560       break;
 561
 562    case TGSI_OPCODE_EXP:
 563       return FALSE;
 564
 565    case TGSI_OPCODE_LOG:
 566       return FALSE;
 567
 568    case TGSI_OPCODE_MUL:
 569       src0 = emit_fetch(bld, inst, 0);
 570       src1 = emit_fetch(bld, inst, 1);
 571       dst0 = lp_build_mul(&bld->base, src0, src1);
 572       break;
 573
 574    case TGSI_OPCODE_ADD:
 575       src0 = emit_fetch(bld, inst, 0);
 576       src1 = emit_fetch(bld, inst, 1);
 577       dst0 = lp_build_add(&bld->base, src0, src1);
 578       break;
 579
 580    case TGSI_OPCODE_DP3:
 581    /* TGSI_OPCODE_DOT3 */
 582       return FALSE;
 583
 584    case TGSI_OPCODE_DP4:
 585    /* TGSI_OPCODE_DOT4 */
 586       return FALSE;
 587
 588    case TGSI_OPCODE_DST:
 589       return FALSE;
 590
 591    case TGSI_OPCODE_MIN:
 592       src0 = emit_fetch(bld, inst, 0);
 593       src1 = emit_fetch(bld, inst, 1);
 594       dst0 = lp_build_max(&bld->base, src0, src1);
 595       break;
 596
 597    case TGSI_OPCODE_MAX:
 598       src0 = emit_fetch(bld, inst, 0);
 599       src1 = emit_fetch(bld, inst, 1);
 600       dst0 = lp_build_max(&bld->base, src0, src1);
 601       break;
 602
 603    case TGSI_OPCODE_SLT:
 604    /* TGSI_OPCODE_SETLT */
 605       src0 = emit_fetch(bld, inst, 0);
 606       src1 = emit_fetch(bld, inst, 1);
 607       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, src1);
 608       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 609       break;
 610
 611    case TGSI_OPCODE_SGE:
 612    /* TGSI_OPCODE_SETGE */
 613       src0 = emit_fetch(bld, inst, 0);
 614       src1 = emit_fetch(bld, inst, 1);
 615       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, src0, src1);
 616       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 617       break;
 618
 619    case TGSI_OPCODE_MAD:
 620    /* TGSI_OPCODE_MADD */
 621       src0 = emit_fetch(bld, inst, 0);
 622       src1 = emit_fetch(bld, inst, 1);
 623       src2 = emit_fetch(bld, inst, 2);
 624       tmp0 = lp_build_mul(&bld->base, src0, src1);
 625       dst0 = lp_build_add(&bld->base, tmp0, src2);
 626       break;
 627
 628    case TGSI_OPCODE_SUB:
 629       src0 = emit_fetch(bld, inst, 0);
 630       src1 = emit_fetch(bld, inst, 1);
 631       dst0 = lp_build_sub(&bld->base, src0, src1);
 632       break;
 633
 634    case TGSI_OPCODE_LRP:
 635       src0 = emit_fetch(bld, inst, 0);
 636       src1 = emit_fetch(bld, inst, 1);
 637       src2 = emit_fetch(bld, inst, 2);
 638       tmp0 = lp_build_sub(&bld->base, src1, src2);
 639       tmp0 = lp_build_mul(&bld->base, src0, tmp0);
 640       dst0 = lp_build_add(&bld->base, tmp0, src2);
 641       break;
 642
 643    case TGSI_OPCODE_CND:
 644       src0 = emit_fetch(bld, inst, 0);
 645       src1 = emit_fetch(bld, inst, 1);
 646       src2 = emit_fetch(bld, inst, 2);
 647       tmp1 = lp_build_const_vec(bld->base.type, 0.5);
 648       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src2, tmp1);
 649       dst0 = lp_build_select(&bld->base, tmp0, src0, src1);
 650       break;
 651
 652    case TGSI_OPCODE_DP2A:
 653       return FALSE;
 654
 655    case TGSI_OPCODE_FRC:
 656       src0 = emit_fetch(bld, inst, 0);
 657       tmp0 = lp_build_floor(&bld->base, src0);
 658       dst0 = lp_build_sub(&bld->base, src0, tmp0);
 659       break;
 660
 661    case TGSI_OPCODE_CLAMP:
 662       src0 = emit_fetch(bld, inst, 0);
 663       src1 = emit_fetch(bld, inst, 1);
 664       src2 = emit_fetch(bld, inst, 2);
 665       tmp0 = lp_build_max(&bld->base, src0, src1);
 666       dst0 = lp_build_min(&bld->base, tmp0, src2);
 667       break;
 668
 669    case TGSI_OPCODE_FLR:
 670       src0 = emit_fetch(bld, inst, 0);
 671       dst0 = lp_build_floor(&bld->base, src0);
 672       break;
 673
 674    case TGSI_OPCODE_ROUND:
 675       src0 = emit_fetch(bld, inst, 0);
 676       dst0 = lp_build_round(&bld->base, src0);
 677       break;
 678
 679    case TGSI_OPCODE_EX2:
 680       src0 = emit_fetch(bld, inst, 0);
 681       tmp0 = lp_build_swizzle_scalar_aos(&bld->base, src0, TGSI_SWIZZLE_X);
 682       dst0 = lp_build_exp2(&bld->base, tmp0);
 683       break;
 684
 685    case TGSI_OPCODE_LG2:
 686       src0 = emit_fetch(bld, inst, 0);
 687       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 688       dst0 = lp_build_log2(&bld->base, tmp0);
 689       break;
 690
 691    case TGSI_OPCODE_POW:
 692       src0 = emit_fetch(bld, inst, 0);
 693       src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 694       src1 = emit_fetch(bld, inst, 1);
 695       src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
 696       dst0 = lp_build_pow(&bld->base, src0, src1);
 697       break;
 698
 699    case TGSI_OPCODE_XPD:
 700       return FALSE;
 701
 702    case TGSI_OPCODE_ABS:
 703       src0 = emit_fetch(bld, inst, 0);
 704       dst0 = lp_build_abs(&bld->base, src0);
 705       break;
 706
 707    case TGSI_OPCODE_RCC:
 708       /* deprecated? */
 709       assert(0);
 710       return FALSE;
 711
 712    case TGSI_OPCODE_DPH:
 713       return FALSE;
 714
 715    case TGSI_OPCODE_COS:
 716       src0 = emit_fetch(bld, inst, 0);
 717       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 718       dst0 = lp_build_cos(&bld->base, tmp0);
 719       break;
 720
 721    case TGSI_OPCODE_DDX:
 722       return FALSE;
 723
 724    case TGSI_OPCODE_DDY:
 725       return FALSE;
 726
 727    case TGSI_OPCODE_KILP:
 728       /* predicated kill */
 729       return FALSE;
 730
 731    case TGSI_OPCODE_KIL:
 732       /* conditional kill */
 733       return FALSE;
 734
 735    case TGSI_OPCODE_PK2H:
 736       return FALSE;
 737       break;
 738
 739    case TGSI_OPCODE_PK2US:
 740       return FALSE;
 741       break;
 742
 743    case TGSI_OPCODE_PK4B:
 744       return FALSE;
 745       break;
 746
 747    case TGSI_OPCODE_PK4UB:
 748       return FALSE;
 749
 750    case TGSI_OPCODE_RFL:
 751       return FALSE;
 752
 753    case TGSI_OPCODE_SEQ:
 754       src0 = emit_fetch(bld, inst, 0);
 755       src1 = emit_fetch(bld, inst, 1);
 756       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_EQUAL, src0, src1);
 757       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 758       break;
 759
 760    case TGSI_OPCODE_SFL:
 761       dst0 = bld->base.zero;
 762       break;
 763
 764    case TGSI_OPCODE_SGT:
 765       src0 = emit_fetch(bld, inst, 0);
 766       src1 = emit_fetch(bld, inst, 1);
 767       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src0, src1);
 768       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 769       break;
 770
 771    case TGSI_OPCODE_SIN:
 772       src0 = emit_fetch(bld, inst, 0);
 773       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 774       dst0 = lp_build_sin(&bld->base, tmp0);
 775       break;
 776
 777    case TGSI_OPCODE_SLE:
 778       src0 = emit_fetch(bld, inst, 0);
 779       src1 = emit_fetch(bld, inst, 1);
 780       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LEQUAL, src0, src1);
 781       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 782       break;
 783
 784    case TGSI_OPCODE_SNE:
 785       src0 = emit_fetch(bld, inst, 0);
 786       src1 = emit_fetch(bld, inst, 1);
 787       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL, src0, src1);
 788       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 789       break;
 790
 791    case TGSI_OPCODE_STR:
 792       dst0 = bld->base.one;
 793       break;
 794
 795    case TGSI_OPCODE_TEX:
 796       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 797       break;
 798
 799    case TGSI_OPCODE_TXD:
 800       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 801       break;
 802
 803    case TGSI_OPCODE_UP2H:
 804       /* deprecated */
 805       assert (0);
 806       return FALSE;
 807       break;
 808
 809    case TGSI_OPCODE_UP2US:
 810       /* deprecated */
 811       assert(0);
 812       return FALSE;
 813       break;
 814
 815    case TGSI_OPCODE_UP4B:
 816       /* deprecated */
 817       assert(0);
 818       return FALSE;
 819       break;
 820
 821    case TGSI_OPCODE_UP4UB:
 822       /* deprecated */
 823       assert(0);
 824       return FALSE;
 825       break;
 826
 827    case TGSI_OPCODE_X2D:
 828       /* deprecated? */
 829       assert(0);
 830       return FALSE;
 831       break;
 832
 833    case TGSI_OPCODE_ARA:
 834       /* deprecated */
 835       assert(0);
 836       return FALSE;
 837       break;
 838
 839    case TGSI_OPCODE_ARR:
 840       src0 = emit_fetch(bld, inst, 0);
 841       dst0 = lp_build_round(&bld->base, src0);
 842       break;
 843
 844    case TGSI_OPCODE_BRA:
 845       /* deprecated */
 846       assert(0);
 847       return FALSE;
 848       break;
 849
 850    case TGSI_OPCODE_CAL:
 851       return FALSE;
 852
 853    case TGSI_OPCODE_RET:
 854       return FALSE;
 855
 856    case TGSI_OPCODE_END:
 857       *pc = -1;
 858       break;
 859
 860    case TGSI_OPCODE_SSG:
 861    /* TGSI_OPCODE_SGN */
 862       tmp0 = emit_fetch(bld, inst, 0);
 863       dst0 = lp_build_sgn(&bld->base, tmp0);
 864       break;
 865
 866    case TGSI_OPCODE_CMP:
 867       src0 = emit_fetch(bld, inst, 0);
 868       src1 = emit_fetch(bld, inst, 1);
 869       src2 = emit_fetch(bld, inst, 2);
 870       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, bld->base.zero);
 871       dst0 = lp_build_select(&bld->base, tmp0, src1, src2);
 872       break;
 873
 874    case TGSI_OPCODE_SCS:
 875       return FALSE;
 876
 877    case TGSI_OPCODE_TXB:
 878       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 879       break;
 880
 881    case TGSI_OPCODE_NRM:
 882       /* fall-through */
 883    case TGSI_OPCODE_NRM4:
 884       return FALSE;
 885
 886    case TGSI_OPCODE_DIV:
 887       /* deprecated */
 888       assert(0);
 889       return FALSE;
 890       break;
 891
 892    case TGSI_OPCODE_DP2:
 893       return FALSE;
 894
 895    case TGSI_OPCODE_TXL:
 896       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 897       break;
 898
 899    case TGSI_OPCODE_TXP:
 900       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 901       break;
 902
 903    case TGSI_OPCODE_BRK:
 904       return FALSE;
 905
 906    case TGSI_OPCODE_IF:
 907       return FALSE;
 908
 909    case TGSI_OPCODE_BGNLOOP:
 910       return FALSE;
 911
 912    case TGSI_OPCODE_BGNSUB:
 913       return FALSE;
 914
 915    case TGSI_OPCODE_ELSE:
 916       return FALSE;
 917
 918    case TGSI_OPCODE_ENDIF:
 919       return FALSE;
 920
 921    case TGSI_OPCODE_ENDLOOP:
 922       return FALSE;
 923
 924    case TGSI_OPCODE_ENDSUB:
 925       return FALSE;
 926
 927    case TGSI_OPCODE_PUSHA:
 928       /* deprecated? */
 929       assert(0);
 930       return FALSE;
 931       break;
 932
 933    case TGSI_OPCODE_POPA:
 934       /* deprecated? */
 935       assert(0);
 936       return FALSE;
 937       break;
 938
 939    case TGSI_OPCODE_CEIL:
 940       src0 = emit_fetch(bld, inst, 0);
 941       dst0 = lp_build_ceil(&bld->base, src0);
 942       break;
 943
 944    case TGSI_OPCODE_I2F:
 945       /* deprecated? */
 946       assert(0);
 947       return FALSE;
 948       break;
 949
 950    case TGSI_OPCODE_NOT:
 951       /* deprecated? */
 952       assert(0);
 953       return FALSE;
 954       break;
 955
 956    case TGSI_OPCODE_TRUNC:
 957       src0 = emit_fetch(bld, inst, 0);
 958       dst0 = lp_build_trunc(&bld->base, src0);
 959       break;
 960
 961    case TGSI_OPCODE_SHL:
 962       /* deprecated? */
 963       assert(0);
 964       return FALSE;
 965       break;
 966
 967    case TGSI_OPCODE_ISHR:
 968       /* deprecated? */
 969       assert(0);
 970       return FALSE;
 971       break;
 972
 973    case TGSI_OPCODE_AND:
 974       /* deprecated? */
 975       assert(0);
 976       return FALSE;
 977       break;
 978
 979    case TGSI_OPCODE_OR:
 980       /* deprecated? */
 981       assert(0);
 982       return FALSE;
 983       break;
 984
 985    case TGSI_OPCODE_MOD:
 986       /* deprecated? */
 987       assert(0);
 988       return FALSE;
 989       break;
 990
 991    case TGSI_OPCODE_XOR:
 992       /* deprecated? */
 993       assert(0);
 994       return FALSE;
 995       break;
 996
 997    case TGSI_OPCODE_SAD:
 998       /* deprecated? */
 999       assert(0);
1000       return FALSE;
1001       break;
1002
1003    case TGSI_OPCODE_TXF:
1004       /* deprecated? */
1005       assert(0);
1006       return FALSE;
1007       break;
1008
1009    case TGSI_OPCODE_TXQ:
1010       /* deprecated? */
1011       assert(0);
1012       return FALSE;
1013       break;
1014
1015    case TGSI_OPCODE_CONT:
1016       return FALSE;
1017
1018    case TGSI_OPCODE_EMIT:
1019       return FALSE;
1020       break;
1021
1022    case TGSI_OPCODE_ENDPRIM:
1023       return FALSE;
1024       break;
1025
1026    case TGSI_OPCODE_NOP:
1027       break;
1028
1029    default:
1030       return FALSE;
1031    }
1032
1033    if (info->num_dst) {
1034       emit_store(bld, inst, 0, dst0);
1035    }
1036
1037    return TRUE;
1038 }
1039
1040
1041 void
1042 lp_build_tgsi_aos(LLVMBuilderRef builder,
1043                   const struct tgsi_token *tokens,
1044                   struct lp_type type,
1045                   const unsigned char swizzles[4],
1046                   LLVMValueRef consts_ptr,
1047                   const LLVMValueRef *inputs,
1048                   LLVMValueRef *outputs,
1049                   struct lp_build_sampler_aos *sampler,
1050                   const struct tgsi_shader_info *info)
1051 {
1052    struct lp_build_tgsi_aos_context bld;
1053    struct tgsi_parse_context parse;
1054    uint num_immediates = 0;
1055    uint num_instructions = 0;
1056    unsigned chan;
1057    int pc = 0;
1058
1059    /* Setup build context */
1060    memset(&bld, 0, sizeof bld);
1061    lp_build_context_init(&bld.base, builder, type);
1062    lp_build_context_init(&bld.int_bld, builder, lp_int_type(type));
1063
1064    for (chan = 0; chan < 4; ++chan) {
1065       bld.swizzles[chan] = swizzles[chan];
1066       bld.inv_swizzles[swizzles[chan]] = chan;
1067    }
1068
1069    bld.inputs = inputs;
1070    bld.outputs = outputs;
1071    bld.consts_ptr = consts_ptr;
1072    bld.sampler = sampler;
1073    bld.indirect_files = info->indirect_files;
1074    bld.instructions = (struct tgsi_full_instruction *)
1075                       MALLOC(LP_MAX_INSTRUCTIONS * sizeof(struct tgsi_full_instruction));
1076    bld.max_instructions = LP_MAX_INSTRUCTIONS;
1077
1078    if (!bld.instructions) {
1079       return;
1080    }
1081
1082    tgsi_parse_init(&parse, tokens);
1083
1084    while (!tgsi_parse_end_of_tokens(&parse)) {
1085       tgsi_parse_token(&parse);
1086
1087       switch(parse.FullToken.Token.Type) {
1088       case TGSI_TOKEN_TYPE_DECLARATION:
1089          /* Inputs already interpolated */
1090          emit_declaration(&bld, &parse.FullToken.FullDeclaration);
1091          break;
1092
1093       case TGSI_TOKEN_TYPE_INSTRUCTION:
1094          {
1095             /* save expanded instruction */
1096             if (num_instructions == bld.max_instructions) {
1097                struct tgsi_full_instruction *instructions;
1098                instructions = REALLOC(bld.instructions,
1099                                       bld.max_instructions
1100                                       * sizeof(struct tgsi_full_instruction),
1101                                       (bld.max_instructions + LP_MAX_INSTRUCTIONS)
1102                                       * sizeof(struct tgsi_full_instruction));
1103                if (!instructions) {
1104                   break;
1105                }
1106                bld.instructions = instructions;
1107                bld.max_instructions += LP_MAX_INSTRUCTIONS;
1108             }
1109
1110             memcpy(bld.instructions + num_instructions,
1111                    &parse.FullToken.FullInstruction,
1112                    sizeof(bld.instructions[0]));
1113
1114             num_instructions++;
1115          }
1116
1117          break;
1118
1119       case TGSI_TOKEN_TYPE_IMMEDIATE:
1120          /* simply copy the immediate values into the next immediates[] slot */
1121          {
1122             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1123             float imm[4];
1124             assert(size <= 4);
1125             assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
1126             for (chan = 0; chan < 4; ++chan) {
1127                imm[chan] = 0.0f;
1128             }
1129             for (chan = 0; chan < size; ++chan) {
1130                unsigned swizzle = bld.swizzles[chan];
1131                imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
1132             }
1133             bld.immediates[num_immediates] =
1134                      lp_build_const_aos(type,
1135                                         imm[0], imm[1], imm[2], imm[3],
1136                                         NULL);
1137             num_immediates++;
1138          }
1139          break;
1140
1141       case TGSI_TOKEN_TYPE_PROPERTY:
1142          break;
1143
1144       default:
1145          assert(0);
1146       }
1147    }
1148
1149    while (pc != -1) {
1150       struct tgsi_full_instruction *instr = bld.instructions + pc;
1151       const struct tgsi_opcode_info *opcode_info =
1152          tgsi_get_opcode_info(instr->Instruction.Opcode);
1153       if (!emit_instruction(&bld, instr, opcode_info, &pc))
1154          _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1155                        opcode_info->mnemonic);
1156    }
1157
1158    if (0) {
1159       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1160       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1161       debug_printf("11111111111111111111111111111 \n");
1162       tgsi_dump(tokens, 0);
1163       lp_debug_dump_value(function);
1164       debug_printf("2222222222222222222222222222 \n");
1165    }
1166    tgsi_parse_free(&parse);
1167
1168    if (0) {
1169       LLVMModuleRef module = LLVMGetGlobalParent(
1170          LLVMGetBasicBlockParent(LLVMGetInsertBlock(bld.base.builder)));
1171       LLVMDumpModule(module);
1172    }
1173
1174    FREE(bld.instructions);
1175 }
1176