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_gather.h"
  54 #include "lp_bld_logic.h"
  55 #include "lp_bld_swizzle.h"
  56 #include "lp_bld_flow.h"
  57 #include "lp_bld_quad.h"
  58 #include "lp_bld_tgsi.h"
  59 #include "lp_bld_limits.h"
  60 #include "lp_bld_debug.h"
  61
  62
  63 #define LP_MAX_INSTRUCTIONS 256
  64
  65
  66 struct lp_build_tgsi_aos_context
  67 {
  68    struct lp_build_context base;
  69
  70    /* Builder for integer masks and indices */
  71    struct lp_build_context int_bld;
  72
  73    /*
  74     * AoS swizzle used:
  75     * - swizzles[0] = red index
  76     * - swizzles[1] = green index
  77     * - swizzles[2] = blue index
  78     * - swizzles[3] = alpha index
  79     */
  80    unsigned char swizzles[4];
  81    unsigned char inv_swizzles[4];
  82
  83    LLVMValueRef consts_ptr;
  84    const LLVMValueRef *inputs;
  85    LLVMValueRef *outputs;
  86
  87    struct lp_build_sampler_aos *sampler;
  88
  89    LLVMValueRef immediates[LP_MAX_TGSI_IMMEDIATES];
  90    LLVMValueRef temps[LP_MAX_TGSI_TEMPS];
  91    LLVMValueRef addr[LP_MAX_TGSI_ADDRS];
  92    LLVMValueRef preds[LP_MAX_TGSI_PREDS];
  93
  94    /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
  95     * set in the indirect_files field.
  96     * The temps[] array above is unused then.
  97     */
  98    LLVMValueRef temps_array;
  99
 100    /** bitmask indicating which register files are accessed indirectly */
 101    unsigned indirect_files;
 102
 103    struct tgsi_full_instruction *instructions;
 104    uint max_instructions;
 105 };
 106
 107
 108 /**
 109  * Wrapper around lp_build_swizzle_aos which translates swizzles to another
 110  * ordering.
 111  */
 112 static LLVMValueRef
 113 swizzle_aos(struct lp_build_tgsi_aos_context *bld,
 114             LLVMValueRef a,
 115             unsigned swizzle_x,
 116             unsigned swizzle_y,
 117             unsigned swizzle_z,
 118             unsigned swizzle_w)
 119 {
 120    unsigned char swizzles[4];
 121
 122    assert(swizzle_x < 4);
 123    assert(swizzle_y < 4);
 124    assert(swizzle_z < 4);
 125    assert(swizzle_w < 4);
 126
 127    swizzles[bld->inv_swizzles[0]] = bld->swizzles[swizzle_x];
 128    swizzles[bld->inv_swizzles[1]] = bld->swizzles[swizzle_y];
 129    swizzles[bld->inv_swizzles[2]] = bld->swizzles[swizzle_z];
 130    swizzles[bld->inv_swizzles[3]] = bld->swizzles[swizzle_w];
 131
 132    return lp_build_swizzle_aos(&bld->base, a, swizzles);
 133 }
 134
 135
 136 static LLVMValueRef
 137 swizzle_scalar_aos(struct lp_build_tgsi_aos_context *bld,
 138                    LLVMValueRef a,
 139                    unsigned chan)
 140 {
 141    chan = bld->swizzles[chan];
 142    return lp_build_swizzle_scalar_aos(&bld->base, a, chan);
 143 }
 144
 145
 146 /**
 147  * Register fetch.
 148  */
 149 static LLVMValueRef
 150 emit_fetch(
 151    struct lp_build_tgsi_aos_context *bld,
 152    const struct tgsi_full_instruction *inst,
 153    unsigned src_op)
 154 {
 155    struct lp_type type = bld->base.type;
 156    const struct tgsi_full_src_register *reg = &inst->Src[src_op];
 157    LLVMValueRef res;
 158    unsigned chan;
 159
 160    assert(!reg->Register.Indirect);
 161
 162    /*
 163     * Fetch the from the register file.
 164     */
 165
 166    switch (reg->Register.File) {
 167    case TGSI_FILE_CONSTANT:
 168       /*
 169        * Get the constants components
 170        */
 171
 172       res = bld->base.undef;
 173       for (chan = 0; chan < 4; ++chan) {
 174          LLVMValueRef index;
 175          LLVMValueRef scalar_ptr;
 176          LLVMValueRef scalar;
 177          LLVMValueRef swizzle;
 178
 179          index = LLVMConstInt(LLVMInt32Type(),
 180                               reg->Register.Index*4 + chan,
 181                               0);
 182
 183          scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr,
 184                                    &index, 1, "");
 185
 186          scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
 187
 188          lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
 189
 190          /*
 191           * NOTE: constants array is always assumed to be RGBA
 192           */
 193
 194          swizzle = LLVMConstInt(LLVMInt32Type(), chan, 0);
 195
 196          res = LLVMBuildInsertElement(bld->base.builder, res, scalar, swizzle, "");
 197       }
 198
 199       /*
 200        * Broadcast the first quaternion to all others.
 201        *
 202        * XXX: could be factored into a reusable function.
 203        */
 204
 205       if (type.length > 4) {
 206          LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 207          unsigned i;
 208
 209          for (chan = 0; chan < 4; ++chan) {
 210             shuffles[chan] = LLVMConstInt(LLVMInt32Type(), chan, 0);
 211          }
 212
 213          for (i = 4; i < type.length; ++i) {
 214             shuffles[i] = shuffles[i % 4];
 215          }
 216
 217          res = LLVMBuildShuffleVector(bld->base.builder,
 218                                       res, bld->base.undef,
 219                                       LLVMConstVector(shuffles, type.length),
 220                                       "");
 221       }
 222       break;
 223
 224    case TGSI_FILE_IMMEDIATE:
 225       res = bld->immediates[reg->Register.Index];
 226       assert(res);
 227       break;
 228
 229    case TGSI_FILE_INPUT:
 230       res = bld->inputs[reg->Register.Index];
 231       assert(res);
 232       break;
 233
 234    case TGSI_FILE_TEMPORARY:
 235       {
 236          LLVMValueRef temp_ptr;
 237          temp_ptr = bld->temps[reg->Register.Index];
 238          res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
 239          if (!res)
 240             return bld->base.undef;
 241       }
 242       break;
 243
 244    default:
 245       assert(0 && "invalid src register in emit_fetch()");
 246       return bld->base.undef;
 247    }
 248
 249    /*
 250     * Apply sign modifier.
 251     */
 252
 253    if (reg->Register.Absolute) {
 254       res = lp_build_abs(&bld->base, res);
 255    }
 256
 257    if(reg->Register.Negate) {
 258       res = lp_build_negate(&bld->base, res);
 259    }
 260
 261    /*
 262     * Swizzle the argument
 263     */
 264
 265    res = swizzle_aos(bld, res,
 266                      reg->Register.SwizzleX,
 267                      reg->Register.SwizzleY,
 268                      reg->Register.SwizzleZ,
 269                      reg->Register.SwizzleW);
 270
 271    return res;
 272 }
 273
 274
 275 /**
 276  * Register store.
 277  */
 278 static void
 279 emit_store(
 280    struct lp_build_tgsi_aos_context *bld,
 281    const struct tgsi_full_instruction *inst,
 282    unsigned index,
 283    LLVMValueRef value)
 284 {
 285    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 286    LLVMValueRef mask = NULL;
 287    LLVMValueRef ptr;
 288
 289    /*
 290     * Saturate the value
 291     */
 292
 293    switch (inst->Instruction.Saturate) {
 294    case TGSI_SAT_NONE:
 295       break;
 296
 297    case TGSI_SAT_ZERO_ONE:
 298       value = lp_build_max(&bld->base, value, bld->base.zero);
 299       value = lp_build_min(&bld->base, value, bld->base.one);
 300       break;
 301
 302    case TGSI_SAT_MINUS_PLUS_ONE:
 303       value = lp_build_max(&bld->base, value, lp_build_const_vec(bld->base.type, -1.0));
 304       value = lp_build_min(&bld->base, value, bld->base.one);
 305       break;
 306
 307    default:
 308       assert(0);
 309    }
 310
 311    /*
 312     * Translate the register file
 313     */
 314
 315    assert(!reg->Register.Indirect);
 316
 317    switch (reg->Register.File) {
 318    case TGSI_FILE_OUTPUT:
 319       ptr = bld->outputs[reg->Register.Index];
 320       break;
 321
 322    case TGSI_FILE_TEMPORARY:
 323       ptr = bld->temps[reg->Register.Index];
 324       break;
 325
 326    case TGSI_FILE_ADDRESS:
 327       ptr = bld->addr[reg->Indirect.Index];
 328       break;
 329
 330    case TGSI_FILE_PREDICATE:
 331       ptr = bld->preds[reg->Register.Index];
 332       break;
 333
 334    default:
 335       assert(0);
 336       return;
 337    }
 338
 339    /*
 340     * Predicate
 341     */
 342
 343    if (inst->Instruction.Predicate) {
 344       LLVMValueRef pred;
 345
 346       assert(inst->Predicate.Index < LP_MAX_TGSI_PREDS);
 347
 348       pred = LLVMBuildLoad(bld->base.builder,
 349                            bld->preds[inst->Predicate.Index], "");
 350
 351       /*
 352        * Convert the value to an integer mask.
 353        */
 354       pred = lp_build_compare(bld->base.builder,
 355                                bld->base.type,
 356                                PIPE_FUNC_NOTEQUAL,
 357                                pred,
 358                                bld->base.zero);
 359
 360       if (inst->Predicate.Negate) {
 361          pred = LLVMBuildNot(bld->base.builder, pred, "");
 362       }
 363
 364       pred = swizzle_aos(bld, pred,
 365                          inst->Predicate.SwizzleX,
 366                          inst->Predicate.SwizzleY,
 367                          inst->Predicate.SwizzleZ,
 368                          inst->Predicate.SwizzleW);
 369
 370       if (mask) {
 371          mask = LLVMBuildAnd(bld->base.builder, mask, pred, "");
 372       } else {
 373          mask = pred;
 374       }
 375    }
 376
 377    /*
 378     * Writemask
 379     */
 380
 381    if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
 382       LLVMValueRef writemask;
 383
 384       writemask = lp_build_const_mask_aos(bld->base.type, reg->Register.WriteMask);
 385
 386       if (mask) {
 387          mask = LLVMBuildAnd(bld->base.builder, mask, writemask, "");
 388       } else {
 389          mask = writemask;
 390       }
 391    }
 392
 393    if (mask) {
 394       LLVMValueRef orig_value;
 395
 396       orig_value = LLVMBuildLoad(bld->base.builder, ptr, "");
 397       value = lp_build_select(&bld->base,
 398                               mask, value, orig_value);
 399    }
 400
 401    LLVMBuildStore(bld->base.builder, value, ptr);
 402 }
 403
 404
 405 /**
 406  * High-level instruction translators.
 407  */
 408
 409 static LLVMValueRef
 410 emit_tex(struct lp_build_tgsi_aos_context *bld,
 411          const struct tgsi_full_instruction *inst,
 412          enum lp_build_tex_modifier modifier)
 413 {
 414    unsigned target;
 415    unsigned unit;
 416    LLVMValueRef coords;
 417    LLVMValueRef ddx;
 418    LLVMValueRef ddy;
 419
 420    if (!bld->sampler) {
 421       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 422       return bld->base.undef;
 423    }
 424
 425    target = inst->Texture.Texture;
 426
 427    coords = emit_fetch( bld, inst, 0 );
 428
 429    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 430       ddx = emit_fetch( bld, inst, 1 );
 431       ddy = emit_fetch( bld, inst, 2 );
 432       unit = inst->Src[3].Register.Index;
 433    }  else {
 434 #if 0
 435       ddx = lp_build_ddx( &bld->base, coords );
 436       ddy = lp_build_ddy( &bld->base, coords );
 437 #else
 438       /* TODO */
 439       ddx = bld->base.one;
 440       ddy = bld->base.one;
 441 #endif
 442       unit = inst->Src[1].Register.Index;
 443    }
 444
 445    return bld->sampler->emit_fetch_texel(bld->sampler,
 446                                          &bld->base,
 447                                          target, unit,
 448                                          coords, ddx, ddy,
 449                                          modifier);
 450 }
 451
 452
 453 static void
 454 emit_declaration(
 455    struct lp_build_tgsi_aos_context *bld,
 456    const struct tgsi_full_declaration *decl)
 457 {
 458    LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
 459
 460    unsigned first = decl->Range.First;
 461    unsigned last = decl->Range.Last;
 462    unsigned idx;
 463
 464    for (idx = first; idx <= last; ++idx) {
 465       switch (decl->Declaration.File) {
 466       case TGSI_FILE_TEMPORARY:
 467          assert(idx < LP_MAX_TGSI_TEMPS);
 468          if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 469             LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
 470                                                    last + 1, 0);
 471             bld->temps_array = lp_build_array_alloca(bld->base.builder,
 472                                                      vec_type, array_size, "");
 473          } else {
 474             bld->temps[idx] = lp_build_alloca(bld->base.builder,
 475                                               vec_type, "");
 476          }
 477          break;
 478
 479       case TGSI_FILE_OUTPUT:
 480          bld->outputs[idx] = lp_build_alloca(bld->base.builder,
 481                                              vec_type, "");
 482          break;
 483
 484       case TGSI_FILE_ADDRESS:
 485          assert(idx < LP_MAX_TGSI_ADDRS);
 486          bld->addr[idx] = lp_build_alloca(bld->base.builder,
 487                                           vec_type, "");
 488          break;
 489
 490       case TGSI_FILE_PREDICATE:
 491          assert(idx < LP_MAX_TGSI_PREDS);
 492          bld->preds[idx] = lp_build_alloca(bld->base.builder,
 493                                            vec_type, "");
 494          break;
 495
 496       default:
 497          /* don't need to declare other vars */
 498          break;
 499       }
 500    }
 501 }
 502
 503
 504 /**
 505  * Emit LLVM for one TGSI instruction.
 506  * \param return TRUE for success, FALSE otherwise
 507  */
 508 static boolean
 509 emit_instruction(
 510    struct lp_build_tgsi_aos_context *bld,
 511    const struct tgsi_full_instruction *inst,
 512    const struct tgsi_opcode_info *info,
 513    int *pc)
 514 {
 515    LLVMValueRef src0, src1, src2;
 516    LLVMValueRef tmp0, tmp1;
 517    LLVMValueRef dst0;
 518
 519    /*
 520     * Stores and write masks are handled in a general fashion after the long
 521     * instruction opcode switch statement.
 522     *
 523     * Although not stricitly necessary, we avoid generating instructions for
 524     * channels which won't be stored, in cases where's that easy. For some
 525     * complex instructions, like texture sampling, it is more convenient to
 526     * assume a full writemask and then let LLVM optimization passes eliminate
 527     * redundant code.
 528     */
 529
 530    (*pc)++;
 531
 532    assert(info->num_dst <= 1);
 533    if (info->num_dst) {
 534       dst0 = bld->base.undef;
 535    }
 536
 537    switch (inst->Instruction.Opcode) {
 538    case TGSI_OPCODE_ARL:
 539       src0 = emit_fetch(bld, inst, 0);
 540       dst0 = lp_build_floor(&bld->base, src0);
 541       break;
 542
 543    case TGSI_OPCODE_MOV:
 544       dst0 = emit_fetch(bld, inst, 0);
 545       break;
 546
 547    case TGSI_OPCODE_LIT:
 548       return FALSE;
 549
 550    case TGSI_OPCODE_RCP:
 551    /* TGSI_OPCODE_RECIP */
 552       src0 = emit_fetch(bld, inst, 0);
 553       dst0 = lp_build_rcp(&bld->base, src0);
 554       break;
 555
 556    case TGSI_OPCODE_RSQ:
 557    /* TGSI_OPCODE_RECIPSQRT */
 558       src0 = emit_fetch(bld, inst, 0);
 559       tmp0 = lp_build_abs(&bld->base, src0);
 560       dst0 = lp_build_rsqrt(&bld->base, tmp0);
 561       break;
 562
 563    case TGSI_OPCODE_EXP:
 564       return FALSE;
 565
 566    case TGSI_OPCODE_LOG:
 567       return FALSE;
 568
 569    case TGSI_OPCODE_MUL:
 570       src0 = emit_fetch(bld, inst, 0);
 571       src1 = emit_fetch(bld, inst, 1);
 572       dst0 = lp_build_mul(&bld->base, src0, src1);
 573       break;
 574
 575    case TGSI_OPCODE_ADD:
 576       src0 = emit_fetch(bld, inst, 0);
 577       src1 = emit_fetch(bld, inst, 1);
 578       dst0 = lp_build_add(&bld->base, src0, src1);
 579       break;
 580
 581    case TGSI_OPCODE_DP3:
 582    /* TGSI_OPCODE_DOT3 */
 583       return FALSE;
 584
 585    case TGSI_OPCODE_DP4:
 586    /* TGSI_OPCODE_DOT4 */
 587       return FALSE;
 588
 589    case TGSI_OPCODE_DST:
 590       return FALSE;
 591
 592    case TGSI_OPCODE_MIN:
 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_MAX:
 599       src0 = emit_fetch(bld, inst, 0);
 600       src1 = emit_fetch(bld, inst, 1);
 601       dst0 = lp_build_max(&bld->base, src0, src1);
 602       break;
 603
 604    case TGSI_OPCODE_SLT:
 605    /* TGSI_OPCODE_SETLT */
 606       src0 = emit_fetch(bld, inst, 0);
 607       src1 = emit_fetch(bld, inst, 1);
 608       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, src1);
 609       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 610       break;
 611
 612    case TGSI_OPCODE_SGE:
 613    /* TGSI_OPCODE_SETGE */
 614       src0 = emit_fetch(bld, inst, 0);
 615       src1 = emit_fetch(bld, inst, 1);
 616       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, src0, src1);
 617       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 618       break;
 619
 620    case TGSI_OPCODE_MAD:
 621    /* TGSI_OPCODE_MADD */
 622       src0 = emit_fetch(bld, inst, 0);
 623       src1 = emit_fetch(bld, inst, 1);
 624       src2 = emit_fetch(bld, inst, 2);
 625       tmp0 = lp_build_mul(&bld->base, src0, src1);
 626       dst0 = lp_build_add(&bld->base, tmp0, src2);
 627       break;
 628
 629    case TGSI_OPCODE_SUB:
 630       src0 = emit_fetch(bld, inst, 0);
 631       src1 = emit_fetch(bld, inst, 1);
 632       dst0 = lp_build_sub(&bld->base, src0, src1);
 633       break;
 634
 635    case TGSI_OPCODE_LRP:
 636       src0 = emit_fetch(bld, inst, 0);
 637       src1 = emit_fetch(bld, inst, 1);
 638       src2 = emit_fetch(bld, inst, 2);
 639       tmp0 = lp_build_sub(&bld->base, src1, src2);
 640       tmp0 = lp_build_mul(&bld->base, src0, tmp0);
 641       dst0 = lp_build_add(&bld->base, tmp0, src2);
 642       break;
 643
 644    case TGSI_OPCODE_CND:
 645       src0 = emit_fetch(bld, inst, 0);
 646       src1 = emit_fetch(bld, inst, 1);
 647       src2 = emit_fetch(bld, inst, 2);
 648       tmp1 = lp_build_const_vec(bld->base.type, 0.5);
 649       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src2, tmp1);
 650       dst0 = lp_build_select(&bld->base, tmp0, src0, src1);
 651       break;
 652
 653    case TGSI_OPCODE_DP2A:
 654       return FALSE;
 655
 656    case TGSI_OPCODE_FRC:
 657       src0 = emit_fetch(bld, inst, 0);
 658       tmp0 = lp_build_floor(&bld->base, src0);
 659       dst0 = lp_build_sub(&bld->base, src0, tmp0);
 660       break;
 661
 662    case TGSI_OPCODE_CLAMP:
 663       src0 = emit_fetch(bld, inst, 0);
 664       src1 = emit_fetch(bld, inst, 1);
 665       src2 = emit_fetch(bld, inst, 2);
 666       tmp0 = lp_build_max(&bld->base, src0, src1);
 667       dst0 = lp_build_min(&bld->base, tmp0, src2);
 668       break;
 669
 670    case TGSI_OPCODE_FLR:
 671       src0 = emit_fetch(bld, inst, 0);
 672       dst0 = lp_build_floor(&bld->base, src0);
 673       break;
 674
 675    case TGSI_OPCODE_ROUND:
 676       src0 = emit_fetch(bld, inst, 0);
 677       dst0 = lp_build_round(&bld->base, src0);
 678       break;
 679
 680    case TGSI_OPCODE_EX2:
 681       src0 = emit_fetch(bld, inst, 0);
 682       tmp0 = lp_build_swizzle_scalar_aos(&bld->base, src0, TGSI_SWIZZLE_X);
 683       dst0 = lp_build_exp2(&bld->base, tmp0);
 684       break;
 685
 686    case TGSI_OPCODE_LG2:
 687       src0 = emit_fetch(bld, inst, 0);
 688       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 689       dst0 = lp_build_log2(&bld->base, tmp0);
 690       break;
 691
 692    case TGSI_OPCODE_POW:
 693       src0 = emit_fetch(bld, inst, 0);
 694       src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 695       src1 = emit_fetch(bld, inst, 1);
 696       src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
 697       dst0 = lp_build_pow(&bld->base, src0, src1);
 698       break;
 699
 700    case TGSI_OPCODE_XPD:
 701       return FALSE;
 702
 703    case TGSI_OPCODE_ABS:
 704       src0 = emit_fetch(bld, inst, 0);
 705       dst0 = lp_build_abs(&bld->base, src0);
 706       break;
 707
 708    case TGSI_OPCODE_RCC:
 709       /* deprecated? */
 710       assert(0);
 711       return FALSE;
 712
 713    case TGSI_OPCODE_DPH:
 714       return FALSE;
 715
 716    case TGSI_OPCODE_COS:
 717       src0 = emit_fetch(bld, inst, 0);
 718       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 719       dst0 = lp_build_cos(&bld->base, tmp0);
 720       break;
 721
 722    case TGSI_OPCODE_DDX:
 723       return FALSE;
 724
 725    case TGSI_OPCODE_DDY:
 726       return FALSE;
 727
 728    case TGSI_OPCODE_KILP:
 729       /* predicated kill */
 730       return FALSE;
 731
 732    case TGSI_OPCODE_KIL:
 733       /* conditional kill */
 734       return FALSE;
 735
 736    case TGSI_OPCODE_PK2H:
 737       return FALSE;
 738       break;
 739
 740    case TGSI_OPCODE_PK2US:
 741       return FALSE;
 742       break;
 743
 744    case TGSI_OPCODE_PK4B:
 745       return FALSE;
 746       break;
 747
 748    case TGSI_OPCODE_PK4UB:
 749       return FALSE;
 750
 751    case TGSI_OPCODE_RFL:
 752       return FALSE;
 753
 754    case TGSI_OPCODE_SEQ:
 755       src0 = emit_fetch(bld, inst, 0);
 756       src1 = emit_fetch(bld, inst, 1);
 757       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_EQUAL, src0, src1);
 758       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 759       break;
 760
 761    case TGSI_OPCODE_SFL:
 762       dst0 = bld->base.zero;
 763       break;
 764
 765    case TGSI_OPCODE_SGT:
 766       src0 = emit_fetch(bld, inst, 0);
 767       src1 = emit_fetch(bld, inst, 1);
 768       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src0, src1);
 769       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 770       break;
 771
 772    case TGSI_OPCODE_SIN:
 773       src0 = emit_fetch(bld, inst, 0);
 774       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 775       dst0 = lp_build_sin(&bld->base, tmp0);
 776       break;
 777
 778    case TGSI_OPCODE_SLE:
 779       src0 = emit_fetch(bld, inst, 0);
 780       src1 = emit_fetch(bld, inst, 1);
 781       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LEQUAL, src0, src1);
 782       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 783       break;
 784
 785    case TGSI_OPCODE_SNE:
 786       src0 = emit_fetch(bld, inst, 0);
 787       src1 = emit_fetch(bld, inst, 1);
 788       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL, src0, src1);
 789       dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
 790       break;
 791
 792    case TGSI_OPCODE_STR:
 793       dst0 = bld->base.one;
 794       break;
 795
 796    case TGSI_OPCODE_TEX:
 797       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 798       break;
 799
 800    case TGSI_OPCODE_TXD:
 801       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 802       break;
 803
 804    case TGSI_OPCODE_UP2H:
 805       /* deprecated */
 806       assert (0);
 807       return FALSE;
 808       break;
 809
 810    case TGSI_OPCODE_UP2US:
 811       /* deprecated */
 812       assert(0);
 813       return FALSE;
 814       break;
 815
 816    case TGSI_OPCODE_UP4B:
 817       /* deprecated */
 818       assert(0);
 819       return FALSE;
 820       break;
 821
 822    case TGSI_OPCODE_UP4UB:
 823       /* deprecated */
 824       assert(0);
 825       return FALSE;
 826       break;
 827
 828    case TGSI_OPCODE_X2D:
 829       /* deprecated? */
 830       assert(0);
 831       return FALSE;
 832       break;
 833
 834    case TGSI_OPCODE_ARA:
 835       /* deprecated */
 836       assert(0);
 837       return FALSE;
 838       break;
 839
 840    case TGSI_OPCODE_ARR:
 841       src0 = emit_fetch(bld, inst, 0);
 842       dst0 = lp_build_round(&bld->base, src0);
 843       break;
 844
 845    case TGSI_OPCODE_BRA:
 846       /* deprecated */
 847       assert(0);
 848       return FALSE;
 849       break;
 850
 851    case TGSI_OPCODE_CAL:
 852       return FALSE;
 853
 854    case TGSI_OPCODE_RET:
 855       return FALSE;
 856
 857    case TGSI_OPCODE_END:
 858       *pc = -1;
 859       break;
 860
 861    case TGSI_OPCODE_SSG:
 862    /* TGSI_OPCODE_SGN */
 863       tmp0 = emit_fetch(bld, inst, 0);
 864       dst0 = lp_build_sgn(&bld->base, tmp0);
 865       break;
 866
 867    case TGSI_OPCODE_CMP:
 868       src0 = emit_fetch(bld, inst, 0);
 869       src1 = emit_fetch(bld, inst, 1);
 870       src2 = emit_fetch(bld, inst, 2);
 871       tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, bld->base.zero);
 872       dst0 = lp_build_select(&bld->base, tmp0, src1, src2);
 873       break;
 874
 875    case TGSI_OPCODE_SCS:
 876       return FALSE;
 877
 878    case TGSI_OPCODE_TXB:
 879       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 880       break;
 881
 882    case TGSI_OPCODE_NRM:
 883       /* fall-through */
 884    case TGSI_OPCODE_NRM4:
 885       return FALSE;
 886
 887    case TGSI_OPCODE_DIV:
 888       /* deprecated */
 889       assert(0);
 890       return FALSE;
 891       break;
 892
 893    case TGSI_OPCODE_DP2:
 894       return FALSE;
 895
 896    case TGSI_OPCODE_TXL:
 897       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 898       break;
 899
 900    case TGSI_OPCODE_TXP:
 901       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 902       break;
 903
 904    case TGSI_OPCODE_BRK:
 905       return FALSE;
 906
 907    case TGSI_OPCODE_IF:
 908       return FALSE;
 909
 910    case TGSI_OPCODE_BGNLOOP:
 911       return FALSE;
 912
 913    case TGSI_OPCODE_BGNSUB:
 914       return FALSE;
 915
 916    case TGSI_OPCODE_ELSE:
 917       return FALSE;
 918
 919    case TGSI_OPCODE_ENDIF:
 920       return FALSE;
 921
 922    case TGSI_OPCODE_ENDLOOP:
 923       return FALSE;
 924
 925    case TGSI_OPCODE_ENDSUB:
 926       return FALSE;
 927
 928    case TGSI_OPCODE_PUSHA:
 929       /* deprecated? */
 930       assert(0);
 931       return FALSE;
 932       break;
 933
 934    case TGSI_OPCODE_POPA:
 935       /* deprecated? */
 936       assert(0);
 937       return FALSE;
 938       break;
 939
 940    case TGSI_OPCODE_CEIL:
 941       src0 = emit_fetch(bld, inst, 0);
 942       dst0 = lp_build_ceil(&bld->base, src0);
 943       break;
 944
 945    case TGSI_OPCODE_I2F:
 946       /* deprecated? */
 947       assert(0);
 948       return FALSE;
 949       break;
 950
 951    case TGSI_OPCODE_NOT:
 952       /* deprecated? */
 953       assert(0);
 954       return FALSE;
 955       break;
 956
 957    case TGSI_OPCODE_TRUNC:
 958       src0 = emit_fetch(bld, inst, 0);
 959       dst0 = lp_build_trunc(&bld->base, src0);
 960       break;
 961
 962    case TGSI_OPCODE_SHL:
 963       /* deprecated? */
 964       assert(0);
 965       return FALSE;
 966       break;
 967
 968    case TGSI_OPCODE_ISHR:
 969       /* deprecated? */
 970       assert(0);
 971       return FALSE;
 972       break;
 973
 974    case TGSI_OPCODE_AND:
 975       /* deprecated? */
 976       assert(0);
 977       return FALSE;
 978       break;
 979
 980    case TGSI_OPCODE_OR:
 981       /* deprecated? */
 982       assert(0);
 983       return FALSE;
 984       break;
 985
 986    case TGSI_OPCODE_MOD:
 987       /* deprecated? */
 988       assert(0);
 989       return FALSE;
 990       break;
 991
 992    case TGSI_OPCODE_XOR:
 993       /* deprecated? */
 994       assert(0);
 995       return FALSE;
 996       break;
 997
 998    case TGSI_OPCODE_SAD:
 999       /* deprecated? */
1000       assert(0);
1001       return FALSE;
1002       break;
1003
1004    case TGSI_OPCODE_TXF:
1005       /* deprecated? */
1006       assert(0);
1007       return FALSE;
1008       break;
1009
1010    case TGSI_OPCODE_TXQ:
1011       /* deprecated? */
1012       assert(0);
1013       return FALSE;
1014       break;
1015
1016    case TGSI_OPCODE_CONT:
1017       return FALSE;
1018
1019    case TGSI_OPCODE_EMIT:
1020       return FALSE;
1021       break;
1022
1023    case TGSI_OPCODE_ENDPRIM:
1024       return FALSE;
1025       break;
1026
1027    case TGSI_OPCODE_NOP:
1028       break;
1029
1030    default:
1031       return FALSE;
1032    }
1033
1034    if (info->num_dst) {
1035       emit_store(bld, inst, 0, dst0);
1036    }
1037
1038    return TRUE;
1039 }
1040
1041
1042 void
1043 lp_build_tgsi_aos(LLVMBuilderRef builder,
1044                   const struct tgsi_token *tokens,
1045                   struct lp_type type,
1046                   const unsigned char swizzles[4],
1047                   LLVMValueRef consts_ptr,
1048                   const LLVMValueRef *inputs,
1049                   LLVMValueRef *outputs,
1050                   struct lp_build_sampler_aos *sampler,
1051                   const struct tgsi_shader_info *info)
1052 {
1053    struct lp_build_tgsi_aos_context bld;
1054    struct tgsi_parse_context parse;
1055    uint num_immediates = 0;
1056    uint num_instructions = 0;
1057    unsigned chan;
1058    int pc = 0;
1059
1060    /* Setup build context */
1061    memset(&bld, 0, sizeof bld);
1062    lp_build_context_init(&bld.base, builder, type);
1063    lp_build_context_init(&bld.int_bld, builder, lp_int_type(type));
1064
1065    for (chan = 0; chan < 4; ++chan) {
1066       bld.swizzles[chan] = swizzles[chan];
1067       bld.inv_swizzles[swizzles[chan]] = chan;
1068    }
1069
1070    bld.inputs = inputs;
1071    bld.outputs = outputs;
1072    bld.consts_ptr = consts_ptr;
1073    bld.sampler = sampler;
1074    bld.indirect_files = info->indirect_files;
1075    bld.instructions = (struct tgsi_full_instruction *)
1076                       MALLOC(LP_MAX_INSTRUCTIONS * sizeof(struct tgsi_full_instruction));
1077    bld.max_instructions = LP_MAX_INSTRUCTIONS;
1078
1079    if (!bld.instructions) {
1080       return;
1081    }
1082
1083    tgsi_parse_init(&parse, tokens);
1084
1085    while (!tgsi_parse_end_of_tokens(&parse)) {
1086       tgsi_parse_token(&parse);
1087
1088       switch(parse.FullToken.Token.Type) {
1089       case TGSI_TOKEN_TYPE_DECLARATION:
1090          /* Inputs already interpolated */
1091          emit_declaration(&bld, &parse.FullToken.FullDeclaration);
1092          break;
1093
1094       case TGSI_TOKEN_TYPE_INSTRUCTION:
1095          {
1096             /* save expanded instruction */
1097             if (num_instructions == bld.max_instructions) {
1098                struct tgsi_full_instruction *instructions;
1099                instructions = REALLOC(bld.instructions,
1100                                       bld.max_instructions
1101                                       * sizeof(struct tgsi_full_instruction),
1102                                       (bld.max_instructions + LP_MAX_INSTRUCTIONS)
1103                                       * sizeof(struct tgsi_full_instruction));
1104                if (!instructions) {
1105                   break;
1106                }
1107                bld.instructions = instructions;
1108                bld.max_instructions += LP_MAX_INSTRUCTIONS;
1109             }
1110
1111             memcpy(bld.instructions + num_instructions,
1112                    &parse.FullToken.FullInstruction,
1113                    sizeof(bld.instructions[0]));
1114
1115             num_instructions++;
1116          }
1117
1118          break;
1119
1120       case TGSI_TOKEN_TYPE_IMMEDIATE:
1121          /* simply copy the immediate values into the next immediates[] slot */
1122          {
1123             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1124             float imm[4];
1125             assert(size <= 4);
1126             assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
1127             for (chan = 0; chan < 4; ++chan) {
1128                imm[chan] = 0.0f;
1129             }
1130             for (chan = 0; chan < size; ++chan) {
1131                unsigned swizzle = bld.swizzles[chan];
1132                imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
1133             }
1134             bld.immediates[num_immediates] =
1135                      lp_build_const_aos(type,
1136                                         imm[0], imm[1], imm[2], imm[3],
1137                                         NULL);
1138             num_immediates++;
1139          }
1140          break;
1141
1142       case TGSI_TOKEN_TYPE_PROPERTY:
1143          break;
1144
1145       default:
1146          assert(0);
1147       }
1148    }
1149
1150    while (pc != -1) {
1151       struct tgsi_full_instruction *instr = bld.instructions + pc;
1152       const struct tgsi_opcode_info *opcode_info =
1153          tgsi_get_opcode_info(instr->Instruction.Opcode);
1154       if (!emit_instruction(&bld, instr, opcode_info, &pc))
1155          _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1156                        opcode_info->mnemonic);
1157    }
1158
1159    if (0) {
1160       LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
1161       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1162       debug_printf("11111111111111111111111111111 \n");
1163       tgsi_dump(tokens, 0);
1164       lp_debug_dump_value(function);
1165       debug_printf("2222222222222222222222222222 \n");
1166    }
1167    tgsi_parse_free(&parse);
1168
1169    if (0) {
1170       LLVMModuleRef module = LLVMGetGlobalParent(
1171          LLVMGetBasicBlockParent(LLVMGetInsertBlock(bld.base.builder)));
1172       LLVMDumpModule(module);
1173    }
1174
1175    FREE(bld.instructions);
1176 }
1177