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 = lp_build_const_int32(bld->base.gallivm, reg->Register.Index * 4 + chan);
 179
 180          scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr,
 181                                    &index, 1, "");
 182
 183          scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
 184
 185          lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
 186
 187          /*
 188           * NOTE: constants array is always assumed to be RGBA
 189           */
 190
 191          swizzle = lp_build_const_int32(bld->base.gallivm, chan);
 192
 193          res = LLVMBuildInsertElement(bld->base.builder, res, scalar, swizzle, "");
 194       }
 195
 196       /*
 197        * Broadcast the first quaternion to all others.
 198        *
 199        * XXX: could be factored into a reusable function.
 200        */
 201
 202       if (type.length > 4) {
 203          LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 204          unsigned i;
 205
 206          for (chan = 0; chan < 4; ++chan) {
 207             shuffles[chan] = lp_build_const_int32(bld->base.gallivm, chan);
 208          }
 209
 210          for (i = 4; i < type.length; ++i) {
 211             shuffles[i] = shuffles[i % 4];
 212          }
 213
 214          res = LLVMBuildShuffleVector(bld->base.builder,
 215                                       res, bld->base.undef,
 216                                       LLVMConstVector(shuffles, type.length),
 217                                       "");
 218       }
 219       break;
 220
 221    case TGSI_FILE_IMMEDIATE:
 222       res = bld->immediates[reg->Register.Index];
 223       assert(res);
 224       break;
 225
 226    case TGSI_FILE_INPUT:
 227       res = bld->inputs[reg->Register.Index];
 228       assert(res);
 229       break;
 230
 231    case TGSI_FILE_TEMPORARY:
 232       {
 233          LLVMValueRef temp_ptr;
 234          temp_ptr = bld->temps[reg->Register.Index];
 235          res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
 236          if (!res)
 237             return bld->base.undef;
 238       }
 239       break;
 240
 241    default:
 242       assert(0 && "invalid src register in emit_fetch()");
 243       return bld->base.undef;
 244    }
 245
 246    /*
 247     * Apply sign modifier.
 248     */
 249
 250    if (reg->Register.Absolute) {
 251       res = lp_build_abs(&bld->base, res);
 252    }
 253
 254    if(reg->Register.Negate) {
 255       res = lp_build_negate(&bld->base, res);
 256    }
 257
 258    /*
 259     * Swizzle the argument
 260     */
 261
 262    res = swizzle_aos(bld, res,
 263                      reg->Register.SwizzleX,
 264                      reg->Register.SwizzleY,
 265                      reg->Register.SwizzleZ,
 266                      reg->Register.SwizzleW);
 267
 268    return res;
 269 }
 270
 271
 272 /**
 273  * Register store.
 274  */
 275 static void
 276 emit_store(
 277    struct lp_build_tgsi_aos_context *bld,
 278    const struct tgsi_full_instruction *inst,
 279    unsigned index,
 280    LLVMValueRef value)
 281 {
 282    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 283    LLVMValueRef mask = NULL;
 284    LLVMValueRef ptr;
 285
 286    /*
 287     * Saturate the value
 288     */
 289
 290    switch (inst->Instruction.Saturate) {
 291    case TGSI_SAT_NONE:
 292       break;
 293
 294    case TGSI_SAT_ZERO_ONE:
 295       value = lp_build_max(&bld->base, value, bld->base.zero);
 296       value = lp_build_min(&bld->base, value, bld->base.one);
 297       break;
 298
 299    case TGSI_SAT_MINUS_PLUS_ONE:
 300       value = lp_build_max(&bld->base, value, lp_build_const_vec(bld->base.gallivm, bld->base.type, -1.0));
 301       value = lp_build_min(&bld->base, value, bld->base.one);
 302       break;
 303
 304    default:
 305       assert(0);
 306    }
 307
 308    /*
 309     * Translate the register file
 310     */
 311
 312    assert(!reg->Register.Indirect);
 313
 314    switch (reg->Register.File) {
 315    case TGSI_FILE_OUTPUT:
 316       ptr = bld->outputs[reg->Register.Index];
 317       break;
 318
 319    case TGSI_FILE_TEMPORARY:
 320       ptr = bld->temps[reg->Register.Index];
 321       break;
 322
 323    case TGSI_FILE_ADDRESS:
 324       ptr = bld->addr[reg->Indirect.Index];
 325       break;
 326
 327    case TGSI_FILE_PREDICATE:
 328       ptr = bld->preds[reg->Register.Index];
 329       break;
 330
 331    default:
 332       assert(0);
 333       return;
 334    }
 335
 336    /*
 337     * Predicate
 338     */
 339
 340    if (inst->Instruction.Predicate) {
 341       LLVMValueRef pred;
 342
 343       assert(inst->Predicate.Index < LP_MAX_TGSI_PREDS);
 344
 345       pred = LLVMBuildLoad(bld->base.builder,
 346                            bld->preds[inst->Predicate.Index], "");
 347
 348       /*
 349        * Convert the value to an integer mask.
 350        */
 351       pred = lp_build_compare(bld->base.gallivm,
 352                                bld->base.type,
 353                                PIPE_FUNC_NOTEQUAL,
 354                                pred,
 355                                bld->base.zero);
 356
 357       if (inst->Predicate.Negate) {
 358          pred = LLVMBuildNot(bld->base.builder, pred, "");
 359       }
 360
 361       pred = swizzle_aos(bld, pred,
 362                          inst->Predicate.SwizzleX,
 363                          inst->Predicate.SwizzleY,
 364                          inst->Predicate.SwizzleZ,
 365                          inst->Predicate.SwizzleW);
 366
 367       if (mask) {
 368          mask = LLVMBuildAnd(bld->base.builder, mask, pred, "");
 369       } else {
 370          mask = pred;
 371       }
 372    }
 373
 374    /*
 375     * Writemask
 376     */
 377
 378    if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
 379       LLVMValueRef writemask;
 380
 381       writemask = lp_build_const_mask_aos(bld->base.gallivm, bld->base.type,
 382                                           reg->Register.WriteMask);
 383
 384       if (mask) {
 385          mask = LLVMBuildAnd(bld->base.builder, mask, writemask, "");
 386       } else {
 387          mask = writemask;
 388       }
 389    }
 390
 391    if (mask) {
 392       LLVMValueRef orig_value;
 393
 394       orig_value = LLVMBuildLoad(bld->base.builder, ptr, "");
 395       value = lp_build_select(&bld->base,
 396                               mask, value, orig_value);
 397    }
 398
 399    LLVMBuildStore(bld->base.builder, value, ptr);
 400 }
 401
 402
 403 /**
 404  * High-level instruction translators.
 405  */
 406
 407 static LLVMValueRef
 408 emit_tex(struct lp_build_tgsi_aos_context *bld,
 409          const struct tgsi_full_instruction *inst,
 410          enum lp_build_tex_modifier modifier)
 411 {
 412    unsigned target;
 413    unsigned unit;
 414    LLVMValueRef coords;
 415    LLVMValueRef ddx;
 416    LLVMValueRef ddy;
 417
 418    if (!bld->sampler) {
 419       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 420       return bld->base.undef;
 421    }
 422
 423    target = inst->Texture.Texture;
 424
 425    coords = emit_fetch( bld, inst, 0 );
 426
 427    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 428       ddx = emit_fetch( bld, inst, 1 );
 429       ddy = emit_fetch( bld, inst, 2 );
 430       unit = inst->Src[3].Register.Index;
 431    }  else {
 432 #if 0
 433       ddx = lp_build_ddx( &bld->base, coords );
 434       ddy = lp_build_ddy( &bld->base, coords );
 435 #else
 436       /* TODO */
 437       ddx = bld->base.one;
 438       ddy = bld->base.one;
 439 #endif
 440       unit = inst->Src[1].Register.Index;
 441    }
 442
 443    return bld->sampler->emit_fetch_texel(bld->sampler,
 444                                          &bld->base,
 445                                          target, unit,
 446                                          coords, ddx, ddy,
 447                                          modifier);
 448 }
 449
 450
 451 static void
 452 emit_declaration(
 453    struct lp_build_tgsi_aos_context *bld,
 454    const struct tgsi_full_declaration *decl)
 455 {
 456    struct gallivm_state *gallivm = bld->base.gallivm;
 457    LLVMTypeRef vec_type = lp_build_vec_type(bld->base.gallivm, 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 = lp_build_const_int32(gallivm, last + 1);
 469             bld->temps_array = lp_build_array_alloca(bld->base.gallivm,
 470                                                      vec_type, array_size, "");
 471          } else {
 472             bld->temps[idx] = lp_build_alloca(gallivm, vec_type, "");
 473          }
 474          break;
 475
 476       case TGSI_FILE_OUTPUT:
 477          bld->outputs[idx] = lp_build_alloca(gallivm, vec_type, "");
 478          break;
 479
 480       case TGSI_FILE_ADDRESS:
 481          assert(idx < LP_MAX_TGSI_ADDRS);
 482          bld->addr[idx] = lp_build_alloca(gallivm, vec_type, "");
 483          break;
 484
 485       case TGSI_FILE_PREDICATE:
 486          assert(idx < LP_MAX_TGSI_PREDS);
 487          bld->preds[idx] = lp_build_alloca(gallivm, vec_type, "");
 488          break;
 489
 490       default:
 491          /* don't need to declare other vars */
 492          break;
 493       }
 494    }
 495 }
 496
 497
 498 /**
 499  * Emit LLVM for one TGSI instruction.
 500  * \param return TRUE for success, FALSE otherwise
 501  */
 502 static boolean
 503 emit_instruction(
 504    struct lp_build_tgsi_aos_context *bld,
 505    const struct tgsi_full_instruction *inst,
 506    const struct tgsi_opcode_info *info,
 507    int *pc)
 508 {
 509    LLVMValueRef src0, src1, src2;
 510    LLVMValueRef tmp0, tmp1;
 511    LLVMValueRef dst0 = NULL;
 512
 513    /*
 514     * Stores and write masks are handled in a general fashion after the long
 515     * instruction opcode switch statement.
 516     *
 517     * Although not stricitly necessary, we avoid generating instructions for
 518     * channels which won't be stored, in cases where's that easy. For some
 519     * complex instructions, like texture sampling, it is more convenient to
 520     * assume a full writemask and then let LLVM optimization passes eliminate
 521     * redundant code.
 522     */
 523
 524    (*pc)++;
 525
 526    assert(info->num_dst <= 1);
 527    if (info->num_dst) {
 528       dst0 = bld->base.undef;
 529    }
 530
 531    switch (inst->Instruction.Opcode) {
 532    case TGSI_OPCODE_ARL:
 533       src0 = emit_fetch(bld, inst, 0);
 534       dst0 = lp_build_floor(&bld->base, src0);
 535       break;
 536
 537    case TGSI_OPCODE_MOV:
 538       dst0 = emit_fetch(bld, inst, 0);
 539       break;
 540
 541    case TGSI_OPCODE_LIT:
 542       return FALSE;
 543
 544    case TGSI_OPCODE_RCP:
 545    /* TGSI_OPCODE_RECIP */
 546       src0 = emit_fetch(bld, inst, 0);
 547       dst0 = lp_build_rcp(&bld->base, src0);
 548       break;
 549
 550    case TGSI_OPCODE_RSQ:
 551    /* TGSI_OPCODE_RECIPSQRT */
 552       src0 = emit_fetch(bld, inst, 0);
 553       tmp0 = lp_build_abs(&bld->base, src0);
 554       dst0 = lp_build_rsqrt(&bld->base, tmp0);
 555       break;
 556
 557    case TGSI_OPCODE_EXP:
 558       return FALSE;
 559
 560    case TGSI_OPCODE_LOG:
 561       return FALSE;
 562
 563    case TGSI_OPCODE_MUL:
 564       src0 = emit_fetch(bld, inst, 0);
 565       src1 = emit_fetch(bld, inst, 1);
 566       dst0 = lp_build_mul(&bld->base, src0, src1);
 567       break;
 568
 569    case TGSI_OPCODE_ADD:
 570       src0 = emit_fetch(bld, inst, 0);
 571       src1 = emit_fetch(bld, inst, 1);
 572       dst0 = lp_build_add(&bld->base, src0, src1);
 573       break;
 574
 575    case TGSI_OPCODE_DP3:
 576    /* TGSI_OPCODE_DOT3 */
 577       return FALSE;
 578
 579    case TGSI_OPCODE_DP4:
 580    /* TGSI_OPCODE_DOT4 */
 581       return FALSE;
 582
 583    case TGSI_OPCODE_DST:
 584       return FALSE;
 585
 586    case TGSI_OPCODE_MIN:
 587       src0 = emit_fetch(bld, inst, 0);
 588       src1 = emit_fetch(bld, inst, 1);
 589       dst0 = lp_build_max(&bld->base, src0, src1);
 590       break;
 591
 592    case TGSI_OPCODE_MAX:
 593       src0 = emit_fetch(bld, inst, 0);
 594       src1 = emit_fetch(bld, inst, 1);
 595       dst0 = lp_build_max(&bld->base, src0, src1);
 596       break;
 597
 598    case TGSI_OPCODE_SLT:
 599    /* TGSI_OPCODE_SETLT */
 600       src0 = emit_fetch(bld, inst, 0);
 601       src1 = emit_fetch(bld, inst, 1);
 602       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, src1);
 603       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 604       break;
 605
 606    case TGSI_OPCODE_SGE:
 607    /* TGSI_OPCODE_SETGE */
 608       src0 = emit_fetch(bld, inst, 0);
 609       src1 = emit_fetch(bld, inst, 1);
 610       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, src0, src1);
 611       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 612       break;
 613
 614    case TGSI_OPCODE_MAD:
 615    /* TGSI_OPCODE_MADD */
 616       src0 = emit_fetch(bld, inst, 0);
 617       src1 = emit_fetch(bld, inst, 1);
 618       src2 = emit_fetch(bld, inst, 2);
 619       tmp0 = lp_build_mul(&bld->base, src0, src1);
 620       dst0 = lp_build_add(&bld->base, tmp0, src2);
 621       break;
 622
 623    case TGSI_OPCODE_SUB:
 624       src0 = emit_fetch(bld, inst, 0);
 625       src1 = emit_fetch(bld, inst, 1);
 626       dst0 = lp_build_sub(&bld->base, src0, src1);
 627       break;
 628
 629    case TGSI_OPCODE_LRP:
 630       src0 = emit_fetch(bld, inst, 0);
 631       src1 = emit_fetch(bld, inst, 1);
 632       src2 = emit_fetch(bld, inst, 2);
 633       tmp0 = lp_build_sub(&bld->base, src1, src2);
 634       tmp0 = lp_build_mul(&bld->base, src0, tmp0);
 635       dst0 = lp_build_add(&bld->base, tmp0, src2);
 636       break;
 637
 638    case TGSI_OPCODE_CND:
 639       src0 = emit_fetch(bld, inst, 0);
 640       src1 = emit_fetch(bld, inst, 1);
 641       src2 = emit_fetch(bld, inst, 2);
 642       tmp1 = lp_build_const_vec(bld->base.gallivm, bld->base.type, 0.5);
 643       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src2, tmp1);
 644       dst0 = lp_build_select(&bld->base, tmp0, src0, src1);
 645       break;
 646
 647    case TGSI_OPCODE_DP2A:
 648       return FALSE;
 649
 650    case TGSI_OPCODE_FRC:
 651       src0 = emit_fetch(bld, inst, 0);
 652       tmp0 = lp_build_floor(&bld->base, src0);
 653       dst0 = lp_build_sub(&bld->base, src0, tmp0);
 654       break;
 655
 656    case TGSI_OPCODE_CLAMP:
 657       src0 = emit_fetch(bld, inst, 0);
 658       src1 = emit_fetch(bld, inst, 1);
 659       src2 = emit_fetch(bld, inst, 2);
 660       tmp0 = lp_build_max(&bld->base, src0, src1);
 661       dst0 = lp_build_min(&bld->base, tmp0, src2);
 662       break;
 663
 664    case TGSI_OPCODE_FLR:
 665       src0 = emit_fetch(bld, inst, 0);
 666       dst0 = lp_build_floor(&bld->base, src0);
 667       break;
 668
 669    case TGSI_OPCODE_ROUND:
 670       src0 = emit_fetch(bld, inst, 0);
 671       dst0 = lp_build_round(&bld->base, src0);
 672       break;
 673
 674    case TGSI_OPCODE_EX2:
 675       src0 = emit_fetch(bld, inst, 0);
 676       tmp0 = lp_build_swizzle_scalar_aos(&bld->base, src0, TGSI_SWIZZLE_X);
 677       dst0 = lp_build_exp2(&bld->base, tmp0);
 678       break;
 679
 680    case TGSI_OPCODE_LG2:
 681       src0 = emit_fetch(bld, inst, 0);
 682       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 683       dst0 = lp_build_log2(&bld->base, tmp0);
 684       break;
 685
 686    case TGSI_OPCODE_POW:
 687       src0 = emit_fetch(bld, inst, 0);
 688       src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 689       src1 = emit_fetch(bld, inst, 1);
 690       src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
 691       dst0 = lp_build_pow(&bld->base, src0, src1);
 692       break;
 693
 694    case TGSI_OPCODE_XPD:
 695       return FALSE;
 696
 697    case TGSI_OPCODE_ABS:
 698       src0 = emit_fetch(bld, inst, 0);
 699       dst0 = lp_build_abs(&bld->base, src0);
 700       break;
 701
 702    case TGSI_OPCODE_RCC:
 703       /* deprecated? */
 704       assert(0);
 705       return FALSE;
 706
 707    case TGSI_OPCODE_DPH:
 708       return FALSE;
 709
 710    case TGSI_OPCODE_COS:
 711       src0 = emit_fetch(bld, inst, 0);
 712       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 713       dst0 = lp_build_cos(&bld->base, tmp0);
 714       break;
 715
 716    case TGSI_OPCODE_DDX:
 717       return FALSE;
 718
 719    case TGSI_OPCODE_DDY:
 720       return FALSE;
 721
 722    case TGSI_OPCODE_KILP:
 723       /* predicated kill */
 724       return FALSE;
 725
 726    case TGSI_OPCODE_KIL:
 727       /* conditional kill */
 728       return FALSE;
 729
 730    case TGSI_OPCODE_PK2H:
 731       return FALSE;
 732       break;
 733
 734    case TGSI_OPCODE_PK2US:
 735       return FALSE;
 736       break;
 737
 738    case TGSI_OPCODE_PK4B:
 739       return FALSE;
 740       break;
 741
 742    case TGSI_OPCODE_PK4UB:
 743       return FALSE;
 744
 745    case TGSI_OPCODE_RFL:
 746       return FALSE;
 747
 748    case TGSI_OPCODE_SEQ:
 749       src0 = emit_fetch(bld, inst, 0);
 750       src1 = emit_fetch(bld, inst, 1);
 751       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_EQUAL, src0, src1);
 752       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 753       break;
 754
 755    case TGSI_OPCODE_SFL:
 756       dst0 = bld->base.zero;
 757       break;
 758
 759    case TGSI_OPCODE_SGT:
 760       src0 = emit_fetch(bld, inst, 0);
 761       src1 = emit_fetch(bld, inst, 1);
 762       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src0, src1);
 763       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 764       break;
 765
 766    case TGSI_OPCODE_SIN:
 767       src0 = emit_fetch(bld, inst, 0);
 768       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 769       dst0 = lp_build_sin(&bld->base, tmp0);
 770       break;
 771
 772    case TGSI_OPCODE_SLE:
 773       src0 = emit_fetch(bld, inst, 0);
 774       src1 = emit_fetch(bld, inst, 1);
 775       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LEQUAL, src0, src1);
 776       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 777       break;
 778
 779    case TGSI_OPCODE_SNE:
 780       src0 = emit_fetch(bld, inst, 0);
 781       src1 = emit_fetch(bld, inst, 1);
 782       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL, src0, src1);
 783       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 784       break;
 785
 786    case TGSI_OPCODE_STR:
 787       dst0 = bld->base.one;
 788       break;
 789
 790    case TGSI_OPCODE_TEX:
 791       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 792       break;
 793
 794    case TGSI_OPCODE_TXD:
 795       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 796       break;
 797
 798    case TGSI_OPCODE_UP2H:
 799       /* deprecated */
 800       assert (0);
 801       return FALSE;
 802       break;
 803
 804    case TGSI_OPCODE_UP2US:
 805       /* deprecated */
 806       assert(0);
 807       return FALSE;
 808       break;
 809
 810    case TGSI_OPCODE_UP4B:
 811       /* deprecated */
 812       assert(0);
 813       return FALSE;
 814       break;
 815
 816    case TGSI_OPCODE_UP4UB:
 817       /* deprecated */
 818       assert(0);
 819       return FALSE;
 820       break;
 821
 822    case TGSI_OPCODE_X2D:
 823       /* deprecated? */
 824       assert(0);
 825       return FALSE;
 826       break;
 827
 828    case TGSI_OPCODE_ARA:
 829       /* deprecated */
 830       assert(0);
 831       return FALSE;
 832       break;
 833
 834    case TGSI_OPCODE_ARR:
 835       src0 = emit_fetch(bld, inst, 0);
 836       dst0 = lp_build_round(&bld->base, src0);
 837       break;
 838
 839    case TGSI_OPCODE_BRA:
 840       /* deprecated */
 841       assert(0);
 842       return FALSE;
 843       break;
 844
 845    case TGSI_OPCODE_CAL:
 846       return FALSE;
 847
 848    case TGSI_OPCODE_RET:
 849       return FALSE;
 850
 851    case TGSI_OPCODE_END:
 852       *pc = -1;
 853       break;
 854
 855    case TGSI_OPCODE_SSG:
 856    /* TGSI_OPCODE_SGN */
 857       tmp0 = emit_fetch(bld, inst, 0);
 858       dst0 = lp_build_sgn(&bld->base, tmp0);
 859       break;
 860
 861    case TGSI_OPCODE_CMP:
 862       src0 = emit_fetch(bld, inst, 0);
 863       src1 = emit_fetch(bld, inst, 1);
 864       src2 = emit_fetch(bld, inst, 2);
 865       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, bld->base.zero);
 866       dst0 = lp_build_select(&bld->base, tmp0, src1, src2);
 867       break;
 868
 869    case TGSI_OPCODE_SCS:
 870       return FALSE;
 871
 872    case TGSI_OPCODE_TXB:
 873       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 874       break;
 875
 876    case TGSI_OPCODE_NRM:
 877       /* fall-through */
 878    case TGSI_OPCODE_NRM4:
 879       return FALSE;
 880
 881    case TGSI_OPCODE_DIV:
 882       /* deprecated */
 883       assert(0);
 884       return FALSE;
 885       break;
 886
 887    case TGSI_OPCODE_DP2:
 888       return FALSE;
 889
 890    case TGSI_OPCODE_TXL:
 891       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 892       break;
 893
 894    case TGSI_OPCODE_TXP:
 895       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 896       break;
 897
 898    case TGSI_OPCODE_BRK:
 899       return FALSE;
 900
 901    case TGSI_OPCODE_IF:
 902       return FALSE;
 903
 904    case TGSI_OPCODE_BGNLOOP:
 905       return FALSE;
 906
 907    case TGSI_OPCODE_BGNSUB:
 908       return FALSE;
 909
 910    case TGSI_OPCODE_ELSE:
 911       return FALSE;
 912
 913    case TGSI_OPCODE_ENDIF:
 914       return FALSE;
 915
 916    case TGSI_OPCODE_ENDLOOP:
 917       return FALSE;
 918
 919    case TGSI_OPCODE_ENDSUB:
 920       return FALSE;
 921
 922    case TGSI_OPCODE_PUSHA:
 923       /* deprecated? */
 924       assert(0);
 925       return FALSE;
 926       break;
 927
 928    case TGSI_OPCODE_POPA:
 929       /* deprecated? */
 930       assert(0);
 931       return FALSE;
 932       break;
 933
 934    case TGSI_OPCODE_CEIL:
 935       src0 = emit_fetch(bld, inst, 0);
 936       dst0 = lp_build_ceil(&bld->base, src0);
 937       break;
 938
 939    case TGSI_OPCODE_I2F:
 940       /* deprecated? */
 941       assert(0);
 942       return FALSE;
 943       break;
 944
 945    case TGSI_OPCODE_NOT:
 946       /* deprecated? */
 947       assert(0);
 948       return FALSE;
 949       break;
 950
 951    case TGSI_OPCODE_TRUNC:
 952       src0 = emit_fetch(bld, inst, 0);
 953       dst0 = lp_build_trunc(&bld->base, src0);
 954       break;
 955
 956    case TGSI_OPCODE_SHL:
 957       /* deprecated? */
 958       assert(0);
 959       return FALSE;
 960       break;
 961
 962    case TGSI_OPCODE_ISHR:
 963       /* deprecated? */
 964       assert(0);
 965       return FALSE;
 966       break;
 967
 968    case TGSI_OPCODE_AND:
 969       /* deprecated? */
 970       assert(0);
 971       return FALSE;
 972       break;
 973
 974    case TGSI_OPCODE_OR:
 975       /* deprecated? */
 976       assert(0);
 977       return FALSE;
 978       break;
 979
 980    case TGSI_OPCODE_MOD:
 981       /* deprecated? */
 982       assert(0);
 983       return FALSE;
 984       break;
 985
 986    case TGSI_OPCODE_XOR:
 987       /* deprecated? */
 988       assert(0);
 989       return FALSE;
 990       break;
 991
 992    case TGSI_OPCODE_SAD:
 993       /* deprecated? */
 994       assert(0);
 995       return FALSE;
 996       break;
 997
 998    case TGSI_OPCODE_TXF:
 999       /* deprecated? */
1000       assert(0);
1001       return FALSE;
1002       break;
1003
1004    case TGSI_OPCODE_TXQ:
1005       /* deprecated? */
1006       assert(0);
1007       return FALSE;
1008       break;
1009
1010    case TGSI_OPCODE_CONT:
1011       return FALSE;
1012
1013    case TGSI_OPCODE_EMIT:
1014       return FALSE;
1015       break;
1016
1017    case TGSI_OPCODE_ENDPRIM:
1018       return FALSE;
1019       break;
1020
1021    case TGSI_OPCODE_NOP:
1022       break;
1023
1024    default:
1025       return FALSE;
1026    }
1027
1028    if (info->num_dst) {
1029       emit_store(bld, inst, 0, dst0);
1030    }
1031
1032    return TRUE;
1033 }
1034
1035
1036 void
1037 lp_build_tgsi_aos(struct gallivm_state *gallivm,
1038                   const struct tgsi_token *tokens,
1039                   struct lp_type type,
1040                   const unsigned char swizzles[4],
1041                   LLVMValueRef consts_ptr,
1042                   const LLVMValueRef *inputs,
1043                   LLVMValueRef *outputs,
1044                   struct lp_build_sampler_aos *sampler,
1045                   const struct tgsi_shader_info *info)
1046 {
1047    struct lp_build_tgsi_aos_context bld;
1048    struct tgsi_parse_context parse;
1049    uint num_immediates = 0;
1050    uint num_instructions = 0;
1051    unsigned chan;
1052    int pc = 0;
1053
1054    /* Setup build context */
1055    memset(&bld, 0, sizeof bld);
1056    lp_build_context_init(&bld.base, gallivm, type);
1057    lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
1058
1059    for (chan = 0; chan < 4; ++chan) {
1060       bld.swizzles[chan] = swizzles[chan];
1061       bld.inv_swizzles[swizzles[chan]] = chan;
1062    }
1063
1064    bld.inputs = inputs;
1065    bld.outputs = outputs;
1066    bld.consts_ptr = consts_ptr;
1067    bld.sampler = sampler;
1068    bld.indirect_files = info->indirect_files;
1069    bld.instructions = (struct tgsi_full_instruction *)
1070                       MALLOC(LP_MAX_INSTRUCTIONS * sizeof(struct tgsi_full_instruction));
1071    bld.max_instructions = LP_MAX_INSTRUCTIONS;
1072
1073    if (!bld.instructions) {
1074       return;
1075    }
1076
1077    tgsi_parse_init(&parse, tokens);
1078
1079    while (!tgsi_parse_end_of_tokens(&parse)) {
1080       tgsi_parse_token(&parse);
1081
1082       switch(parse.FullToken.Token.Type) {
1083       case TGSI_TOKEN_TYPE_DECLARATION:
1084          /* Inputs already interpolated */
1085          emit_declaration(&bld, &parse.FullToken.FullDeclaration);
1086          break;
1087
1088       case TGSI_TOKEN_TYPE_INSTRUCTION:
1089          {
1090             /* save expanded instruction */
1091             if (num_instructions == bld.max_instructions) {
1092                struct tgsi_full_instruction *instructions;
1093                instructions = REALLOC(bld.instructions,
1094                                       bld.max_instructions
1095                                       * sizeof(struct tgsi_full_instruction),
1096                                       (bld.max_instructions + LP_MAX_INSTRUCTIONS)
1097                                       * sizeof(struct tgsi_full_instruction));
1098                if (!instructions) {
1099                   break;
1100                }
1101                bld.instructions = instructions;
1102                bld.max_instructions += LP_MAX_INSTRUCTIONS;
1103             }
1104
1105             memcpy(bld.instructions + num_instructions,
1106                    &parse.FullToken.FullInstruction,
1107                    sizeof(bld.instructions[0]));
1108
1109             num_instructions++;
1110          }
1111
1112          break;
1113
1114       case TGSI_TOKEN_TYPE_IMMEDIATE:
1115          /* simply copy the immediate values into the next immediates[] slot */
1116          {
1117             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1118             float imm[4];
1119             assert(size <= 4);
1120             assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
1121             for (chan = 0; chan < 4; ++chan) {
1122                imm[chan] = 0.0f;
1123             }
1124             for (chan = 0; chan < size; ++chan) {
1125                unsigned swizzle = bld.swizzles[chan];
1126                imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
1127             }
1128             bld.immediates[num_immediates] =
1129                      lp_build_const_aos(gallivm, type,
1130                                         imm[0], imm[1], imm[2], imm[3],
1131                                         NULL);
1132             num_immediates++;
1133          }
1134          break;
1135
1136       case TGSI_TOKEN_TYPE_PROPERTY:
1137          break;
1138
1139       default:
1140          assert(0);
1141       }
1142    }
1143
1144    while (pc != -1) {
1145       struct tgsi_full_instruction *instr = bld.instructions + pc;
1146       const struct tgsi_opcode_info *opcode_info =
1147          tgsi_get_opcode_info(instr->Instruction.Opcode);
1148       if (!emit_instruction(&bld, instr, opcode_info, &pc))
1149          _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1150                        opcode_info->mnemonic);
1151    }
1152
1153    if (0) {
1154       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1155       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1156       debug_printf("11111111111111111111111111111 \n");
1157       tgsi_dump(tokens, 0);
1158       lp_debug_dump_value(function);
1159       debug_printf("2222222222222222222222222222 \n");
1160    }
1161    tgsi_parse_free(&parse);
1162
1163    if (0) {
1164       LLVMModuleRef module = LLVMGetGlobalParent(
1165          LLVMGetBasicBlockParent(LLVMGetInsertBlock(bld.base.builder)));
1166       LLVMDumpModule(module);
1167    }
1168
1169    FREE(bld.instructions);
1170 }
1171