src/gallium/auxiliary/gallivm/tgsitollvm.cpp

   1 #include "tgsitollvm.h"
   2
   3 #include "gallivm.h"
   4 #include "gallivm_p.h"
   5
   6 #include "storage.h"
   7 #include "instructions.h"
   8 #include "storagesoa.h"
   9 #include "instructionssoa.h"
  10
  11 #include "pipe/p_shader_tokens.h"
  12
  13 #include "tgsi/util/tgsi_parse.h"
  14 #include "tgsi/exec/tgsi_exec.h"
  15 #include "tgsi/util/tgsi_util.h"
  16 #include "tgsi/util/tgsi_build.h"
  17 #include "tgsi/util/tgsi_dump.h"
  18
  19
  20 #include <llvm/Module.h>
  21 #include <llvm/CallingConv.h>
  22 #include <llvm/Constants.h>
  23 #include <llvm/DerivedTypes.h>
  24 #include <llvm/Instructions.h>
  25 #include <llvm/ModuleProvider.h>
  26 #include <llvm/Pass.h>
  27 #include <llvm/PassManager.h>
  28 #include <llvm/ParameterAttributes.h>
  29 #include <llvm/Support/PatternMatch.h>
  30 #include <llvm/ExecutionEngine/JIT.h>
  31 #include <llvm/ExecutionEngine/Interpreter.h>
  32 #include <llvm/ExecutionEngine/GenericValue.h>
  33 #include <llvm/Support/MemoryBuffer.h>
  34 #include <llvm/LinkAllPasses.h>
  35 #include <llvm/Analysis/Verifier.h>
  36 #include <llvm/Analysis/LoopPass.h>
  37 #include <llvm/Target/TargetData.h>
  38 #include <llvm/Bitcode/ReaderWriter.h>
  39 #include <llvm/Transforms/Utils/Cloning.h>
  40
  41
  42 #include <sstream>
  43 #include <fstream>
  44 #include <iostream>
  45
  46 using namespace llvm;
  47
  48 static inline FunctionType *vertexShaderFunctionType()
  49 {
  50    //Function takes three arguments,
  51    // the calling code has to make sure the types it will
  52    // pass are castable to the following:
  53    // [4 x <4 x float>] inputs,
  54    // [4 x <4 x float>] output,
  55    // [4 x [4 x float]] consts,
  56    // [4 x <4 x float>] temps
  57
  58    std::vector<const Type*> funcArgs;
  59    VectorType *vectorType = VectorType::get(Type::FloatTy, 4);
  60    ArrayType *vectorArray = ArrayType::get(vectorType, 4);
  61    PointerType *vectorArrayPtr = PointerType::get(vectorArray, 0);
  62
  63    ArrayType   *floatArray     = ArrayType::get(Type::FloatTy, 4);
  64    ArrayType   *constsArray    = ArrayType::get(floatArray, 4);
  65    PointerType *constsArrayPtr = PointerType::get(constsArray, 0);
  66
  67    funcArgs.push_back(vectorArrayPtr);//inputs
  68    funcArgs.push_back(vectorArrayPtr);//output
  69    funcArgs.push_back(constsArrayPtr);//consts
  70    funcArgs.push_back(vectorArrayPtr);//temps
  71
  72    FunctionType *functionType = FunctionType::get(
  73       /*Result=*/Type::VoidTy,
  74       /*Params=*/funcArgs,
  75       /*isVarArg=*/false);
  76
  77    return functionType;
  78 }
  79
  80 static inline void
  81 add_interpolator(struct gallivm_ir *ir,
  82                  struct gallivm_interpolate *interp)
  83 {
  84    ir->interpolators[ir->num_interp] = *interp;
  85    ++ir->num_interp;
  86 }
  87
  88 static void
  89 translate_declaration(struct gallivm_ir *prog,
  90                       llvm::Module *module,
  91                       Storage *storage,
  92                       struct tgsi_full_declaration *decl,
  93                       struct tgsi_full_declaration *fd)
  94 {
  95    if (decl->Declaration.File == TGSI_FILE_INPUT) {
  96       unsigned first, last, mask;
  97       uint interp_method;
  98
  99       assert(decl->Declaration.Declare == TGSI_DECLARE_RANGE);
 100
 101       first = decl->u.DeclarationRange.First;
 102       last = decl->u.DeclarationRange.Last;
 103       mask = decl->Declaration.UsageMask;
 104
 105       /* Do not touch WPOS.xy */
 106       if (first == 0) {
 107          mask &= ~TGSI_WRITEMASK_XY;
 108          if (mask == TGSI_WRITEMASK_NONE) {
 109             first++;
 110             if (first > last) {
 111                return;
 112             }
 113          }
 114       }
 115
 116       interp_method = decl->Interpolation.Interpolate;
 117
 118       if (mask == TGSI_WRITEMASK_XYZW) {
 119          unsigned i, j;
 120
 121          for (i = first; i <= last; i++) {
 122             for (j = 0; j < NUM_CHANNELS; j++) {
 123                //interp( mach, i, j );
 124                struct gallivm_interpolate interp;
 125                interp.type = interp_method;
 126                interp.attrib = i;
 127                interp.chan = j;
 128                add_interpolator(prog, &interp);
 129             }
 130          }
 131       } else {
 132          unsigned i, j;
 133          for( j = 0; j < NUM_CHANNELS; j++ ) {
 134             if( mask & (1 << j) ) {
 135                for( i = first; i <= last; i++ ) {
 136                   struct gallivm_interpolate interp;
 137                   interp.type = interp_method;
 138                   interp.attrib = i;
 139                   interp.chan = j;
 140                   add_interpolator(prog, &interp);
 141                }
 142             }
 143          }
 144       }
 145    }
 146 }
 147
 148 static void
 149 translate_declarationir(struct gallivm_ir *,
 150                       llvm::Module *,
 151                       StorageSoa *storage,
 152                       struct tgsi_full_declaration *decl,
 153                       struct tgsi_full_declaration *)
 154 {
 155    if (decl->Declaration.File == TGSI_FILE_ADDRESS) {
 156       int idx = decl->u.DeclarationRange.First;
 157       storage->addAddress(idx);
 158    }
 159 }
 160
 161 static void
 162 translate_immediate(Storage *storage,
 163                     struct tgsi_full_immediate *imm)
 164 {
 165    float vec[4];
 166    int i;
 167    for (i = 0; i < imm->Immediate.Size - 1; ++i) {
 168       switch (imm->Immediate.DataType) {
 169       case TGSI_IMM_FLOAT32:
 170          vec[i] = imm->u.ImmediateFloat32[i].Float;
 171          break;
 172       default:
 173          assert(0);
 174       }
 175    }
 176    storage->addImmediate(vec);
 177 }
 178
 179
 180 static void
 181 translate_immediateir(StorageSoa *storage,
 182                       struct tgsi_full_immediate *imm)
 183 {
 184    float vec[4];
 185    int i;
 186    for (i = 0; i < imm->Immediate.Size - 1; ++i) {
 187       switch (imm->Immediate.DataType) {
 188       case TGSI_IMM_FLOAT32:
 189          vec[i] = imm->u.ImmediateFloat32[i].Float;
 190          break;
 191       default:
 192          assert(0);
 193       }
 194    }
 195    storage->addImmediate(vec);
 196 }
 197
 198 static inline int
 199 swizzleInt(struct tgsi_full_src_register *src)
 200 {
 201    int swizzle = 0;
 202    int start = 1000;
 203
 204    for (int k = 0; k < 4; ++k) {
 205       swizzle += tgsi_util_get_full_src_register_extswizzle(src, k) * start;
 206       start /= 10;
 207    }
 208    return swizzle;
 209 }
 210
 211 static inline llvm::Value *
 212 swizzleVector(llvm::Value *val, struct tgsi_full_src_register *src,
 213               Storage *storage)
 214 {
 215    int swizzle = swizzleInt(src);
 216
 217    if (gallivm_is_swizzle(swizzle)) {
 218       /*fprintf(stderr, "XXXXXXXX swizzle = %d\n", swizzle);*/
 219       val = storage->shuffleVector(val, swizzle);
 220    }
 221    return val;
 222 }
 223
 224 static void
 225 translate_instruction(llvm::Module *module,
 226                       Storage *storage,
 227                       Instructions *instr,
 228                       struct tgsi_full_instruction *inst,
 229                       struct tgsi_full_instruction *fi,
 230                       unsigned instno)
 231 {
 232    llvm::Value *inputs[4];
 233    inputs[0] = 0;
 234    inputs[1] = 0;
 235    inputs[2] = 0;
 236    inputs[3] = 0;
 237
 238    for (int i = 0; i < inst->Instruction.NumSrcRegs; ++i) {
 239       struct tgsi_full_src_register *src = &inst->FullSrcRegisters[i];
 240       llvm::Value *val = 0;
 241       llvm::Value *indIdx = 0;
 242
 243       if (src->SrcRegister.Indirect) {
 244          indIdx = storage->addrElement(src->SrcRegisterInd.Index);
 245          indIdx = storage->extractIndex(indIdx);
 246       }
 247       if (src->SrcRegister.File == TGSI_FILE_CONSTANT) {
 248          val = storage->constElement(src->SrcRegister.Index, indIdx);
 249       } else if (src->SrcRegister.File == TGSI_FILE_INPUT) {
 250          val = storage->inputElement(src->SrcRegister.Index, indIdx);
 251       } else if (src->SrcRegister.File == TGSI_FILE_TEMPORARY) {
 252          val = storage->tempElement(src->SrcRegister.Index);
 253       } else if (src->SrcRegister.File == TGSI_FILE_OUTPUT) {
 254          val = storage->outputElement(src->SrcRegister.Index, indIdx);
 255       } else if (src->SrcRegister.File == TGSI_FILE_IMMEDIATE) {
 256          val = storage->immediateElement(src->SrcRegister.Index);
 257       } else {
 258          fprintf(stderr, "ERROR: not supported llvm source %d\n", src->SrcRegister.File);
 259          return;
 260       }
 261
 262       inputs[i] = swizzleVector(val, src, storage);
 263    }
 264
 265    /*if (inputs[0])
 266      instr->printVector(inputs[0]);
 267      if (inputs[1])
 268      instr->printVector(inputs[1]);*/
 269    llvm::Value *out = 0;
 270    switch (inst->Instruction.Opcode) {
 271    case TGSI_OPCODE_ARL: {
 272       out = instr->arl(inputs[0]);
 273    }
 274       break;
 275    case TGSI_OPCODE_MOV: {
 276       out = inputs[0];
 277    }
 278       break;
 279    case TGSI_OPCODE_LIT: {
 280       out = instr->lit(inputs[0]);
 281    }
 282       break;
 283    case TGSI_OPCODE_RCP: {
 284       out = instr->rcp(inputs[0]);
 285    }
 286       break;
 287    case TGSI_OPCODE_RSQ: {
 288       out = instr->rsq(inputs[0]);
 289    }
 290       break;
 291    case TGSI_OPCODE_EXP:
 292       break;
 293    case TGSI_OPCODE_LOG:
 294       break;
 295    case TGSI_OPCODE_MUL: {
 296       out = instr->mul(inputs[0], inputs[1]);
 297    }
 298       break;
 299    case TGSI_OPCODE_ADD: {
 300       out = instr->add(inputs[0], inputs[1]);
 301    }
 302       break;
 303    case TGSI_OPCODE_DP3: {
 304       out = instr->dp3(inputs[0], inputs[1]);
 305    }
 306       break;
 307    case TGSI_OPCODE_DP4: {
 308       out = instr->dp4(inputs[0], inputs[1]);
 309    }
 310       break;
 311    case TGSI_OPCODE_DST: {
 312       out = instr->dst(inputs[0], inputs[1]);
 313    }
 314       break;
 315    case TGSI_OPCODE_MIN: {
 316       out = instr->min(inputs[0], inputs[1]);
 317    }
 318       break;
 319    case TGSI_OPCODE_MAX: {
 320       out = instr->max(inputs[0], inputs[1]);
 321    }
 322       break;
 323    case TGSI_OPCODE_SLT: {
 324       out = instr->slt(inputs[0], inputs[1]);
 325    }
 326       break;
 327    case TGSI_OPCODE_SGE: {
 328       out = instr->sge(inputs[0], inputs[1]);
 329    }
 330       break;
 331    case TGSI_OPCODE_MAD: {
 332       out = instr->madd(inputs[0], inputs[1], inputs[2]);
 333    }
 334       break;
 335    case TGSI_OPCODE_SUB: {
 336       out = instr->sub(inputs[0], inputs[1]);
 337    }
 338       break;
 339    case TGSI_OPCODE_LERP: {
 340       out = instr->lerp(inputs[0], inputs[1], inputs[2]);
 341    }
 342       break;
 343    case TGSI_OPCODE_CND:
 344       break;
 345    case TGSI_OPCODE_CND0:
 346       break;
 347    case TGSI_OPCODE_DOT2ADD:
 348       break;
 349    case TGSI_OPCODE_INDEX:
 350       break;
 351    case TGSI_OPCODE_NEGATE:
 352       break;
 353    case TGSI_OPCODE_FRAC: {
 354       out = instr->frc(inputs[0]);
 355    }
 356       break;
 357    case TGSI_OPCODE_CLAMP:
 358       break;
 359    case TGSI_OPCODE_FLOOR: {
 360       out = instr->floor(inputs[0]);
 361    }
 362       break;
 363    case TGSI_OPCODE_ROUND:
 364       break;
 365    case TGSI_OPCODE_EXPBASE2: {
 366       out = instr->ex2(inputs[0]);
 367    }
 368       break;
 369    case TGSI_OPCODE_LOGBASE2: {
 370       out = instr->lg2(inputs[0]);
 371    }
 372       break;
 373    case TGSI_OPCODE_POWER: {
 374       out = instr->pow(inputs[0], inputs[1]);
 375    }
 376       break;
 377    case TGSI_OPCODE_CROSSPRODUCT: {
 378       out = instr->cross(inputs[0], inputs[1]);
 379    }
 380       break;
 381    case TGSI_OPCODE_MULTIPLYMATRIX:
 382       break;
 383    case TGSI_OPCODE_ABS: {
 384       out = instr->abs(inputs[0]);
 385    }
 386       break;
 387    case TGSI_OPCODE_RCC:
 388       break;
 389    case TGSI_OPCODE_DPH: {
 390       out = instr->dph(inputs[0], inputs[1]);
 391    }
 392       break;
 393    case TGSI_OPCODE_COS: {
 394       out = instr->cos(inputs[0]);
 395    }
 396       break;
 397    case TGSI_OPCODE_DDX:
 398       break;
 399    case TGSI_OPCODE_DDY:
 400       break;
 401    case TGSI_OPCODE_KILP: {
 402       out = instr->kilp(inputs[0]);
 403       storage->setKilElement(out);
 404       return;
 405    }
 406       break;
 407    case TGSI_OPCODE_PK2H:
 408       break;
 409    case TGSI_OPCODE_PK2US:
 410       break;
 411    case TGSI_OPCODE_PK4B:
 412       break;
 413    case TGSI_OPCODE_PK4UB:
 414       break;
 415    case TGSI_OPCODE_RFL:
 416       break;
 417    case TGSI_OPCODE_SEQ:
 418       break;
 419    case TGSI_OPCODE_SFL:
 420       break;
 421    case TGSI_OPCODE_SGT: {
 422       out = instr->sgt(inputs[0], inputs[1]);
 423    }
 424       break;
 425    case TGSI_OPCODE_SIN: {
 426       out = instr->sin(inputs[0]);
 427    }
 428       break;
 429    case TGSI_OPCODE_SLE:
 430       break;
 431    case TGSI_OPCODE_SNE:
 432       break;
 433    case TGSI_OPCODE_STR:
 434       break;
 435    case TGSI_OPCODE_TEX:
 436       break;
 437    case TGSI_OPCODE_TXD:
 438       break;
 439    case TGSI_OPCODE_UP2H:
 440       break;
 441    case TGSI_OPCODE_UP2US:
 442       break;
 443    case TGSI_OPCODE_UP4B:
 444       break;
 445    case TGSI_OPCODE_UP4UB:
 446       break;
 447    case TGSI_OPCODE_X2D:
 448       break;
 449    case TGSI_OPCODE_ARA:
 450       break;
 451    case TGSI_OPCODE_ARR:
 452       break;
 453    case TGSI_OPCODE_BRA:
 454       break;
 455    case TGSI_OPCODE_CAL: {
 456       instr->cal(inst->InstructionExtLabel.Label, storage->inputPtr());
 457       return;
 458    }
 459       break;
 460    case TGSI_OPCODE_RET: {
 461       instr->end();
 462       return;
 463    }
 464       break;
 465    case TGSI_OPCODE_SSG:
 466       break;
 467    case TGSI_OPCODE_CMP: {
 468       out = instr->cmp(inputs[0], inputs[1], inputs[2]);
 469    }
 470       break;
 471    case TGSI_OPCODE_SCS: {
 472       out = instr->scs(inputs[0]);
 473    }
 474       break;
 475    case TGSI_OPCODE_TXB:
 476       break;
 477    case TGSI_OPCODE_NRM:
 478       break;
 479    case TGSI_OPCODE_DIV:
 480       break;
 481    case TGSI_OPCODE_DP2:
 482       break;
 483    case TGSI_OPCODE_TXL:
 484       break;
 485    case TGSI_OPCODE_BRK: {
 486       instr->brk();
 487       return;
 488    }
 489       break;
 490    case TGSI_OPCODE_IF: {
 491       instr->ifop(inputs[0]);
 492       storage->setCurrentBlock(instr->currentBlock());
 493       return;  //just update the state
 494    }
 495       break;
 496    case TGSI_OPCODE_LOOP:
 497       break;
 498    case TGSI_OPCODE_REP:
 499       break;
 500    case TGSI_OPCODE_ELSE: {
 501       instr->elseop();
 502       storage->setCurrentBlock(instr->currentBlock());
 503       return; //only state update
 504    }
 505       break;
 506    case TGSI_OPCODE_ENDIF: {
 507       instr->endif();
 508       storage->setCurrentBlock(instr->currentBlock());
 509       return; //just update the state
 510    }
 511       break;
 512    case TGSI_OPCODE_ENDLOOP:
 513       break;
 514    case TGSI_OPCODE_ENDREP:
 515       break;
 516    case TGSI_OPCODE_PUSHA:
 517       break;
 518    case TGSI_OPCODE_POPA:
 519       break;
 520    case TGSI_OPCODE_CEIL:
 521       break;
 522    case TGSI_OPCODE_I2F:
 523       break;
 524    case TGSI_OPCODE_NOT:
 525       break;
 526    case TGSI_OPCODE_TRUNC: {
 527       out = instr->trunc(inputs[0]);
 528    }
 529       break;
 530    case TGSI_OPCODE_SHL:
 531       break;
 532    case TGSI_OPCODE_SHR:
 533       break;
 534    case TGSI_OPCODE_AND:
 535       break;
 536    case TGSI_OPCODE_OR:
 537       break;
 538    case TGSI_OPCODE_MOD:
 539       break;
 540    case TGSI_OPCODE_XOR:
 541       break;
 542    case TGSI_OPCODE_SAD:
 543       break;
 544    case TGSI_OPCODE_TXF:
 545       break;
 546    case TGSI_OPCODE_TXQ:
 547       break;
 548    case TGSI_OPCODE_CONT:
 549       break;
 550    case TGSI_OPCODE_EMIT:
 551       break;
 552    case TGSI_OPCODE_ENDPRIM:
 553       break;
 554    case TGSI_OPCODE_BGNLOOP2: {
 555       instr->beginLoop();
 556       storage->setCurrentBlock(instr->currentBlock());
 557       return;
 558    }
 559       break;
 560    case TGSI_OPCODE_BGNSUB: {
 561       instr->bgnSub(instno);
 562       storage->setCurrentBlock(instr->currentBlock());
 563       storage->pushTemps();
 564       return;
 565    }
 566       break;
 567    case TGSI_OPCODE_ENDLOOP2: {
 568       instr->endLoop();
 569       storage->setCurrentBlock(instr->currentBlock());
 570       return;
 571    }
 572       break;
 573    case TGSI_OPCODE_ENDSUB: {
 574       instr->endSub();
 575       storage->setCurrentBlock(instr->currentBlock());
 576       storage->popArguments();
 577       storage->popTemps();
 578       return;
 579    }
 580       break;
 581    case TGSI_OPCODE_NOISE1:
 582       break;
 583    case TGSI_OPCODE_NOISE2:
 584       break;
 585    case TGSI_OPCODE_NOISE3:
 586       break;
 587    case TGSI_OPCODE_NOISE4:
 588       break;
 589    case TGSI_OPCODE_NOP:
 590       break;
 591    case TGSI_OPCODE_M4X3:
 592       break;
 593    case TGSI_OPCODE_M3X4:
 594       break;
 595    case TGSI_OPCODE_M3X3:
 596       break;
 597    case TGSI_OPCODE_M3X2:
 598       break;
 599    case TGSI_OPCODE_NRM4:
 600       break;
 601    case TGSI_OPCODE_CALLNZ:
 602       break;
 603    case TGSI_OPCODE_IFC:
 604       break;
 605    case TGSI_OPCODE_BREAKC:
 606       break;
 607    case TGSI_OPCODE_KIL:
 608       break;
 609    case TGSI_OPCODE_END:
 610       instr->end();
 611       return;
 612       break;
 613    default:
 614       fprintf(stderr, "ERROR: Unknown opcode %d\n",
 615               inst->Instruction.Opcode);
 616       assert(0);
 617       break;
 618    }
 619
 620    if (!out) {
 621       fprintf(stderr, "ERROR: unsupported opcode %d\n",
 622               inst->Instruction.Opcode);
 623       assert(!"Unsupported opcode");
 624    }
 625
 626    /* # not sure if we need this */
 627    switch( inst->Instruction.Saturate ) {
 628    case TGSI_SAT_NONE:
 629       break;
 630    case TGSI_SAT_ZERO_ONE:
 631       /*TXT( "_SAT" );*/
 632       break;
 633    case TGSI_SAT_MINUS_PLUS_ONE:
 634       /*TXT( "_SAT[-1,1]" );*/
 635       break;
 636    default:
 637       assert( 0 );
 638    }
 639
 640    /* store results  */
 641    for (int i = 0; i < inst->Instruction.NumDstRegs; ++i) {
 642       struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
 643
 644       if (dst->DstRegister.File == TGSI_FILE_OUTPUT) {
 645          storage->setOutputElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
 646       } else if (dst->DstRegister.File == TGSI_FILE_TEMPORARY) {
 647          storage->setTempElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
 648       } else if (dst->DstRegister.File == TGSI_FILE_ADDRESS) {
 649          storage->setAddrElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
 650       } else {
 651          fprintf(stderr, "ERROR: unsupported LLVM destination!");
 652          assert(!"wrong destination");
 653       }
 654    }
 655 }
 656
 657
 658 static void
 659 translate_instructionir(llvm::Module *module,
 660                         StorageSoa *storage,
 661                         InstructionsSoa *instr,
 662                         struct tgsi_full_instruction *inst,
 663                         struct tgsi_full_instruction *fi,
 664                         unsigned instno)
 665 {
 666    std::vector< std::vector<llvm::Value*> > inputs(inst->Instruction.NumSrcRegs);
 667
 668    for (int i = 0; i < inst->Instruction.NumSrcRegs; ++i) {
 669       struct tgsi_full_src_register *src = &inst->FullSrcRegisters[i];
 670       std::vector<llvm::Value*> val;
 671       llvm::Value *indIdx = 0;
 672       int swizzle = swizzleInt(src);
 673
 674       if (src->SrcRegister.Indirect) {
 675          indIdx = storage->addrElement(src->SrcRegisterInd.Index);
 676       }
 677
 678       val = storage->load((enum tgsi_file_type)src->SrcRegister.File,
 679                           src->SrcRegister.Index, swizzle, indIdx);
 680
 681       inputs[i] = val;
 682    }
 683
 684    std::vector<llvm::Value*> out(4);
 685    switch (inst->Instruction.Opcode) {
 686    case TGSI_OPCODE_ARL: {
 687       out = instr->arl(inputs[0]);
 688    }
 689       break;
 690    case TGSI_OPCODE_MOV: {
 691       out = inputs[0];
 692    }
 693       break;
 694    case TGSI_OPCODE_LIT: {
 695    }
 696       break;
 697    case TGSI_OPCODE_RCP: {
 698    }
 699       break;
 700    case TGSI_OPCODE_RSQ: {
 701    }
 702       break;
 703    case TGSI_OPCODE_EXP:
 704       break;
 705    case TGSI_OPCODE_LOG:
 706       break;
 707    case TGSI_OPCODE_MUL: {
 708       out = instr->mul(inputs[0], inputs[1]);
 709    }
 710       break;
 711    case TGSI_OPCODE_ADD: {
 712       out = instr->add(inputs[0], inputs[1]);
 713    }
 714       break;
 715    case TGSI_OPCODE_DP3: {
 716       out = instr->dp3(inputs[0], inputs[1]);
 717    }
 718       break;
 719    case TGSI_OPCODE_DP4: {
 720       out = instr->dp4(inputs[0], inputs[1]);
 721    }
 722       break;
 723    case TGSI_OPCODE_DST: {
 724    }
 725       break;
 726    case TGSI_OPCODE_MIN: {
 727    }
 728       break;
 729    case TGSI_OPCODE_MAX: {
 730    }
 731       break;
 732    case TGSI_OPCODE_SLT: {
 733    }
 734       break;
 735    case TGSI_OPCODE_SGE: {
 736    }
 737       break;
 738    case TGSI_OPCODE_MAD: {
 739       out = instr->madd(inputs[0], inputs[1], inputs[2]);
 740    }
 741       break;
 742    case TGSI_OPCODE_SUB: {
 743    }
 744       break;
 745    case TGSI_OPCODE_LERP: {
 746    }
 747       break;
 748    case TGSI_OPCODE_CND:
 749       break;
 750    case TGSI_OPCODE_CND0:
 751       break;
 752    case TGSI_OPCODE_DOT2ADD:
 753       break;
 754    case TGSI_OPCODE_INDEX:
 755       break;
 756    case TGSI_OPCODE_NEGATE:
 757       break;
 758    case TGSI_OPCODE_FRAC: {
 759    }
 760       break;
 761    case TGSI_OPCODE_CLAMP:
 762       break;
 763    case TGSI_OPCODE_FLOOR: {
 764    }
 765       break;
 766    case TGSI_OPCODE_ROUND:
 767       break;
 768    case TGSI_OPCODE_EXPBASE2: {
 769    }
 770       break;
 771    case TGSI_OPCODE_LOGBASE2: {
 772    }
 773       break;
 774    case TGSI_OPCODE_POWER: {
 775       out = instr->pow(inputs[0], inputs[1]);
 776    }
 777       break;
 778    case TGSI_OPCODE_CROSSPRODUCT: {
 779    }
 780       break;
 781    case TGSI_OPCODE_MULTIPLYMATRIX:
 782       break;
 783    case TGSI_OPCODE_ABS: {
 784    }
 785       break;
 786    case TGSI_OPCODE_RCC:
 787       break;
 788    case TGSI_OPCODE_DPH: {
 789    }
 790       break;
 791    case TGSI_OPCODE_COS: {
 792    }
 793       break;
 794    case TGSI_OPCODE_DDX:
 795       break;
 796    case TGSI_OPCODE_DDY:
 797       break;
 798    case TGSI_OPCODE_KILP: {
 799    }
 800       break;
 801    case TGSI_OPCODE_PK2H:
 802       break;
 803    case TGSI_OPCODE_PK2US:
 804       break;
 805    case TGSI_OPCODE_PK4B:
 806       break;
 807    case TGSI_OPCODE_PK4UB:
 808       break;
 809    case TGSI_OPCODE_RFL:
 810       break;
 811    case TGSI_OPCODE_SEQ:
 812       break;
 813    case TGSI_OPCODE_SFL:
 814       break;
 815    case TGSI_OPCODE_SGT: {
 816    }
 817       break;
 818    case TGSI_OPCODE_SIN: {
 819    }
 820       break;
 821    case TGSI_OPCODE_SLE:
 822       break;
 823    case TGSI_OPCODE_SNE:
 824       break;
 825    case TGSI_OPCODE_STR:
 826       break;
 827    case TGSI_OPCODE_TEX:
 828       break;
 829    case TGSI_OPCODE_TXD:
 830       break;
 831    case TGSI_OPCODE_UP2H:
 832       break;
 833    case TGSI_OPCODE_UP2US:
 834       break;
 835    case TGSI_OPCODE_UP4B:
 836       break;
 837    case TGSI_OPCODE_UP4UB:
 838       break;
 839    case TGSI_OPCODE_X2D:
 840       break;
 841    case TGSI_OPCODE_ARA:
 842       break;
 843    case TGSI_OPCODE_ARR:
 844       break;
 845    case TGSI_OPCODE_BRA:
 846       break;
 847    case TGSI_OPCODE_CAL: {
 848    }
 849       break;
 850    case TGSI_OPCODE_RET: {
 851    }
 852       break;
 853    case TGSI_OPCODE_SSG:
 854       break;
 855    case TGSI_OPCODE_CMP: {
 856    }
 857       break;
 858    case TGSI_OPCODE_SCS: {
 859    }
 860       break;
 861    case TGSI_OPCODE_TXB:
 862       break;
 863    case TGSI_OPCODE_NRM:
 864       break;
 865    case TGSI_OPCODE_DIV:
 866       break;
 867    case TGSI_OPCODE_DP2:
 868       break;
 869    case TGSI_OPCODE_TXL:
 870       break;
 871    case TGSI_OPCODE_BRK: {
 872    }
 873       break;
 874    case TGSI_OPCODE_IF: {
 875    }
 876       break;
 877    case TGSI_OPCODE_LOOP:
 878       break;
 879    case TGSI_OPCODE_REP:
 880       break;
 881    case TGSI_OPCODE_ELSE: {
 882    }
 883       break;
 884    case TGSI_OPCODE_ENDIF: {
 885    }
 886       break;
 887    case TGSI_OPCODE_ENDLOOP:
 888       break;
 889    case TGSI_OPCODE_ENDREP:
 890       break;
 891    case TGSI_OPCODE_PUSHA:
 892       break;
 893    case TGSI_OPCODE_POPA:
 894       break;
 895    case TGSI_OPCODE_CEIL:
 896       break;
 897    case TGSI_OPCODE_I2F:
 898       break;
 899    case TGSI_OPCODE_NOT:
 900       break;
 901    case TGSI_OPCODE_TRUNC: {
 902    }
 903       break;
 904    case TGSI_OPCODE_SHL:
 905       break;
 906    case TGSI_OPCODE_SHR:
 907       break;
 908    case TGSI_OPCODE_AND:
 909       break;
 910    case TGSI_OPCODE_OR:
 911       break;
 912    case TGSI_OPCODE_MOD:
 913       break;
 914    case TGSI_OPCODE_XOR:
 915       break;
 916    case TGSI_OPCODE_SAD:
 917       break;
 918    case TGSI_OPCODE_TXF:
 919       break;
 920    case TGSI_OPCODE_TXQ:
 921       break;
 922    case TGSI_OPCODE_CONT:
 923       break;
 924    case TGSI_OPCODE_EMIT:
 925       break;
 926    case TGSI_OPCODE_ENDPRIM:
 927       break;
 928    case TGSI_OPCODE_BGNLOOP2: {
 929    }
 930       break;
 931    case TGSI_OPCODE_BGNSUB: {
 932    }
 933       break;
 934    case TGSI_OPCODE_ENDLOOP2: {
 935    }
 936       break;
 937    case TGSI_OPCODE_ENDSUB: {
 938    }
 939       break;
 940    case TGSI_OPCODE_NOISE1:
 941       break;
 942    case TGSI_OPCODE_NOISE2:
 943       break;
 944    case TGSI_OPCODE_NOISE3:
 945       break;
 946    case TGSI_OPCODE_NOISE4:
 947       break;
 948    case TGSI_OPCODE_NOP:
 949       break;
 950    case TGSI_OPCODE_M4X3:
 951       break;
 952    case TGSI_OPCODE_M3X4:
 953       break;
 954    case TGSI_OPCODE_M3X3:
 955       break;
 956    case TGSI_OPCODE_M3X2:
 957       break;
 958    case TGSI_OPCODE_NRM4:
 959       break;
 960    case TGSI_OPCODE_CALLNZ:
 961       break;
 962    case TGSI_OPCODE_IFC:
 963       break;
 964    case TGSI_OPCODE_BREAKC:
 965       break;
 966    case TGSI_OPCODE_KIL:
 967       break;
 968    case TGSI_OPCODE_END:
 969       instr->end();
 970       return;
 971       break;
 972    default:
 973       fprintf(stderr, "ERROR: Unknown opcode %d\n",
 974               inst->Instruction.Opcode);
 975       assert(0);
 976       break;
 977    }
 978
 979    if (!out[0]) {
 980       fprintf(stderr, "ERROR: unsupported opcode %d\n",
 981               inst->Instruction.Opcode);
 982       assert(!"Unsupported opcode");
 983    }
 984
 985    /* store results  */
 986    for (int i = 0; i < inst->Instruction.NumDstRegs; ++i) {
 987       struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
 988
 989       storage->store((enum tgsi_file_type)dst->DstRegister.File,
 990                      dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
 991    }
 992 }
 993
 994 llvm::Module *
 995 tgsi_to_llvm(struct gallivm_ir *ir, const struct tgsi_token *tokens)
 996 {
 997    llvm::Module *mod = new Module("shader");
 998    struct tgsi_parse_context parse;
 999    struct tgsi_full_instruction fi;
1000    struct tgsi_full_declaration fd;
1001    unsigned instno = 0;
1002    Function* shader = mod->getFunction("execute_shader");
1003    std::ostringstream stream;
1004    if (ir->type == GALLIVM_VS) {
1005       stream << "vs_shader";
1006    } else {
1007       stream << "fs_shader";
1008    }
1009    stream << ir->id;
1010    std::string func_name = stream.str();
1011    shader->setName(func_name.c_str());
1012
1013    Function::arg_iterator args = shader->arg_begin();
1014    Value *ptr_INPUT = args++;
1015    ptr_INPUT->setName("input");
1016
1017    BasicBlock *label_entry = new BasicBlock("entry", shader, 0);
1018
1019    tgsi_parse_init(&parse, tokens);
1020
1021    fi = tgsi_default_full_instruction();
1022    fd = tgsi_default_full_declaration();
1023    Storage storage(label_entry, ptr_INPUT);
1024    Instructions instr(mod, shader, label_entry, &storage);
1025    while(!tgsi_parse_end_of_tokens(&parse)) {
1026       tgsi_parse_token(&parse);
1027
1028       switch (parse.FullToken.Token.Type) {
1029       case TGSI_TOKEN_TYPE_DECLARATION:
1030          translate_declaration(ir, mod, &storage,
1031                                &parse.FullToken.FullDeclaration,
1032                                &fd);
1033          break;
1034
1035       case TGSI_TOKEN_TYPE_IMMEDIATE:
1036          translate_immediate(&storage,
1037                              &parse.FullToken.FullImmediate);
1038          break;
1039
1040       case TGSI_TOKEN_TYPE_INSTRUCTION:
1041          translate_instruction(mod, &storage, &instr,
1042                                &parse.FullToken.FullInstruction,
1043                                &fi, instno);
1044          ++instno;
1045          break;
1046
1047       default:
1048          assert(0);
1049       }
1050    }
1051
1052    tgsi_parse_free(&parse);
1053
1054    ir->num_consts = storage.numConsts();
1055    return mod;
1056 }
1057
1058 llvm::Module * tgsi_to_llvmir(struct gallivm_ir *ir,
1059                               const struct tgsi_token *tokens)
1060 {
1061    llvm::Module *mod = new Module("shader");
1062    struct tgsi_parse_context parse;
1063    struct tgsi_full_instruction fi;
1064    struct tgsi_full_declaration fd;
1065    unsigned instno = 0;
1066    std::ostringstream stream;
1067    if (ir->type == GALLIVM_VS) {
1068       stream << "vs_shader";
1069    } else {
1070       stream << "fs_shader";
1071    }
1072    //stream << ir->id;
1073    std::string func_name = stream.str();
1074    Function *shader = llvm::cast<Function>(mod->getOrInsertFunction(
1075                                               func_name.c_str(),
1076                                               vertexShaderFunctionType()));
1077
1078    Function::arg_iterator args = shader->arg_begin();
1079    Value *input = args++;
1080    input->setName("inputs");
1081    Value *output = args++;
1082    output->setName("outputs");
1083    Value *consts = args++;
1084    consts->setName("consts");
1085    Value *temps = args++;
1086    temps->setName("temps");
1087
1088    BasicBlock *label_entry = new BasicBlock("entry", shader, 0);
1089
1090    tgsi_parse_init(&parse, tokens);
1091
1092    fi = tgsi_default_full_instruction();
1093    fd = tgsi_default_full_declaration();
1094
1095    StorageSoa storage(label_entry, input, output, consts, temps);
1096    InstructionsSoa instr(mod, shader, label_entry, &storage);
1097
1098    while(!tgsi_parse_end_of_tokens(&parse)) {
1099       tgsi_parse_token(&parse);
1100
1101       switch (parse.FullToken.Token.Type) {
1102       case TGSI_TOKEN_TYPE_DECLARATION:
1103          translate_declarationir(ir, mod, &storage,
1104                                  &parse.FullToken.FullDeclaration,
1105                                  &fd);
1106          break;
1107
1108       case TGSI_TOKEN_TYPE_IMMEDIATE:
1109          translate_immediateir(&storage,
1110                              &parse.FullToken.FullImmediate);
1111          break;
1112
1113       case TGSI_TOKEN_TYPE_INSTRUCTION:
1114          storage.declareImmediates();
1115          translate_instructionir(mod, &storage, &instr,
1116                                  &parse.FullToken.FullInstruction,
1117                                  &fi, instno);
1118          ++instno;
1119          break;
1120
1121       default:
1122          assert(0);
1123       }
1124    }
1125
1126    tgsi_parse_free(&parse);
1127
1128    return mod;
1129 }