src/OpenGL/compiler/OutputASM.cpp

   1 // Copyright 2016 The SwiftShader Authors. All Rights Reserved.
   2 //
   3 // Licensed under the Apache License, Version 2.0 (the "License");
   4 // you may not use this file except in compliance with the License.
   5 // You may obtain a copy of the License at
   6 //
   7 //    http://www.apache.org/licenses/LICENSE-2.0
   8 //
   9 // Unless required by applicable law or agreed to in writing, software
  10 // distributed under the License is distributed on an "AS IS" BASIS,
  11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12 // See the License for the specific language governing permissions and
  13 // limitations under the License.
  14
  15 #include "OutputASM.h"
  16 #include "Common/Math.hpp"
  17
  18 #include "common/debug.h"
  19 #include "InfoSink.h"
  20
  21 #include "libGLESv2/Shader.h"
  22
  23 #include <GLES2/gl2.h>
  24 #include <GLES2/gl2ext.h>
  25 #include <GLES3/gl3.h>
  26
  27 namespace glsl
  28 {
  29         // Integer to TString conversion
  30         TString str(int i)
  31         {
  32                 char buffer[20];
  33                 sprintf(buffer, "%d", i);
  34                 return buffer;
  35         }
  36
  37         class Temporary : public TIntermSymbol
  38         {
  39         public:
  40                 Temporary(OutputASM *assembler) : TIntermSymbol(TSymbolTableLevel::nextUniqueId(), "tmp", TType(EbtFloat, EbpHigh, EvqTemporary, 4, 1, false)), assembler(assembler)
  41                 {
  42                 }
  43
  44                 ~Temporary()
  45                 {
  46                         assembler->freeTemporary(this);
  47                 }
  48
  49         private:
  50                 OutputASM *const assembler;
  51         };
  52
  53         class Constant : public TIntermConstantUnion
  54         {
  55         public:
  56                 Constant(float x, float y, float z, float w) : TIntermConstantUnion(constants, TType(EbtFloat, EbpHigh, EvqConstExpr, 4, 1, false))
  57                 {
  58                         constants[0].setFConst(x);
  59                         constants[1].setFConst(y);
  60                         constants[2].setFConst(z);
  61                         constants[3].setFConst(w);
  62                 }
  63
  64                 Constant(bool b) : TIntermConstantUnion(constants, TType(EbtBool, EbpHigh, EvqConstExpr, 1, 1, false))
  65                 {
  66                         constants[0].setBConst(b);
  67                 }
  68
  69                 Constant(int i) : TIntermConstantUnion(constants, TType(EbtInt, EbpHigh, EvqConstExpr, 1, 1, false))
  70                 {
  71                         constants[0].setIConst(i);
  72                 }
  73
  74                 ~Constant()
  75                 {
  76                 }
  77
  78         private:
  79                 ConstantUnion constants[4];
  80         };
  81
  82         Uniform::Uniform(GLenum type, GLenum precision, const std::string &name, int arraySize, int registerIndex, int blockId, const BlockMemberInfo& blockMemberInfo) :
  83                 type(type), precision(precision), name(name), arraySize(arraySize), registerIndex(registerIndex), blockId(blockId), blockInfo(blockMemberInfo)
  84         {
  85         }
  86
  87         UniformBlock::UniformBlock(const std::string& name, unsigned int dataSize, unsigned int arraySize,
  88                                    TLayoutBlockStorage layout, bool isRowMajorLayout, int registerIndex, int blockId) :
  89                 name(name), dataSize(dataSize), arraySize(arraySize), layout(layout),
  90                 isRowMajorLayout(isRowMajorLayout), registerIndex(registerIndex), blockId(blockId)
  91         {
  92         }
  93
  94         BlockLayoutEncoder::BlockLayoutEncoder(bool rowMajor)
  95                 : mCurrentOffset(0), isRowMajor(rowMajor)
  96         {
  97         }
  98
  99         BlockMemberInfo BlockLayoutEncoder::encodeType(const TType &type)
 100         {
 101                 int arrayStride;
 102                 int matrixStride;
 103
 104                 getBlockLayoutInfo(type, type.getArraySize(), isRowMajor, &arrayStride, &matrixStride);
 105
 106                 const BlockMemberInfo memberInfo(static_cast<int>(mCurrentOffset * BytesPerComponent),
 107                                                  static_cast<int>(arrayStride * BytesPerComponent),
 108                                                  static_cast<int>(matrixStride * BytesPerComponent),
 109                                                  (matrixStride > 0) && isRowMajor);
 110
 111                 advanceOffset(type, type.getArraySize(), isRowMajor, arrayStride, matrixStride);
 112
 113                 return memberInfo;
 114         }
 115
 116         // static
 117         size_t BlockLayoutEncoder::getBlockRegister(const BlockMemberInfo &info)
 118         {
 119                 return (info.offset / BytesPerComponent) / ComponentsPerRegister;
 120         }
 121
 122         // static
 123         size_t BlockLayoutEncoder::getBlockRegisterElement(const BlockMemberInfo &info)
 124         {
 125                 return (info.offset / BytesPerComponent) % ComponentsPerRegister;
 126         }
 127
 128         void BlockLayoutEncoder::nextRegister()
 129         {
 130                 mCurrentOffset = sw::align(mCurrentOffset, ComponentsPerRegister);
 131         }
 132
 133         Std140BlockEncoder::Std140BlockEncoder(bool rowMajor) : BlockLayoutEncoder(rowMajor)
 134         {
 135         }
 136
 137         void Std140BlockEncoder::enterAggregateType()
 138         {
 139                 nextRegister();
 140         }
 141
 142         void Std140BlockEncoder::exitAggregateType()
 143         {
 144                 nextRegister();
 145         }
 146
 147         void Std140BlockEncoder::getBlockLayoutInfo(const TType &type, unsigned int arraySize, bool isRowMajorMatrix, int *arrayStrideOut, int *matrixStrideOut)
 148         {
 149                 size_t baseAlignment = 0;
 150                 int matrixStride = 0;
 151                 int arrayStride = 0;
 152
 153                 if(type.isMatrix())
 154                 {
 155                         baseAlignment = ComponentsPerRegister;
 156                         matrixStride = ComponentsPerRegister;
 157
 158                         if(arraySize > 0)
 159                         {
 160                                 const int numRegisters = isRowMajorMatrix ? type.getSecondarySize() : type.getNominalSize();
 161                                 arrayStride = ComponentsPerRegister * numRegisters;
 162                         }
 163                 }
 164                 else if(arraySize > 0)
 165                 {
 166                         baseAlignment = ComponentsPerRegister;
 167                         arrayStride = ComponentsPerRegister;
 168                 }
 169                 else
 170                 {
 171                         const size_t numComponents = type.getElementSize();
 172                         baseAlignment = (numComponents == 3 ? 4u : numComponents);
 173                 }
 174
 175                 mCurrentOffset = sw::align(mCurrentOffset, baseAlignment);
 176
 177                 *matrixStrideOut = matrixStride;
 178                 *arrayStrideOut = arrayStride;
 179         }
 180
 181         void Std140BlockEncoder::advanceOffset(const TType &type, unsigned int arraySize, bool isRowMajorMatrix, int arrayStride, int matrixStride)
 182         {
 183                 if(arraySize > 0)
 184                 {
 185                         mCurrentOffset += arrayStride * arraySize;
 186                 }
 187                 else if(type.isMatrix())
 188                 {
 189                         ASSERT(matrixStride == ComponentsPerRegister);
 190                         const int numRegisters = isRowMajorMatrix ? type.getSecondarySize() : type.getNominalSize();
 191                         mCurrentOffset += ComponentsPerRegister * numRegisters;
 192                 }
 193                 else
 194                 {
 195                         mCurrentOffset += type.getElementSize();
 196                 }
 197         }
 198
 199         Attribute::Attribute()
 200         {
 201                 type = GL_NONE;
 202                 arraySize = 0;
 203                 registerIndex = 0;
 204         }
 205
 206         Attribute::Attribute(GLenum type, const std::string &name, int arraySize, int location, int registerIndex)
 207         {
 208                 this->type = type;
 209                 this->name = name;
 210                 this->arraySize = arraySize;
 211                 this->location = location;
 212                 this->registerIndex = registerIndex;
 213         }
 214
 215         sw::PixelShader *Shader::getPixelShader() const
 216         {
 217                 return 0;
 218         }
 219
 220         sw::VertexShader *Shader::getVertexShader() const
 221         {
 222                 return 0;
 223         }
 224
 225         OutputASM::TextureFunction::TextureFunction(const TString& nodeName) : method(IMPLICIT), proj(false), offset(false)
 226         {
 227                 TString name = TFunction::unmangleName(nodeName);
 228
 229                 if(name == "texture2D" || name == "textureCube" || name == "texture" || name == "texture3D")
 230                 {
 231                         method = IMPLICIT;
 232                 }
 233                 else if(name == "texture2DProj" || name == "textureProj")
 234                 {
 235                         method = IMPLICIT;
 236                         proj = true;
 237                 }
 238                 else if(name == "texture2DLod" || name == "textureCubeLod" || name == "textureLod")
 239                 {
 240                         method = LOD;
 241                 }
 242                 else if(name == "texture2DProjLod" || name == "textureProjLod")
 243                 {
 244                         method = LOD;
 245                         proj = true;
 246                 }
 247                 else if(name == "textureSize")
 248                 {
 249                         method = SIZE;
 250                 }
 251                 else if(name == "textureOffset")
 252                 {
 253                         method = IMPLICIT;
 254                         offset = true;
 255                 }
 256                 else if(name == "textureProjOffset")
 257                 {
 258                         method = IMPLICIT;
 259                         offset = true;
 260                         proj = true;
 261                 }
 262                 else if(name == "textureLodOffset")
 263                 {
 264                         method = LOD;
 265                         offset = true;
 266                 }
 267                 else if(name == "textureProjLodOffset")
 268                 {
 269                         method = LOD;
 270                         proj = true;
 271                         offset = true;
 272                 }
 273                 else if(name == "texelFetch")
 274                 {
 275                         method = FETCH;
 276                 }
 277                 else if(name == "texelFetchOffset")
 278                 {
 279                         method = FETCH;
 280                         offset = true;
 281                 }
 282                 else if(name == "textureGrad")
 283                 {
 284                         method = GRAD;
 285                 }
 286                 else if(name == "textureGradOffset")
 287                 {
 288                         method = GRAD;
 289                         offset = true;
 290                 }
 291                 else if(name == "textureProjGrad")
 292                 {
 293                         method = GRAD;
 294                         proj = true;
 295                 }
 296                 else if(name == "textureProjGradOffset")
 297                 {
 298                         method = GRAD;
 299                         proj = true;
 300                         offset = true;
 301                 }
 302                 else UNREACHABLE(0);
 303         }
 304
 305         OutputASM::OutputASM(TParseContext &context, Shader *shaderObject) : TIntermTraverser(true, true, true), shaderObject(shaderObject), mContext(context)
 306         {
 307                 shader = 0;
 308                 pixelShader = 0;
 309                 vertexShader = 0;
 310
 311                 if(shaderObject)
 312                 {
 313                         shader = shaderObject->getShader();
 314                         pixelShader = shaderObject->getPixelShader();
 315                         vertexShader = shaderObject->getVertexShader();
 316                 }
 317
 318                 functionArray.push_back(Function(0, "main(", 0, 0));
 319                 currentFunction = 0;
 320                 outputQualifier = EvqOutput; // Set outputQualifier to any value other than EvqFragColor or EvqFragData
 321         }
 322
 323         OutputASM::~OutputASM()
 324         {
 325         }
 326
 327         void OutputASM::output()
 328         {
 329                 if(shader)
 330                 {
 331                         emitShader(GLOBAL);
 332
 333                         if(functionArray.size() > 1)   // Only call main() when there are other functions
 334                         {
 335                                 Instruction *callMain = emit(sw::Shader::OPCODE_CALL);
 336                                 callMain->dst.type = sw::Shader::PARAMETER_LABEL;
 337                                 callMain->dst.index = 0;   // main()
 338
 339                                 emit(sw::Shader::OPCODE_RET);
 340                         }
 341
 342                         emitShader(FUNCTION);
 343                 }
 344         }
 345
 346         void OutputASM::emitShader(Scope scope)
 347         {
 348                 emitScope = scope;
 349                 currentScope = GLOBAL;
 350                 mContext.getTreeRoot()->traverse(this);
 351         }
 352
 353         void OutputASM::freeTemporary(Temporary *temporary)
 354         {
 355                 free(temporaries, temporary);
 356         }
 357
 358         sw::Shader::Opcode OutputASM::getOpcode(sw::Shader::Opcode op, TIntermTyped *in) const
 359         {
 360                 TBasicType baseType = in->getType().getBasicType();
 361
 362                 switch(op)
 363                 {
 364                 case sw::Shader::OPCODE_NEG:
 365                         switch(baseType)
 366                         {
 367                         case EbtInt:
 368                         case EbtUInt:
 369                                 return sw::Shader::OPCODE_INEG;
 370                         case EbtFloat:
 371                         default:
 372                                 return op;
 373                         }
 374                 case sw::Shader::OPCODE_ABS:
 375                         switch(baseType)
 376                         {
 377                         case EbtInt:
 378                                 return sw::Shader::OPCODE_IABS;
 379                         case EbtFloat:
 380                         default:
 381                                 return op;
 382                         }
 383                 case sw::Shader::OPCODE_SGN:
 384                         switch(baseType)
 385                         {
 386                         case EbtInt:
 387                                 return sw::Shader::OPCODE_ISGN;
 388                         case EbtFloat:
 389                         default:
 390                                 return op;
 391                         }
 392                 case sw::Shader::OPCODE_ADD:
 393                         switch(baseType)
 394                         {
 395                         case EbtInt:
 396                         case EbtUInt:
 397                                 return sw::Shader::OPCODE_IADD;
 398                         case EbtFloat:
 399                         default:
 400                                 return op;
 401                         }
 402                 case sw::Shader::OPCODE_SUB:
 403                         switch(baseType)
 404                         {
 405                         case EbtInt:
 406                         case EbtUInt:
 407                                 return sw::Shader::OPCODE_ISUB;
 408                         case EbtFloat:
 409                         default:
 410                                 return op;
 411                         }
 412                 case sw::Shader::OPCODE_MUL:
 413                         switch(baseType)
 414                         {
 415                         case EbtInt:
 416                         case EbtUInt:
 417                                 return sw::Shader::OPCODE_IMUL;
 418                         case EbtFloat:
 419                         default:
 420                                 return op;
 421                         }
 422                 case sw::Shader::OPCODE_DIV:
 423                         switch(baseType)
 424                         {
 425                         case EbtInt:
 426                                 return sw::Shader::OPCODE_IDIV;
 427                         case EbtUInt:
 428                                 return sw::Shader::OPCODE_UDIV;
 429                         case EbtFloat:
 430                         default:
 431                                 return op;
 432                         }
 433                 case sw::Shader::OPCODE_IMOD:
 434                         return baseType == EbtUInt ? sw::Shader::OPCODE_UMOD : op;
 435                 case sw::Shader::OPCODE_ISHR:
 436                         return baseType == EbtUInt ? sw::Shader::OPCODE_USHR : op;
 437                 case sw::Shader::OPCODE_MIN:
 438                         switch(baseType)
 439                         {
 440                         case EbtInt:
 441                                 return sw::Shader::OPCODE_IMIN;
 442                         case EbtUInt:
 443                                 return sw::Shader::OPCODE_UMIN;
 444                         case EbtFloat:
 445                         default:
 446                                 return op;
 447                         }
 448                 case sw::Shader::OPCODE_MAX:
 449                         switch(baseType)
 450                         {
 451                         case EbtInt:
 452                                 return sw::Shader::OPCODE_IMAX;
 453                         case EbtUInt:
 454                                 return sw::Shader::OPCODE_UMAX;
 455                         case EbtFloat:
 456                         default:
 457                                 return op;
 458                         }
 459                 default:
 460                         return op;
 461                 }
 462         }
 463
 464         void OutputASM::visitSymbol(TIntermSymbol *symbol)
 465         {
 466                 // Vertex varyings don't have to be actively used to successfully link
 467                 // against pixel shaders that use them. So make sure they're declared.
 468                 if(symbol->getQualifier() == EvqVaryingOut || symbol->getQualifier() == EvqInvariantVaryingOut || symbol->getQualifier() == EvqVertexOut)
 469                 {
 470                         if(symbol->getBasicType() != EbtInvariant)   // Typeless declarations are not new varyings
 471                         {
 472                                 declareVarying(symbol, -1);
 473                         }
 474                 }
 475
 476                 TInterfaceBlock* block = symbol->getType().getInterfaceBlock();
 477                 // OpenGL ES 3.0.4 spec, section 2.12.6 Uniform Variables:
 478                 // "All members of a named uniform block declared with a shared or std140 layout qualifier
 479                 // are considered active, even if they are not referenced in any shader in the program.
 480                 // The uniform block itself is also considered active, even if no member of the block is referenced."
 481                 if(block && ((block->blockStorage() == EbsShared) || (block->blockStorage() == EbsStd140)))
 482                 {
 483                         uniformRegister(symbol);
 484                 }
 485         }
 486
 487         bool OutputASM::visitBinary(Visit visit, TIntermBinary *node)
 488         {
 489                 if(currentScope != emitScope)
 490                 {
 491                         return false;
 492                 }
 493
 494                 TIntermTyped *result = node;
 495                 TIntermTyped *left = node->getLeft();
 496                 TIntermTyped *right = node->getRight();
 497                 const TType &leftType = left->getType();
 498                 const TType &rightType = right->getType();
 499                 const TType &resultType = node->getType();
 500
 501                 if(isSamplerRegister(result))
 502                 {
 503                         return false;   // Don't traverse, the register index is determined statically
 504                 }
 505
 506                 switch(node->getOp())
 507                 {
 508                 case EOpAssign:
 509                         if(visit == PostVisit)
 510                         {
 511                                 assignLvalue(left, right);
 512                                 copy(result, right);
 513                         }
 514                         break;
 515                 case EOpInitialize:
 516                         if(visit == PostVisit)
 517                         {
 518                                 copy(left, right);
 519                         }
 520                         break;
 521                 case EOpMatrixTimesScalarAssign:
 522                         if(visit == PostVisit)
 523                         {
 524                                 for(int i = 0; i < leftType.getNominalSize(); i++)
 525                                 {
 526                                         emit(sw::Shader::OPCODE_MUL, result, i, left, i, right);
 527                                 }
 528
 529                                 assignLvalue(left, result);
 530                         }
 531                         break;
 532                 case EOpVectorTimesMatrixAssign:
 533                         if(visit == PostVisit)
 534                         {
 535                                 int size = leftType.getNominalSize();
 536
 537                                 for(int i = 0; i < size; i++)
 538                                 {
 539                                         Instruction *dot = emit(sw::Shader::OPCODE_DP(size), result, 0, left, 0, right, i);
 540                                         dot->dst.mask = 1 << i;
 541                                 }
 542
 543                                 assignLvalue(left, result);
 544                         }
 545                         break;
 546                 case EOpMatrixTimesMatrixAssign:
 547                         if(visit == PostVisit)
 548                         {
 549                                 int dim = leftType.getNominalSize();
 550
 551                                 for(int i = 0; i < dim; i++)
 552                                 {
 553                                         Instruction *mul = emit(sw::Shader::OPCODE_MUL, result, i, left, 0, right, i);
 554                                         mul->src[1].swizzle = 0x00;
 555
 556                                         for(int j = 1; j < dim; j++)
 557                                         {
 558                                                 Instruction *mad = emit(sw::Shader::OPCODE_MAD, result, i, left, j, right, i, result, i);
 559                                                 mad->src[1].swizzle = j * 0x55;
 560                                         }
 561                                 }
 562
 563                                 assignLvalue(left, result);
 564                         }
 565                         break;
 566                 case EOpIndexDirect:
 567                         if(visit == PostVisit)
 568                         {
 569                                 int index = right->getAsConstantUnion()->getIConst(0);
 570
 571                                 if(result->isMatrix() || result->isStruct() || result->isInterfaceBlock())
 572                                 {
 573                                         ASSERT(left->isArray());
 574                                         copy(result, left, index * left->elementRegisterCount());
 575                                 }
 576                                 else if(result->isRegister())
 577                                 {
 578                                         int srcIndex = 0;
 579                                         if(left->isRegister())
 580                                         {
 581                                                 srcIndex = 0;
 582                                         }
 583                                         else if(left->isArray())
 584                                         {
 585                                                 srcIndex = index * left->elementRegisterCount();
 586                                         }
 587                                         else if(left->isMatrix())
 588                                         {
 589                                                 ASSERT(index < left->getNominalSize());   // FIXME: Report semantic error
 590                                                 srcIndex = index;
 591                                         }
 592                                         else UNREACHABLE(0);
 593
 594                                         Instruction *mov = emit(sw::Shader::OPCODE_MOV, result, 0, left, srcIndex);
 595
 596                                         if(left->isRegister())
 597                                         {
 598                                                 mov->src[0].swizzle = index;
 599                                         }
 600                                 }
 601                                 else UNREACHABLE(0);
 602                         }
 603                         break;
 604                 case EOpIndexIndirect:
 605                         if(visit == PostVisit)
 606                         {
 607                                 if(left->isArray() || left->isMatrix())
 608                                 {
 609                                         for(int index = 0; index < result->totalRegisterCount(); index++)
 610                                         {
 611                                                 Instruction *mov = emit(sw::Shader::OPCODE_MOV, result, index, left, index);
 612                                                 mov->dst.mask = writeMask(result, index);
 613
 614                                                 if(left->totalRegisterCount() > 1)
 615                                                 {
 616                                                         sw::Shader::SourceParameter relativeRegister;
 617                                                         argument(relativeRegister, right);
 618
 619                                                         mov->src[0].rel.type = relativeRegister.type;
 620                                                         mov->src[0].rel.index = relativeRegister.index;
 621                                                         mov->src[0].rel.scale = result->totalRegisterCount();
 622                                                         mov->src[0].rel.deterministic = !(vertexShader && left->getQualifier() == EvqUniform);
 623                                                 }
 624                                         }
 625                                 }
 626                                 else if(left->isRegister())
 627                                 {
 628                                         emit(sw::Shader::OPCODE_EXTRACT, result, left, right);
 629                                 }
 630                                 else UNREACHABLE(0);
 631                         }
 632                         break;
 633                 case EOpIndexDirectStruct:
 634                 case EOpIndexDirectInterfaceBlock:
 635                         if(visit == PostVisit)
 636                         {
 637                                 ASSERT(leftType.isStruct() || (leftType.isInterfaceBlock()));
 638
 639                                 const TFieldList& fields = (node->getOp() == EOpIndexDirectStruct) ?
 640                                                            leftType.getStruct()->fields() :
 641                                                            leftType.getInterfaceBlock()->fields();
 642                                 int index = right->getAsConstantUnion()->getIConst(0);
 643                                 int fieldOffset = 0;
 644
 645                                 for(int i = 0; i < index; i++)
 646                                 {
 647                                         fieldOffset += fields[i]->type()->totalRegisterCount();
 648                                 }
 649
 650                                 copy(result, left, fieldOffset);
 651                         }
 652                         break;
 653                 case EOpVectorSwizzle:
 654                         if(visit == PostVisit)
 655                         {
 656                                 int swizzle = 0;
 657                                 TIntermAggregate *components = right->getAsAggregate();
 658
 659                                 if(components)
 660                                 {
 661                                         TIntermSequence &sequence = components->getSequence();
 662                                         int component = 0;
 663
 664                                         for(TIntermSequence::iterator sit = sequence.begin(); sit != sequence.end(); sit++)
 665                                         {
 666                                                 TIntermConstantUnion *element = (*sit)->getAsConstantUnion();
 667
 668                                                 if(element)
 669                                                 {
 670                                                         int i = element->getUnionArrayPointer()[0].getIConst();
 671                                                         swizzle |= i << (component * 2);
 672                                                         component++;
 673                                                 }
 674                                                 else UNREACHABLE(0);
 675                                         }
 676                                 }
 677                                 else UNREACHABLE(0);
 678
 679                                 Instruction *mov = emit(sw::Shader::OPCODE_MOV, result, left);
 680                                 mov->src[0].swizzle = swizzle;
 681                         }
 682                         break;
 683                 case EOpAddAssign: if(visit == PostVisit) emitAssign(getOpcode(sw::Shader::OPCODE_ADD, result), result, left, left, right); break;
 684                 case EOpAdd:       if(visit == PostVisit) emitBinary(getOpcode(sw::Shader::OPCODE_ADD, result), result, left, right);       break;
 685                 case EOpSubAssign: if(visit == PostVisit) emitAssign(getOpcode(sw::Shader::OPCODE_SUB, result), result, left, left, right); break;
 686                 case EOpSub:       if(visit == PostVisit) emitBinary(getOpcode(sw::Shader::OPCODE_SUB, result), result, left, right);       break;
 687                 case EOpMulAssign: if(visit == PostVisit) emitAssign(getOpcode(sw::Shader::OPCODE_MUL, result), result, left, left, right); break;
 688                 case EOpMul:       if(visit == PostVisit) emitBinary(getOpcode(sw::Shader::OPCODE_MUL, result), result, left, right);       break;
 689                 case EOpDivAssign: if(visit == PostVisit) emitAssign(getOpcode(sw::Shader::OPCODE_DIV, result), result, left, left, right); break;
 690                 case EOpDiv:       if(visit == PostVisit) emitBinary(getOpcode(sw::Shader::OPCODE_DIV, result), result, left, right);       break;
 691                 case EOpIModAssign:          if(visit == PostVisit) emitAssign(getOpcode(sw::Shader::OPCODE_IMOD, result), result, left, left, right); break;
 692                 case EOpIMod:                if(visit == PostVisit) emitBinary(getOpcode(sw::Shader::OPCODE_IMOD, result), result, left, right);       break;
 693                 case EOpBitShiftLeftAssign:  if(visit == PostVisit) emitAssign(sw::Shader::OPCODE_SHL, result, left, left, right); break;
 694                 case EOpBitShiftLeft:        if(visit == PostVisit) emitBinary(sw::Shader::OPCODE_SHL, result, left, right);       break;
 695                 case EOpBitShiftRightAssign: if(visit == PostVisit) emitAssign(getOpcode(sw::Shader::OPCODE_ISHR, result), result, left, left, right); break;
 696                 case EOpBitShiftRight:       if(visit == PostVisit) emitBinary(getOpcode(sw::Shader::OPCODE_ISHR, result), result, left, right);       break;
 697                 case EOpBitwiseAndAssign:    if(visit == PostVisit) emitAssign(sw::Shader::OPCODE_AND, result, left, left, right); break;
 698                 case EOpBitwiseAnd:          if(visit == PostVisit) emitBinary(sw::Shader::OPCODE_AND, result, left, right);       break;
 699                 case EOpBitwiseXorAssign:    if(visit == PostVisit) emitAssign(sw::Shader::OPCODE_XOR, result, left, left, right); break;
 700                 case EOpBitwiseXor:          if(visit == PostVisit) emitBinary(sw::Shader::OPCODE_XOR, result, left, right);       break;
 701                 case EOpBitwiseOrAssign:     if(visit == PostVisit) emitAssign(sw::Shader::OPCODE_OR, result, left, left, right);  break;
 702                 case EOpBitwiseOr:           if(visit == PostVisit) emitBinary(sw::Shader::OPCODE_OR, result, left, right);        break;
 703                 case EOpEqual:
 704                         if(visit == PostVisit)
 705                         {
 706                                 emitBinary(sw::Shader::OPCODE_EQ, result, left, right);
 707
 708                                 for(int index = 1; index < left->totalRegisterCount(); index++)
 709                                 {
 710                                         Temporary equal(this);
 711                                         emit(sw::Shader::OPCODE_EQ, &equal, 0, left, index, right, index);
 712                                         emit(sw::Shader::OPCODE_AND, result, result, &equal);
 713                                 }
 714                         }
 715                         break;
 716                 case EOpNotEqual:
 717                         if(visit == PostVisit)
 718                         {
 719                                 emitBinary(sw::Shader::OPCODE_NE, result, left, right);
 720
 721                                 for(int index = 1; index < left->totalRegisterCount(); index++)
 722                                 {
 723                                         Temporary notEqual(this);
 724                                         emit(sw::Shader::OPCODE_NE, &notEqual, 0, left, index, right, index);
 725                                         emit(sw::Shader::OPCODE_OR, result, result, &notEqual);
 726                                 }
 727                         }
 728                         break;
 729                 case EOpLessThan:                if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_LT, result, left, right); break;
 730                 case EOpGreaterThan:             if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_GT, result, left, right); break;
 731                 case EOpLessThanEqual:           if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_LE, result, left, right); break;
 732                 case EOpGreaterThanEqual:        if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_GE, result, left, right); break;
 733                 case EOpVectorTimesScalarAssign: if(visit == PostVisit) emitAssign(getOpcode(sw::Shader::OPCODE_MUL, left), result, left, left, right); break;
 734                 case EOpVectorTimesScalar:       if(visit == PostVisit) emit(getOpcode(sw::Shader::OPCODE_MUL, left), result, left, right); break;
 735                 case EOpMatrixTimesScalar:
 736                         if(visit == PostVisit)
 737                         {
 738                                 if(left->isMatrix())
 739                                 {
 740                                         for(int i = 0; i < leftType.getNominalSize(); i++)
 741                                         {
 742                                                 emit(sw::Shader::OPCODE_MUL, result, i, left, i, right, 0);
 743                                         }
 744                                 }
 745                                 else if(right->isMatrix())
 746                                 {
 747                                         for(int i = 0; i < rightType.getNominalSize(); i++)
 748                                         {
 749                                                 emit(sw::Shader::OPCODE_MUL, result, i, left, 0, right, i);
 750                                         }
 751                                 }
 752                                 else UNREACHABLE(0);
 753                         }
 754                         break;
 755                 case EOpVectorTimesMatrix:
 756                         if(visit == PostVisit)
 757                         {
 758                                 sw::Shader::Opcode dpOpcode = sw::Shader::OPCODE_DP(leftType.getNominalSize());
 759
 760                                 int size = rightType.getNominalSize();
 761                                 for(int i = 0; i < size; i++)
 762                                 {
 763                                         Instruction *dot = emit(dpOpcode, result, 0, left, 0, right, i);
 764                                         dot->dst.mask = 1 << i;
 765                                 }
 766                         }
 767                         break;
 768                 case EOpMatrixTimesVector:
 769                         if(visit == PostVisit)
 770                         {
 771                                 Instruction *mul = emit(sw::Shader::OPCODE_MUL, result, left, right);
 772                                 mul->src[1].swizzle = 0x00;
 773
 774                                 int size = rightType.getNominalSize();
 775                                 for(int i = 1; i < size; i++)
 776                                 {
 777                                         Instruction *mad = emit(sw::Shader::OPCODE_MAD, result, 0, left, i, right, 0, result);
 778                                         mad->src[1].swizzle = i * 0x55;
 779                                 }
 780                         }
 781                         break;
 782                 case EOpMatrixTimesMatrix:
 783                         if(visit == PostVisit)
 784                         {
 785                                 int dim = leftType.getNominalSize();
 786
 787                                 int size = rightType.getNominalSize();
 788                                 for(int i = 0; i < size; i++)
 789                                 {
 790                                         Instruction *mul = emit(sw::Shader::OPCODE_MUL, result, i, left, 0, right, i);
 791                                         mul->src[1].swizzle = 0x00;
 792
 793                                         for(int j = 1; j < dim; j++)
 794                                         {
 795                                                 Instruction *mad = emit(sw::Shader::OPCODE_MAD, result, i, left, j, right, i, result, i);
 796                                                 mad->src[1].swizzle = j * 0x55;
 797                                         }
 798                                 }
 799                         }
 800                         break;
 801                 case EOpLogicalOr:
 802                         if(trivial(right, 6))
 803                         {
 804                                 if(visit == PostVisit)
 805                                 {
 806                                         emit(sw::Shader::OPCODE_OR, result, left, right);
 807                                 }
 808                         }
 809                         else   // Short-circuit evaluation
 810                         {
 811                                 if(visit == InVisit)
 812                                 {
 813                                         emit(sw::Shader::OPCODE_MOV, result, left);
 814                                         Instruction *ifnot = emit(sw::Shader::OPCODE_IF, 0, result);
 815                                         ifnot->src[0].modifier = sw::Shader::MODIFIER_NOT;
 816                                 }
 817                                 else if(visit == PostVisit)
 818                                 {
 819                                         emit(sw::Shader::OPCODE_MOV, result, right);
 820                                         emit(sw::Shader::OPCODE_ENDIF);
 821                                 }
 822                         }
 823                         break;
 824                 case EOpLogicalXor:        if(visit == PostVisit) emit(sw::Shader::OPCODE_XOR, result, left, right); break;
 825                 case EOpLogicalAnd:
 826                         if(trivial(right, 6))
 827                         {
 828                                 if(visit == PostVisit)
 829                                 {
 830                                         emit(sw::Shader::OPCODE_AND, result, left, right);
 831                                 }
 832                         }
 833                         else   // Short-circuit evaluation
 834                         {
 835                                 if(visit == InVisit)
 836                                 {
 837                                         emit(sw::Shader::OPCODE_MOV, result, left);
 838                                         emit(sw::Shader::OPCODE_IF, 0, result);
 839                                 }
 840                                 else if(visit == PostVisit)
 841                                 {
 842                                         emit(sw::Shader::OPCODE_MOV, result, right);
 843                                         emit(sw::Shader::OPCODE_ENDIF);
 844                                 }
 845                         }
 846                         break;
 847                 default: UNREACHABLE(node->getOp());
 848                 }
 849
 850                 return true;
 851         }
 852
 853         void OutputASM::emitDeterminant(TIntermTyped *result, TIntermTyped *arg, int size, int col, int row, int outCol, int outRow)
 854         {
 855                 switch(size)
 856                 {
 857                 case 1: // Used for cofactor computation only
 858                         {
 859                                 // For a 2x2 matrix, the cofactor is simply a transposed move or negate
 860                                 bool isMov = (row == col);
 861                                 sw::Shader::Opcode op = isMov ? sw::Shader::OPCODE_MOV : sw::Shader::OPCODE_NEG;
 862                                 Instruction *mov = emit(op, result, outCol, arg, isMov ? 1 - row : row);
 863                                 mov->src[0].swizzle = 0x55 * (isMov ? 1 - col : col);
 864                                 mov->dst.mask = 1 << outRow;
 865                         }
 866                         break;
 867                 case 2:
 868                         {
 869                                 static const unsigned int swizzle[3] = { 0x99, 0x88, 0x44 }; // xy?? : yzyz, xzxz, xyxy
 870
 871                                 bool isCofactor = (col >= 0) && (row >= 0);
 872                                 int col0 = (isCofactor && (col <= 0)) ? 1 : 0;
 873                                 int col1 = (isCofactor && (col <= 1)) ? 2 : 1;
 874                                 bool negate = isCofactor && ((col & 0x01) ^ (row & 0x01));
 875
 876                                 Instruction *det = emit(sw::Shader::OPCODE_DET2, result, outCol, arg, negate ? col1 : col0, arg, negate ? col0 : col1);
 877                                 det->src[0].swizzle = det->src[1].swizzle = swizzle[isCofactor ? row : 2];
 878                                 det->dst.mask = 1 << outRow;
 879                         }
 880                         break;
 881                 case 3:
 882                         {
 883                                 static const unsigned int swizzle[4] = { 0xF9, 0xF8, 0xF4, 0xE4 }; // xyz? : yzww, xzww, xyww, xyzw
 884
 885                                 bool isCofactor = (col >= 0) && (row >= 0);
 886                                 int col0 = (isCofactor && (col <= 0)) ? 1 : 0;
 887                                 int col1 = (isCofactor && (col <= 1)) ? 2 : 1;
 888                                 int col2 = (isCofactor && (col <= 2)) ? 3 : 2;
 889                                 bool negate = isCofactor && ((col & 0x01) ^ (row & 0x01));
 890
 891                                 Instruction *det = emit(sw::Shader::OPCODE_DET3, result, outCol, arg, col0, arg, negate ? col2 : col1, arg, negate ? col1 : col2);
 892                                 det->src[0].swizzle = det->src[1].swizzle = det->src[2].swizzle = swizzle[isCofactor ? row : 3];
 893                                 det->dst.mask = 1 << outRow;
 894                         }
 895                         break;
 896                 case 4:
 897                         {
 898                                 Instruction *det = emit(sw::Shader::OPCODE_DET4, result, outCol, arg, 0, arg, 1, arg, 2, arg, 3);
 899                                 det->dst.mask = 1 << outRow;
 900                         }
 901                         break;
 902                 default:
 903                         UNREACHABLE(size);
 904                         break;
 905                 }
 906         }
 907
 908         bool OutputASM::visitUnary(Visit visit, TIntermUnary *node)
 909         {
 910                 if(currentScope != emitScope)
 911                 {
 912                         return false;
 913                 }
 914
 915                 TIntermTyped *result = node;
 916                 TIntermTyped *arg = node->getOperand();
 917                 TBasicType basicType = arg->getType().getBasicType();
 918
 919                 union
 920                 {
 921                         float f;
 922                         int i;
 923                 } one_value;
 924
 925                 if(basicType == EbtInt || basicType == EbtUInt)
 926                 {
 927                         one_value.i = 1;
 928                 }
 929                 else
 930                 {
 931                         one_value.f = 1.0f;
 932                 }
 933
 934                 Constant one(one_value.f, one_value.f, one_value.f, one_value.f);
 935                 Constant rad(1.74532925e-2f, 1.74532925e-2f, 1.74532925e-2f, 1.74532925e-2f);
 936                 Constant deg(5.72957795e+1f, 5.72957795e+1f, 5.72957795e+1f, 5.72957795e+1f);
 937
 938                 switch(node->getOp())
 939                 {
 940                 case EOpNegative:
 941                         if(visit == PostVisit)
 942                         {
 943                                 sw::Shader::Opcode negOpcode = getOpcode(sw::Shader::OPCODE_NEG, arg);
 944                                 for(int index = 0; index < arg->totalRegisterCount(); index++)
 945                                 {
 946                                         emit(negOpcode, result, index, arg, index);
 947                                 }
 948                         }
 949                         break;
 950                 case EOpVectorLogicalNot: if(visit == PostVisit) emit(sw::Shader::OPCODE_NOT, result, arg); break;
 951                 case EOpLogicalNot:       if(visit == PostVisit) emit(sw::Shader::OPCODE_NOT, result, arg); break;
 952                 case EOpPostIncrement:
 953                         if(visit == PostVisit)
 954                         {
 955                                 copy(result, arg);
 956
 957                                 sw::Shader::Opcode addOpcode = getOpcode(sw::Shader::OPCODE_ADD, arg);
 958                                 for(int index = 0; index < arg->totalRegisterCount(); index++)
 959                                 {
 960                                         emit(addOpcode, arg, index, arg, index, &one);
 961                                 }
 962
 963                                 assignLvalue(arg, arg);
 964                         }
 965                         break;
 966                 case EOpPostDecrement:
 967                         if(visit == PostVisit)
 968                         {
 969                                 copy(result, arg);
 970
 971                                 sw::Shader::Opcode subOpcode = getOpcode(sw::Shader::OPCODE_SUB, arg);
 972                                 for(int index = 0; index < arg->totalRegisterCount(); index++)
 973                                 {
 974                                         emit(subOpcode, arg, index, arg, index, &one);
 975                                 }
 976
 977                                 assignLvalue(arg, arg);
 978                         }
 979                         break;
 980                 case EOpPreIncrement:
 981                         if(visit == PostVisit)
 982                         {
 983                                 sw::Shader::Opcode addOpcode = getOpcode(sw::Shader::OPCODE_ADD, arg);
 984                                 for(int index = 0; index < arg->totalRegisterCount(); index++)
 985                                 {
 986                                         emit(addOpcode, result, index, arg, index, &one);
 987                                 }
 988
 989                                 assignLvalue(arg, result);
 990                         }
 991                         break;
 992                 case EOpPreDecrement:
 993                         if(visit == PostVisit)
 994                         {
 995                                 sw::Shader::Opcode subOpcode = getOpcode(sw::Shader::OPCODE_SUB, arg);
 996                                 for(int index = 0; index < arg->totalRegisterCount(); index++)
 997                                 {
 998                                         emit(subOpcode, result, index, arg, index, &one);
 999                                 }
1000
1001                                 assignLvalue(arg, result);
1002                         }
1003                         break;
1004                 case EOpRadians:          if(visit == PostVisit) emit(sw::Shader::OPCODE_MUL, result, arg, &rad); break;
1005                 case EOpDegrees:          if(visit == PostVisit) emit(sw::Shader::OPCODE_MUL, result, arg, &deg); break;
1006                 case EOpSin:              if(visit == PostVisit) emit(sw::Shader::OPCODE_SIN, result, arg); break;
1007                 case EOpCos:              if(visit == PostVisit) emit(sw::Shader::OPCODE_COS, result, arg); break;
1008                 case EOpTan:              if(visit == PostVisit) emit(sw::Shader::OPCODE_TAN, result, arg); break;
1009                 case EOpAsin:             if(visit == PostVisit) emit(sw::Shader::OPCODE_ASIN, result, arg); break;
1010                 case EOpAcos:             if(visit == PostVisit) emit(sw::Shader::OPCODE_ACOS, result, arg); break;
1011                 case EOpAtan:             if(visit == PostVisit) emit(sw::Shader::OPCODE_ATAN, result, arg); break;
1012                 case EOpSinh:             if(visit == PostVisit) emit(sw::Shader::OPCODE_SINH, result, arg); break;
1013                 case EOpCosh:             if(visit == PostVisit) emit(sw::Shader::OPCODE_COSH, result, arg); break;
1014                 case EOpTanh:             if(visit == PostVisit) emit(sw::Shader::OPCODE_TANH, result, arg); break;
1015                 case EOpAsinh:            if(visit == PostVisit) emit(sw::Shader::OPCODE_ASINH, result, arg); break;
1016                 case EOpAcosh:            if(visit == PostVisit) emit(sw::Shader::OPCODE_ACOSH, result, arg); break;
1017                 case EOpAtanh:            if(visit == PostVisit) emit(sw::Shader::OPCODE_ATANH, result, arg); break;
1018                 case EOpExp:              if(visit == PostVisit) emit(sw::Shader::OPCODE_EXP, result, arg); break;
1019                 case EOpLog:              if(visit == PostVisit) emit(sw::Shader::OPCODE_LOG, result, arg); break;
1020                 case EOpExp2:             if(visit == PostVisit) emit(sw::Shader::OPCODE_EXP2, result, arg); break;
1021                 case EOpLog2:             if(visit == PostVisit) emit(sw::Shader::OPCODE_LOG2, result, arg); break;
1022                 case EOpSqrt:             if(visit == PostVisit) emit(sw::Shader::OPCODE_SQRT, result, arg); break;
1023                 case EOpInverseSqrt:      if(visit == PostVisit) emit(sw::Shader::OPCODE_RSQ, result, arg); break;
1024                 case EOpAbs:              if(visit == PostVisit) emit(getOpcode(sw::Shader::OPCODE_ABS, result), result, arg); break;
1025                 case EOpSign:             if(visit == PostVisit) emit(getOpcode(sw::Shader::OPCODE_SGN, result), result, arg); break;
1026                 case EOpFloor:            if(visit == PostVisit) emit(sw::Shader::OPCODE_FLOOR, result, arg); break;
1027                 case EOpTrunc:            if(visit == PostVisit) emit(sw::Shader::OPCODE_TRUNC, result, arg); break;
1028                 case EOpRound:            if(visit == PostVisit) emit(sw::Shader::OPCODE_ROUND, result, arg); break;
1029                 case EOpRoundEven:        if(visit == PostVisit) emit(sw::Shader::OPCODE_ROUNDEVEN, result, arg); break;
1030                 case EOpCeil:             if(visit == PostVisit) emit(sw::Shader::OPCODE_CEIL, result, arg, result); break;
1031                 case EOpFract:            if(visit == PostVisit) emit(sw::Shader::OPCODE_FRC, result, arg); break;
1032                 case EOpIsNan:            if(visit == PostVisit) emit(sw::Shader::OPCODE_ISNAN, result, arg); break;
1033                 case EOpIsInf:            if(visit == PostVisit) emit(sw::Shader::OPCODE_ISINF, result, arg); break;
1034                 case EOpLength:           if(visit == PostVisit) emit(sw::Shader::OPCODE_LEN(dim(arg)), result, arg); break;
1035                 case EOpNormalize:        if(visit == PostVisit) emit(sw::Shader::OPCODE_NRM(dim(arg)), result, arg); break;
1036                 case EOpDFdx:             if(visit == PostVisit) emit(sw::Shader::OPCODE_DFDX, result, arg); break;
1037                 case EOpDFdy:             if(visit == PostVisit) emit(sw::Shader::OPCODE_DFDY, result, arg); break;
1038                 case EOpFwidth:           if(visit == PostVisit) emit(sw::Shader::OPCODE_FWIDTH, result, arg); break;
1039                 case EOpAny:              if(visit == PostVisit) emit(sw::Shader::OPCODE_ANY, result, arg); break;
1040                 case EOpAll:              if(visit == PostVisit) emit(sw::Shader::OPCODE_ALL, result, arg); break;
1041                 case EOpFloatBitsToInt:   if(visit == PostVisit) emit(sw::Shader::OPCODE_FLOATBITSTOINT, result, arg); break;
1042                 case EOpFloatBitsToUint:  if(visit == PostVisit) emit(sw::Shader::OPCODE_FLOATBITSTOUINT, result, arg); break;
1043                 case EOpIntBitsToFloat:   if(visit == PostVisit) emit(sw::Shader::OPCODE_INTBITSTOFLOAT, result, arg); break;
1044                 case EOpUintBitsToFloat:  if(visit == PostVisit) emit(sw::Shader::OPCODE_UINTBITSTOFLOAT, result, arg); break;
1045                 case EOpPackSnorm2x16:    if(visit == PostVisit) emit(sw::Shader::OPCODE_PACKSNORM2x16, result, arg); break;
1046                 case EOpPackUnorm2x16:    if(visit == PostVisit) emit(sw::Shader::OPCODE_PACKUNORM2x16, result, arg); break;
1047                 case EOpPackHalf2x16:     if(visit == PostVisit) emit(sw::Shader::OPCODE_PACKHALF2x16, result, arg); break;
1048                 case EOpUnpackSnorm2x16:  if(visit == PostVisit) emit(sw::Shader::OPCODE_UNPACKSNORM2x16, result, arg); break;
1049                 case EOpUnpackUnorm2x16:  if(visit == PostVisit) emit(sw::Shader::OPCODE_UNPACKUNORM2x16, result, arg); break;
1050                 case EOpUnpackHalf2x16:   if(visit == PostVisit) emit(sw::Shader::OPCODE_UNPACKHALF2x16, result, arg); break;
1051                 case EOpTranspose:
1052                         if(visit == PostVisit)
1053                         {
1054                                 int numCols = arg->getNominalSize();
1055                                 int numRows = arg->getSecondarySize();
1056                                 for(int i = 0; i < numCols; ++i)
1057                                 {
1058                                         for(int j = 0; j < numRows; ++j)
1059                                         {
1060                                                 Instruction *mov = emit(sw::Shader::OPCODE_MOV, result, j, arg, i);
1061                                                 mov->src[0].swizzle = 0x55 * j;
1062                                                 mov->dst.mask = 1 << i;
1063                                         }
1064                                 }
1065                         }
1066                         break;
1067                 case EOpDeterminant:
1068                         if(visit == PostVisit)
1069                         {
1070                                 int size = arg->getNominalSize();
1071                                 ASSERT(size == arg->getSecondarySize());
1072
1073                                 emitDeterminant(result, arg, size);
1074                         }
1075                         break;
1076                 case EOpInverse:
1077                         if(visit == PostVisit)
1078                         {
1079                                 int size = arg->getNominalSize();
1080                                 ASSERT(size == arg->getSecondarySize());
1081
1082                                 // Compute transposed matrix of cofactors
1083                                 for(int i = 0; i < size; ++i)
1084                                 {
1085                                         for(int j = 0; j < size; ++j)
1086                                         {
1087                                                 // For a 2x2 matrix, the cofactor is simply a transposed move or negate
1088                                                 // For a 3x3 or 4x4 matrix, the cofactor is a transposed determinant
1089                                                 emitDeterminant(result, arg, size - 1, j, i, i, j);
1090                                         }
1091                                 }
1092
1093                                 // Compute 1 / determinant
1094                                 Temporary invDet(this);
1095                                 emitDeterminant(&invDet, arg, size);
1096                                 Constant one(1.0f, 1.0f, 1.0f, 1.0f);
1097                                 Instruction *div = emit(sw::Shader::OPCODE_DIV, &invDet, &one, &invDet);
1098                                 div->src[1].swizzle = 0x00; // xxxx
1099
1100                                 // Divide transposed matrix of cofactors by determinant
1101                                 for(int i = 0; i < size; ++i)
1102                                 {
1103                                         emit(sw::Shader::OPCODE_MUL, result, i, result, i, &invDet);
1104                                 }
1105                         }
1106                         break;
1107                 default: UNREACHABLE(node->getOp());
1108                 }
1109
1110                 return true;
1111         }
1112
1113         bool OutputASM::visitAggregate(Visit visit, TIntermAggregate *node)
1114         {
1115                 if(currentScope != emitScope && node->getOp() != EOpFunction && node->getOp() != EOpSequence)
1116                 {
1117                         return false;
1118                 }
1119
1120                 Constant zero(0.0f, 0.0f, 0.0f, 0.0f);
1121
1122                 TIntermTyped *result = node;
1123                 const TType &resultType = node->getType();
1124                 TIntermSequence &arg = node->getSequence();
1125                 size_t argumentCount = arg.size();
1126
1127                 switch(node->getOp())
1128                 {
1129                 case EOpSequence:             break;
1130                 case EOpDeclaration:          break;
1131                 case EOpInvariantDeclaration: break;
1132                 case EOpPrototype:            break;
1133                 case EOpComma:
1134                         if(visit == PostVisit)
1135                         {
1136                                 copy(result, arg[1]);
1137                         }
1138                         break;
1139                 case EOpFunction:
1140                         if(visit == PreVisit)
1141                         {
1142                                 const TString &name = node->getName();
1143
1144                                 if(emitScope == FUNCTION)
1145                                 {
1146                                         if(functionArray.size() > 1)   // No need for a label when there's only main()
1147                                         {
1148                                                 Instruction *label = emit(sw::Shader::OPCODE_LABEL);
1149                                                 label->dst.type = sw::Shader::PARAMETER_LABEL;
1150
1151                                                 const Function *function = findFunction(name);
1152                                                 ASSERT(function);   // Should have been added during global pass
1153                                                 label->dst.index = function->label;
1154                                                 currentFunction = function->label;
1155                                         }
1156                                 }
1157                                 else if(emitScope == GLOBAL)
1158                                 {
1159                                         if(name != "main(")
1160                                         {
1161                                                 TIntermSequence &arguments = node->getSequence()[0]->getAsAggregate()->getSequence();
1162                                                 functionArray.push_back(Function(functionArray.size(), name, &arguments, node));
1163                                         }
1164                                 }
1165                                 else UNREACHABLE(emitScope);
1166
1167                                 currentScope = FUNCTION;
1168                         }
1169                         else if(visit == PostVisit)
1170                         {
1171                                 if(emitScope == FUNCTION)
1172                                 {
1173                                         if(functionArray.size() > 1)   // No need to return when there's only main()
1174                                         {
1175                                                 emit(sw::Shader::OPCODE_RET);
1176                                         }
1177                                 }
1178
1179                                 currentScope = GLOBAL;
1180                         }
1181                         break;
1182                 case EOpFunctionCall:
1183                         if(visit == PostVisit)
1184                         {
1185                                 if(node->isUserDefined())
1186                                 {
1187                                         const TString &name = node->getName();
1188                                         const Function *function = findFunction(name);
1189
1190                                         if(!function)
1191                                         {
1192                                                 mContext.error(node->getLine(), "function definition not found", name.c_str());
1193                                                 return false;
1194                                         }
1195
1196                                         TIntermSequence &arguments = *function->arg;
1197
1198                                         for(size_t i = 0; i < argumentCount; i++)
1199                                         {
1200                                                 TIntermTyped *in = arguments[i]->getAsTyped();
1201
1202                                                 if(in->getQualifier() == EvqIn ||
1203                                                    in->getQualifier() == EvqInOut ||
1204                                                    in->getQualifier() == EvqConstReadOnly)
1205                                                 {
1206                                                         copy(in, arg[i]);
1207                                                 }
1208                                         }
1209
1210                                         Instruction *call = emit(sw::Shader::OPCODE_CALL);
1211                                         call->dst.type = sw::Shader::PARAMETER_LABEL;
1212                                         call->dst.index = function->label;
1213
1214                                         if(function->ret && function->ret->getType().getBasicType() != EbtVoid)
1215                                         {
1216                                                 copy(result, function->ret);
1217                                         }
1218
1219                                         for(size_t i = 0; i < argumentCount; i++)
1220                                         {
1221                                                 TIntermTyped *argument = arguments[i]->getAsTyped();
1222                                                 TIntermTyped *out = arg[i]->getAsTyped();
1223
1224                                                 if(argument->getQualifier() == EvqOut ||
1225                                                    argument->getQualifier() == EvqInOut)
1226                                                 {
1227                                                         copy(out, argument);
1228                                                 }
1229                                         }
1230                                 }
1231                                 else
1232                                 {
1233                                         const TextureFunction textureFunction(node->getName());
1234                                         TIntermTyped *t = arg[1]->getAsTyped();
1235
1236                                         Temporary coord(this);
1237
1238                                         if(textureFunction.proj)
1239                                         {
1240                                                 TIntermConstantUnion* constant = arg[1]->getAsConstantUnion();
1241                                                 if(constant)
1242                                                 {
1243                                                         float projFactor = 1.0f / constant->getFConst(t->getNominalSize() - 1);
1244                                                         Constant projCoord(constant->getFConst(0) * projFactor,
1245                                                                            constant->getFConst(1) * projFactor,
1246                                                                            constant->getFConst(2) * projFactor,
1247                                                                            0.0f);
1248                                                         emit(sw::Shader::OPCODE_MOV, &coord, &projCoord);
1249                                                 }
1250                                                 else
1251                                                 {
1252                                                         Instruction *rcp = emit(sw::Shader::OPCODE_RCPX, &coord, arg[1]);
1253                                                         rcp->src[0].swizzle = 0x55 * (t->getNominalSize() - 1);
1254                                                         rcp->dst.mask = 0x7;
1255
1256                                                         Instruction *mul = emit(sw::Shader::OPCODE_MUL, &coord, arg[1], &coord);
1257                                                         mul->dst.mask = 0x7;
1258                                                 }
1259                                         }
1260                                         else
1261                                         {
1262                                                 emit(sw::Shader::OPCODE_MOV, &coord, arg[1]);
1263                                         }
1264
1265                                         switch(textureFunction.method)
1266                                         {
1267                                         case TextureFunction::IMPLICIT:
1268                                                 {
1269                                                         TIntermNode* offset = textureFunction.offset ? arg[2] : 0;
1270
1271                                                         if(argumentCount == 2 || (textureFunction.offset && argumentCount == 3))
1272                                                         {
1273                                                                 Instruction *tex = emit(textureFunction.offset ? sw::Shader::OPCODE_TEXOFFSET : sw::Shader::OPCODE_TEX,
1274                                                                                         result, &coord, arg[0], offset);
1275                                                         }
1276                                                         else if(argumentCount == 3 || (textureFunction.offset && argumentCount == 4))   // bias
1277                                                         {
1278                                                                 Instruction *bias = emit(sw::Shader::OPCODE_MOV, &coord, arg[textureFunction.offset ? 3 : 2]);
1279                                                                 bias->dst.mask = 0x8;
1280
1281                                                                 Instruction *tex = emit(textureFunction.offset ? sw::Shader::OPCODE_TEXOFFSET : sw::Shader::OPCODE_TEX,
1282                                                                                         result, &coord, arg[0], offset); // FIXME: Implement an efficient TEXLDB instruction
1283                                                                 tex->bias = true;
1284                                                         }
1285                                                         else UNREACHABLE(argumentCount);
1286                                                 }
1287                                                 break;
1288                                         case TextureFunction::LOD:
1289                                                 {
1290                                                         Instruction *lod = emit(sw::Shader::OPCODE_MOV, &coord, arg[2]);
1291                                                         lod->dst.mask = 0x8;
1292
1293                                                         emit(textureFunction.offset ? sw::Shader::OPCODE_TEXLDLOFFSET : sw::Shader::OPCODE_TEXLDL,
1294                                                              result, &coord, arg[0], textureFunction.offset ? arg[3] : nullptr);
1295                                                 }
1296                                                 break;
1297                                         case TextureFunction::FETCH:
1298                                                 {
1299                                                         if(argumentCount == 3 || (textureFunction.offset && argumentCount == 4))
1300                                                         {
1301                                                                 TIntermNode *offset = textureFunction.offset ? arg[3] : nullptr;
1302
1303                                                                 emit(textureFunction.offset ? sw::Shader::OPCODE_TEXELFETCHOFFSET : sw::Shader::OPCODE_TEXELFETCH,
1304                                                                      result, arg[1], arg[0], arg[2], offset);
1305                                                         }
1306                                                         else UNREACHABLE(argumentCount);
1307                                                 }
1308                                                 break;
1309                                         case TextureFunction::GRAD:
1310                                                 {
1311                                                         if(argumentCount == 4 || (textureFunction.offset && argumentCount == 5))
1312                                                         {
1313                                                                 TIntermNode *offset = textureFunction.offset ? arg[4] : nullptr;
1314
1315                                                                 emit(textureFunction.offset ? sw::Shader::OPCODE_TEXGRADOFFSET : sw::Shader::OPCODE_TEXGRAD,
1316                                                                      result, &coord, arg[0], arg[2], arg[3], offset);
1317                                                         }
1318                                                         else UNREACHABLE(argumentCount);
1319                                                 }
1320                                                 break;
1321                                         case TextureFunction::SIZE:
1322                                                 emit(sw::Shader::OPCODE_TEXSIZE, result, arg[1], arg[0]);
1323                                                 break;
1324                                         default:
1325                                                 UNREACHABLE(textureFunction.method);
1326                                         }
1327                                 }
1328                         }
1329                         break;
1330                 case EOpParameters:
1331                         break;
1332                 case EOpConstructFloat:
1333                 case EOpConstructVec2:
1334                 case EOpConstructVec3:
1335                 case EOpConstructVec4:
1336                 case EOpConstructBool:
1337                 case EOpConstructBVec2:
1338                 case EOpConstructBVec3:
1339                 case EOpConstructBVec4:
1340                 case EOpConstructInt:
1341                 case EOpConstructIVec2:
1342                 case EOpConstructIVec3:
1343                 case EOpConstructIVec4:
1344                 case EOpConstructUInt:
1345                 case EOpConstructUVec2:
1346                 case EOpConstructUVec3:
1347                 case EOpConstructUVec4:
1348                         if(visit == PostVisit)
1349                         {
1350                                 int component = 0;
1351
1352                                 for(size_t i = 0; i < argumentCount; i++)
1353                                 {
1354                                         TIntermTyped *argi = arg[i]->getAsTyped();
1355                                         int size = argi->getNominalSize();
1356
1357                                         if(!argi->isMatrix())
1358                                         {
1359                                                 Instruction *mov = emitCast(result, argi);
1360                                                 mov->dst.mask = (0xF << component) & 0xF;
1361                                                 mov->src[0].swizzle = readSwizzle(argi, size) << (component * 2);
1362
1363                                                 component += size;
1364                                         }
1365                                         else   // Matrix
1366                                         {
1367                                                 int column = 0;
1368
1369                                                 while(component < resultType.getNominalSize())
1370                                                 {
1371                                                         Instruction *mov = emitCast(result, 0, argi, column);
1372                                                         mov->dst.mask = (0xF << component) & 0xF;
1373                                                         mov->src[0].swizzle = readSwizzle(argi, size) << (component * 2);
1374
1375                                                         column++;
1376                                                         component += size;
1377                                                 }
1378                                         }
1379                                 }
1380                         }
1381                         break;
1382                 case EOpConstructMat2:
1383                 case EOpConstructMat2x3:
1384                 case EOpConstructMat2x4:
1385                 case EOpConstructMat3x2:
1386                 case EOpConstructMat3:
1387                 case EOpConstructMat3x4:
1388                 case EOpConstructMat4x2:
1389                 case EOpConstructMat4x3:
1390                 case EOpConstructMat4:
1391                         if(visit == PostVisit)
1392                         {
1393                                 TIntermTyped *arg0 = arg[0]->getAsTyped();
1394                                 const int outCols = result->getNominalSize();
1395                                 const int outRows = result->getSecondarySize();
1396
1397                                 if(arg0->isScalar() && arg.size() == 1)   // Construct scale matrix
1398                                 {
1399                                         for(int i = 0; i < outCols; i++)
1400                                         {
1401                                                 Instruction *init = emit(sw::Shader::OPCODE_MOV, result, i, &zero);
1402                                                 Instruction *mov = emitCast(result, i, arg0, 0);
1403                                                 mov->dst.mask = 1 << i;
1404                                                 ASSERT(mov->src[0].swizzle == 0x00);
1405                                         }
1406                                 }
1407                                 else if(arg0->isMatrix())
1408                                 {
1409                                         const int inCols = arg0->getNominalSize();
1410                                         const int inRows = arg0->getSecondarySize();
1411
1412                                         for(int i = 0; i < outCols; i++)
1413                                         {
1414                                                 if(i >= inCols || outRows > inRows)
1415                                                 {
1416                                                         // Initialize to identity matrix
1417                                                         Constant col((i == 0 ? 1.0f : 0.0f), (i == 1 ? 1.0f : 0.0f), (i == 2 ? 1.0f : 0.0f), (i == 3 ? 1.0f : 0.0f));
1418                                                         Instruction *mov = emitCast(result, i, &col, 0);
1419                                                 }
1420
1421                                                 if(i < inCols)
1422                                                 {
1423                                                         Instruction *mov = emitCast(result, i, arg0, i);
1424                                                         mov->dst.mask = 0xF >> (4 - inRows);
1425                                                 }
1426                                         }
1427                                 }
1428                                 else
1429                                 {
1430                                         int column = 0;
1431                                         int row = 0;
1432
1433                                         for(size_t i = 0; i < argumentCount; i++)
1434                                         {
1435                                                 TIntermTyped *argi = arg[i]->getAsTyped();
1436                                                 int size = argi->getNominalSize();
1437                                                 int element = 0;
1438
1439                                                 while(element < size)
1440                                                 {
1441                                                         Instruction *mov = emitCast(result, column, argi, 0);
1442                                                         mov->dst.mask = (0xF << row) & 0xF;
1443                                                         mov->src[0].swizzle = (readSwizzle(argi, size) << (row * 2)) + 0x55 * element;
1444
1445                                                         int end = row + size - element;
1446                                                         column = end >= outRows ? column + 1 : column;
1447                                                         element = element + outRows - row;
1448                                                         row = end >= outRows ? 0 : end;
1449                                                 }
1450                                         }
1451                                 }
1452                         }
1453                         break;
1454                 case EOpConstructStruct:
1455                         if(visit == PostVisit)
1456                         {
1457                                 int offset = 0;
1458                                 for(size_t i = 0; i < argumentCount; i++)
1459                                 {
1460                                         TIntermTyped *argi = arg[i]->getAsTyped();
1461                                         int size = argi->totalRegisterCount();
1462
1463                                         for(int index = 0; index < size; index++)
1464                                         {
1465                                                 Instruction *mov = emit(sw::Shader::OPCODE_MOV, result, index + offset, argi, index);
1466                                                 mov->dst.mask = writeMask(result, offset + index);
1467                                         }
1468
1469                                         offset += size;
1470                                 }
1471                         }
1472                         break;
1473                 case EOpLessThan:         if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_LT, result, arg[0], arg[1]); break;
1474                 case EOpGreaterThan:      if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_GT, result, arg[0], arg[1]); break;
1475                 case EOpLessThanEqual:    if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_LE, result, arg[0], arg[1]); break;
1476                 case EOpGreaterThanEqual: if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_GE, result, arg[0], arg[1]); break;
1477                 case EOpVectorEqual:      if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_EQ, result, arg[0], arg[1]); break;
1478                 case EOpVectorNotEqual:   if(visit == PostVisit) emitCmp(sw::Shader::CONTROL_NE, result, arg[0], arg[1]); break;
1479                 case EOpMod:              if(visit == PostVisit) emit(sw::Shader::OPCODE_MOD, result, arg[0], arg[1]); break;
1480                 case EOpModf:
1481                         if(visit == PostVisit)
1482                         {
1483                                 TIntermTyped* arg1 = arg[1]->getAsTyped();
1484                                 emit(sw::Shader::OPCODE_TRUNC, arg1, arg[0]);
1485                                 assignLvalue(arg1, arg1);
1486                                 emitBinary(sw::Shader::OPCODE_SUB, result, arg[0], arg1);
1487                         }
1488                         break;
1489                 case EOpPow:              if(visit == PostVisit) emit(sw::Shader::OPCODE_POW, result, arg[0], arg[1]); break;
1490                 case EOpAtan:             if(visit == PostVisit) emit(sw::Shader::OPCODE_ATAN2, result, arg[0], arg[1]); break;
1491                 case EOpMin:              if(visit == PostVisit) emit(getOpcode(sw::Shader::OPCODE_MIN, result), result, arg[0], arg[1]); break;
1492                 case EOpMax:              if(visit == PostVisit) emit(getOpcode(sw::Shader::OPCODE_MAX, result), result, arg[0], arg[1]); break;
1493                 case EOpClamp:
1494                         if(visit == PostVisit)
1495                         {
1496                                 emit(getOpcode(sw::Shader::OPCODE_MAX, result), result, arg[0], arg[1]);
1497                                 emit(getOpcode(sw::Shader::OPCODE_MIN, result), result, result, arg[2]);
1498                         }
1499                         break;
1500                 case EOpMix:         if(visit == PostVisit) emit(sw::Shader::OPCODE_LRP, result, arg[2], arg[1], arg[0]); break;
1501                 case EOpStep:        if(visit == PostVisit) emit(sw::Shader::OPCODE_STEP, result, arg[0], arg[1]); break;
1502                 case EOpSmoothStep:  if(visit == PostVisit) emit(sw::Shader::OPCODE_SMOOTH, result, arg[0], arg[1], arg[2]); break;
1503                 case EOpDistance:    if(visit == PostVisit) emit(sw::Shader::OPCODE_DIST(dim(arg[0])), result, arg[0], arg[1]); break;
1504                 case EOpDot:         if(visit == PostVisit) emit(sw::Shader::OPCODE_DP(dim(arg[0])), result, arg[0], arg[1]); break;
1505                 case EOpCross:       if(visit == PostVisit) emit(sw::Shader::OPCODE_CRS, result, arg[0], arg[1]); break;
1506                 case EOpFaceForward: if(visit == PostVisit) emit(sw::Shader::OPCODE_FORWARD(dim(arg[0])), result, arg[0], arg[1], arg[2]); break;
1507                 case EOpReflect:     if(visit == PostVisit) emit(sw::Shader::OPCODE_REFLECT(dim(arg[0])), result, arg[0], arg[1]); break;
1508                 case EOpRefract:     if(visit == PostVisit) emit(sw::Shader::OPCODE_REFRACT(dim(arg[0])), result, arg[0], arg[1], arg[2]); break;
1509                 case EOpMul:
1510                         if(visit == PostVisit)
1511                         {
1512                                 TIntermTyped *arg0 = arg[0]->getAsTyped();
1513                                 TIntermTyped *arg1 = arg[1]->getAsTyped();
1514                                 ASSERT((arg0->getNominalSize() == arg1->getNominalSize()) && (arg0->getSecondarySize() == arg1->getSecondarySize()));
1515
1516                                 int size = arg0->getNominalSize();
1517                                 for(int i = 0; i < size; i++)
1518                                 {
1519                                         emit(sw::Shader::OPCODE_MUL, result, i, arg[0], i, arg[1], i);
1520                                 }
1521                         }
1522                         break;
1523                 case EOpOuterProduct:
1524                         if(visit == PostVisit)
1525                         {
1526                                 for(int i = 0; i < dim(arg[1]); i++)
1527                                 {
1528                                         Instruction *mul = emit(sw::Shader::OPCODE_MUL, result, i, arg[0], 0, arg[1]);
1529                                         mul->src[1].swizzle = 0x55 * i;
1530                                 }
1531                         }
1532                         break;
1533                 default: UNREACHABLE(node->getOp());
1534                 }
1535
1536                 return true;
1537         }
1538
1539         bool OutputASM::visitSelection(Visit visit, TIntermSelection *node)
1540         {
1541                 if(currentScope != emitScope)
1542                 {
1543                         return false;
1544                 }
1545
1546                 TIntermTyped *condition = node->getCondition();
1547                 TIntermNode *trueBlock = node->getTrueBlock();
1548                 TIntermNode *falseBlock = node->getFalseBlock();
1549                 TIntermConstantUnion *constantCondition = condition->getAsConstantUnion();
1550
1551                 condition->traverse(this);
1552
1553                 if(node->usesTernaryOperator())
1554                 {
1555                         if(constantCondition)
1556                         {
1557                                 bool trueCondition = constantCondition->getUnionArrayPointer()->getBConst();
1558
1559                                 if(trueCondition)
1560                                 {
1561                                         trueBlock->traverse(this);
1562                                         copy(node, trueBlock);
1563                                 }
1564                                 else
1565                                 {
1566                                         falseBlock->traverse(this);
1567                                         copy(node, falseBlock);
1568                                 }
1569                         }
1570                         else if(trivial(node, 6))   // Fast to compute both potential results and no side effects
1571                         {
1572                                 trueBlock->traverse(this);
1573                                 falseBlock->traverse(this);
1574                                 emit(sw::Shader::OPCODE_SELECT, node, condition, trueBlock, falseBlock);
1575                         }
1576                         else
1577                         {
1578                                 emit(sw::Shader::OPCODE_IF, 0, condition);
1579
1580                                 if(trueBlock)
1581                                 {
1582                                         trueBlock->traverse(this);
1583                                         copy(node, trueBlock);
1584                                 }
1585
1586                                 if(falseBlock)
1587                                 {
1588                                         emit(sw::Shader::OPCODE_ELSE);
1589                                         falseBlock->traverse(this);
1590                                         copy(node, falseBlock);
1591                                 }
1592
1593                                 emit(sw::Shader::OPCODE_ENDIF);
1594                         }
1595                 }
1596                 else  // if/else statement
1597                 {
1598                         if(constantCondition)
1599                         {
1600                                 bool trueCondition = constantCondition->getUnionArrayPointer()->getBConst();
1601
1602                                 if(trueCondition)
1603                                 {
1604                                         if(trueBlock)
1605                                         {
1606                                                 trueBlock->traverse(this);
1607                                         }
1608                                 }
1609                                 else
1610                                 {
1611                                         if(falseBlock)
1612                                         {
1613                                                 falseBlock->traverse(this);
1614                                         }
1615                                 }
1616                         }
1617                         else
1618                         {
1619                                 emit(sw::Shader::OPCODE_IF, 0, condition);
1620
1621                                 if(trueBlock)
1622                                 {
1623                                         trueBlock->traverse(this);
1624                                 }
1625
1626                                 if(falseBlock)
1627                                 {
1628                                         emit(sw::Shader::OPCODE_ELSE);
1629                                         falseBlock->traverse(this);
1630                                 }
1631
1632                                 emit(sw::Shader::OPCODE_ENDIF);
1633                         }
1634                 }
1635
1636                 return false;
1637         }
1638
1639         bool OutputASM::visitLoop(Visit visit, TIntermLoop *node)
1640         {
1641                 if(currentScope != emitScope)
1642                 {
1643                         return false;
1644                 }
1645
1646                 unsigned int iterations = loopCount(node);
1647
1648                 if(iterations == 0)
1649                 {
1650                         return false;
1651                 }
1652
1653                 bool unroll = (iterations <= 4);
1654
1655                 if(unroll)
1656                 {
1657                         LoopUnrollable loopUnrollable;
1658                         unroll = loopUnrollable.traverse(node);
1659                 }
1660
1661                 TIntermNode *init = node->getInit();
1662                 TIntermTyped *condition = node->getCondition();
1663                 TIntermTyped *expression = node->getExpression();
1664                 TIntermNode *body = node->getBody();
1665                 Constant True(true);
1666
1667                 if(node->getType() == ELoopDoWhile)
1668                 {
1669                         Temporary iterate(this);
1670                         emit(sw::Shader::OPCODE_MOV, &iterate, &True);
1671
1672                         emit(sw::Shader::OPCODE_WHILE, 0, &iterate);   // FIXME: Implement real do-while
1673
1674                         if(body)
1675                         {
1676                                 body->traverse(this);
1677                         }
1678
1679                         emit(sw::Shader::OPCODE_TEST);
1680
1681                         condition->traverse(this);
1682                         emit(sw::Shader::OPCODE_MOV, &iterate, condition);
1683
1684                         emit(sw::Shader::OPCODE_ENDWHILE);
1685                 }
1686                 else
1687                 {
1688                         if(init)
1689                         {
1690                                 init->traverse(this);
1691                         }
1692
1693                         if(unroll)
1694                         {
1695                                 for(unsigned int i = 0; i < iterations; i++)
1696                                 {
1697                                 //      condition->traverse(this);   // Condition could contain statements, but not in an unrollable loop
1698
1699                                         if(body)
1700                                         {
1701                                                 body->traverse(this);
1702                                         }
1703
1704                                         if(expression)
1705                                         {
1706                                                 expression->traverse(this);
1707                                         }
1708                                 }
1709                         }
1710                         else
1711                         {
1712                                 if(condition)
1713                                 {
1714                                         condition->traverse(this);
1715                                 }
1716                                 else
1717                                 {
1718                                         condition = &True;
1719                                 }
1720
1721                                 emit(sw::Shader::OPCODE_WHILE, 0, condition);
1722
1723                                 if(body)
1724                                 {
1725                                         body->traverse(this);
1726                                 }
1727
1728                                 emit(sw::Shader::OPCODE_TEST);
1729
1730                                 if(expression)
1731                                 {
1732                                         expression->traverse(this);
1733                                 }
1734
1735                                 if(condition)
1736                                 {
1737                                         condition->traverse(this);
1738                                 }
1739
1740                                 emit(sw::Shader::OPCODE_ENDWHILE);
1741                         }
1742                 }
1743
1744                 return false;
1745         }
1746
1747         bool OutputASM::visitBranch(Visit visit, TIntermBranch *node)
1748         {
1749                 if(currentScope != emitScope)
1750                 {
1751                         return false;
1752                 }
1753
1754                 switch(node->getFlowOp())
1755                 {
1756                 case EOpKill:      if(visit == PostVisit) emit(sw::Shader::OPCODE_DISCARD);  break;
1757                 case EOpBreak:     if(visit == PostVisit) emit(sw::Shader::OPCODE_BREAK);    break;
1758                 case EOpContinue:  if(visit == PostVisit) emit(sw::Shader::OPCODE_CONTINUE); break;
1759                 case EOpReturn:
1760                         if(visit == PostVisit)
1761                         {
1762                                 TIntermTyped *value = node->getExpression();
1763
1764                                 if(value)
1765                                 {
1766                                         copy(functionArray[currentFunction].ret, value);
1767                                 }
1768
1769                                 emit(sw::Shader::OPCODE_LEAVE);
1770                         }
1771                         break;
1772                 default: UNREACHABLE(node->getFlowOp());
1773                 }
1774
1775                 return true;
1776         }
1777
1778         bool OutputASM::visitSwitch(Visit visit, TIntermSwitch *node)
1779         {
1780                 if(currentScope != emitScope)
1781                 {
1782                         return false;
1783                 }
1784
1785                 TIntermTyped* switchValue = node->getInit();
1786                 TIntermAggregate* opList = node->getStatementList();
1787
1788                 if(!switchValue || !opList)
1789                 {
1790                         return false;
1791                 }
1792
1793                 switchValue->traverse(this);
1794
1795                 emit(sw::Shader::OPCODE_SWITCH);
1796
1797                 TIntermSequence& sequence = opList->getSequence();
1798                 TIntermSequence::iterator it = sequence.begin();
1799                 TIntermSequence::iterator defaultIt = sequence.end();
1800                 int nbCases = 0;
1801                 for(; it != sequence.end(); ++it)
1802                 {
1803                         TIntermCase* currentCase = (*it)->getAsCaseNode();
1804                         if(currentCase)
1805                         {
1806                                 TIntermSequence::iterator caseIt = it;
1807
1808                                 TIntermTyped* condition = currentCase->getCondition();
1809                                 if(condition) // non default case
1810                                 {
1811                                         if(nbCases != 0)
1812                                         {
1813                                                 emit(sw::Shader::OPCODE_ELSE);
1814                                         }
1815
1816                                         condition->traverse(this);
1817                                         Temporary result(this);
1818                                         emitBinary(sw::Shader::OPCODE_EQ, &result, switchValue, condition);
1819                                         emit(sw::Shader::OPCODE_IF, 0, &result);
1820                                         nbCases++;
1821
1822                                         for(++caseIt; caseIt != sequence.end(); ++caseIt)
1823                                         {
1824                                                 (*caseIt)->traverse(this);
1825                                                 if((*caseIt)->getAsBranchNode()) // Kill, Break, Continue or Return
1826                                                 {
1827                                                         break;
1828                                                 }
1829                                         }
1830                                 }
1831                                 else
1832                                 {
1833                                         defaultIt = it; // The default case might not be the last case, keep it for last
1834                                 }
1835                         }
1836                 }
1837
1838                 // If there's a default case, traverse it here
1839                 if(defaultIt != sequence.end())
1840                 {
1841                         emit(sw::Shader::OPCODE_ELSE);
1842                         for(++defaultIt; defaultIt != sequence.end(); ++defaultIt)
1843                         {
1844                                 (*defaultIt)->traverse(this);
1845                                 if((*defaultIt)->getAsBranchNode()) // Kill, Break, Continue or Return
1846                                 {
1847                                         break;
1848                                 }
1849                         }
1850                 }
1851
1852                 for(int i = 0; i < nbCases; ++i)
1853                 {
1854                         emit(sw::Shader::OPCODE_ENDIF);
1855                 }
1856
1857                 emit(sw::Shader::OPCODE_ENDSWITCH);
1858
1859                 return false;
1860         }
1861
1862         Instruction *OutputASM::emit(sw::Shader::Opcode op, TIntermTyped *dst, TIntermNode *src0, TIntermNode *src1, TIntermNode *src2, TIntermNode *src3, TIntermNode *src4)
1863         {
1864                 return emit(op, dst, 0, src0, 0, src1, 0, src2, 0, src3, 0, src4, 0);
1865         }
1866
1867         Instruction *OutputASM::emit(sw::Shader::Opcode op, TIntermTyped *dst, int dstIndex, TIntermNode *src0, int index0, TIntermNode *src1, int index1,
1868                                      TIntermNode *src2, int index2, TIntermNode *src3, int index3, TIntermNode *src4, int index4)
1869         {
1870                 Instruction *instruction = new Instruction(op);
1871
1872                 if(dst)
1873                 {
1874                         instruction->dst.type = registerType(dst);
1875                         instruction->dst.index = registerIndex(dst) + dstIndex;
1876                         instruction->dst.mask = writeMask(dst);
1877                         instruction->dst.integer = (dst->getBasicType() == EbtInt);
1878                 }
1879
1880                 argument(instruction->src[0], src0, index0);
1881                 argument(instruction->src[1], src1, index1);
1882                 argument(instruction->src[2], src2, index2);
1883                 argument(instruction->src[3], src3, index3);
1884                 argument(instruction->src[4], src4, index4);
1885
1886                 shader->append(instruction);
1887
1888                 return instruction;
1889         }
1890
1891         Instruction *OutputASM::emitCast(TIntermTyped *dst, TIntermTyped *src)
1892         {
1893                 return emitCast(dst, 0, src, 0);
1894         }
1895
1896         Instruction *OutputASM::emitCast(TIntermTyped *dst, int dstIndex, TIntermTyped *src, int srcIndex)
1897         {
1898                 switch(src->getBasicType())
1899                 {
1900                 case EbtBool:
1901                         switch(dst->getBasicType())
1902                         {
1903                         case EbtInt:   return emit(sw::Shader::OPCODE_B2I, dst, dstIndex, src, srcIndex);
1904                         case EbtUInt:  return emit(sw::Shader::OPCODE_B2I, dst, dstIndex, src, srcIndex);
1905                         case EbtFloat: return emit(sw::Shader::OPCODE_B2F, dst, dstIndex, src, srcIndex);
1906                         default:       break;
1907                         }
1908                         break;
1909                 case EbtInt:
1910                         switch(dst->getBasicType())
1911                         {
1912                         case EbtBool:  return emit(sw::Shader::OPCODE_I2B, dst, dstIndex, src, srcIndex);
1913                         case EbtFloat: return emit(sw::Shader::OPCODE_I2F, dst, dstIndex, src, srcIndex);
1914                         default:       break;
1915                         }
1916                         break;
1917                 case EbtUInt:
1918                         switch(dst->getBasicType())
1919                         {
1920                         case EbtBool:  return emit(sw::Shader::OPCODE_I2B, dst, dstIndex, src, srcIndex);
1921                         case EbtFloat: return emit(sw::Shader::OPCODE_U2F, dst, dstIndex, src, srcIndex);
1922                         default:       break;
1923                         }
1924                         break;
1925                 case EbtFloat:
1926                         switch(dst->getBasicType())
1927                         {
1928                         case EbtBool: return emit(sw::Shader::OPCODE_F2B, dst, dstIndex, src, srcIndex);
1929                         case EbtInt:  return emit(sw::Shader::OPCODE_F2I, dst, dstIndex, src, srcIndex);
1930                         case EbtUInt: return emit(sw::Shader::OPCODE_F2U, dst, dstIndex, src, srcIndex);
1931                         default:      break;
1932                         }
1933                         break;
1934                 default:
1935                         break;
1936                 }
1937
1938                 ASSERT((src->getBasicType() == dst->getBasicType()) ||
1939                       ((src->getBasicType() == EbtInt) && (dst->getBasicType() == EbtUInt)) ||
1940                       ((src->getBasicType() == EbtUInt) && (dst->getBasicType() == EbtInt)));
1941
1942                 return emit(sw::Shader::OPCODE_MOV, dst, dstIndex, src, srcIndex);
1943         }
1944
1945         void OutputASM::emitBinary(sw::Shader::Opcode op, TIntermTyped *dst, TIntermNode *src0, TIntermNode *src1, TIntermNode *src2)
1946         {
1947                 for(int index = 0; index < dst->elementRegisterCount(); index++)
1948                 {
1949                         emit(op, dst, index, src0, index, src1, index, src2, index);
1950                 }
1951         }
1952
1953         void OutputASM::emitAssign(sw::Shader::Opcode op, TIntermTyped *result, TIntermTyped *lhs, TIntermTyped *src0, TIntermTyped *src1)
1954         {
1955                 emitBinary(op, result, src0, src1);
1956                 assignLvalue(lhs, result);
1957         }
1958
1959         void OutputASM::emitCmp(sw::Shader::Control cmpOp, TIntermTyped *dst, TIntermNode *left, TIntermNode *right, int index)
1960         {
1961                 sw::Shader::Opcode opcode;
1962                 switch(left->getAsTyped()->getBasicType())
1963                 {
1964                 case EbtBool:
1965                 case EbtInt:
1966                         opcode = sw::Shader::OPCODE_ICMP;
1967                         break;
1968                 case EbtUInt:
1969                         opcode = sw::Shader::OPCODE_UCMP;
1970                         break;
1971                 default:
1972                         opcode = sw::Shader::OPCODE_CMP;
1973                         break;
1974                 }
1975
1976                 Instruction *cmp = emit(opcode, dst, 0, left, index, right, index);
1977                 cmp->control = cmpOp;
1978         }
1979
1980         int componentCount(const TType &type, int registers)
1981         {
1982                 if(registers == 0)
1983                 {
1984                         return 0;
1985                 }
1986
1987                 if(type.isArray() && registers >= type.elementRegisterCount())
1988                 {
1989                         int index = registers / type.elementRegisterCount();
1990                         registers -= index * type.elementRegisterCount();
1991                         return index * type.getElementSize() + componentCount(type, registers);
1992                 }
1993
1994                 if(type.isStruct() || type.isInterfaceBlock())
1995                 {
1996                         const TFieldList& fields = type.getStruct() ? type.getStruct()->fields() : type.getInterfaceBlock()->fields();
1997                         int elements = 0;
1998
1999                         for(TFieldList::const_iterator field = fields.begin(); field != fields.end(); field++)
2000                         {
2001                                 const TType &fieldType = *((*field)->type());
2002
2003                                 if(fieldType.totalRegisterCount() <= registers)
2004                                 {
2005                                         registers -= fieldType.totalRegisterCount();
2006                                         elements += fieldType.getObjectSize();
2007                                 }
2008                                 else   // Register within this field
2009                                 {
2010                                         return elements + componentCount(fieldType, registers);
2011                                 }
2012                         }
2013                 }
2014                 else if(type.isMatrix())
2015                 {
2016                         return registers * type.registerSize();
2017                 }
2018
2019                 UNREACHABLE(0);
2020                 return 0;
2021         }
2022
2023         int registerSize(const TType &type, int registers)
2024         {
2025                 if(registers == 0)
2026                 {
2027                         if(type.isStruct())
2028                         {
2029                                 return registerSize(*((*(type.getStruct()->fields().begin()))->type()), 0);
2030                         }
2031                         else if(type.isInterfaceBlock())
2032                         {
2033                                 return registerSize(*((*(type.getInterfaceBlock()->fields().begin()))->type()), 0);
2034                         }
2035
2036                         return type.registerSize();
2037                 }
2038
2039                 if(type.isArray() && registers >= type.elementRegisterCount())
2040                 {
2041                         int index = registers / type.elementRegisterCount();
2042                         registers -= index * type.elementRegisterCount();
2043                         return registerSize(type, registers);
2044                 }
2045
2046                 if(type.isStruct() || type.isInterfaceBlock())
2047                 {
2048                         const TFieldList& fields = type.getStruct() ? type.getStruct()->fields() : type.getInterfaceBlock()->fields();
2049                         int elements = 0;
2050
2051                         for(TFieldList::const_iterator field = fields.begin(); field != fields.end(); field++)
2052                         {
2053                                 const TType &fieldType = *((*field)->type());
2054
2055                                 if(fieldType.totalRegisterCount() <= registers)
2056                                 {
2057                                         registers -= fieldType.totalRegisterCount();
2058                                         elements += fieldType.getObjectSize();
2059                                 }
2060                                 else   // Register within this field
2061                                 {
2062                                         return registerSize(fieldType, registers);
2063                                 }
2064                         }
2065                 }
2066                 else if(type.isMatrix())
2067                 {
2068                         return registerSize(type, 0);
2069                 }
2070
2071                 UNREACHABLE(0);
2072                 return 0;
2073         }
2074
2075         int OutputASM::getBlockId(TIntermTyped *arg)
2076         {
2077                 if(arg)
2078                 {
2079                         const TType &type = arg->getType();
2080                         TInterfaceBlock* block = type.getInterfaceBlock();
2081                         if(block && (type.getQualifier() == EvqUniform))
2082                         {
2083                                 // Make sure the uniform block is declared
2084                                 uniformRegister(arg);
2085
2086                                 const char* blockName = block->name().c_str();
2087
2088                                 // Fetch uniform block index from array of blocks
2089                                 for(ActiveUniformBlocks::const_iterator it = shaderObject->activeUniformBlocks.begin(); it != shaderObject->activeUniformBlocks.end(); ++it)
2090                                 {
2091                                         if(blockName == it->name)
2092                                         {
2093                                                 return it->blockId;
2094                                         }
2095                                 }
2096
2097                                 ASSERT(false);
2098                         }
2099                 }
2100
2101                 return -1;
2102         }
2103
2104         OutputASM::ArgumentInfo OutputASM::getArgumentInfo(TIntermTyped *arg, int index)
2105         {
2106                 const TType &type = arg->getType();
2107                 int blockId = getBlockId(arg);
2108                 ArgumentInfo argumentInfo(BlockMemberInfo::getDefaultBlockInfo(), type, -1, -1);
2109                 if(blockId != -1)
2110                 {
2111                         argumentInfo.bufferIndex = 0;
2112                         for(int i = 0; i < blockId; ++i)
2113                         {
2114                                 int blockArraySize = shaderObject->activeUniformBlocks[i].arraySize;
2115                                 argumentInfo.bufferIndex += blockArraySize > 0 ? blockArraySize : 1;
2116                         }
2117
2118                         const BlockDefinitionIndexMap& blockDefinition = blockDefinitions[blockId];
2119
2120                         BlockDefinitionIndexMap::const_iterator itEnd = blockDefinition.end();
2121                         BlockDefinitionIndexMap::const_iterator it = itEnd;
2122
2123                         argumentInfo.clampedIndex = index;
2124                         if(type.isInterfaceBlock())
2125                         {
2126                                 // Offset index to the beginning of the selected instance
2127                                 int blockRegisters = type.elementRegisterCount();
2128                                 int bufferOffset = argumentInfo.clampedIndex / blockRegisters;
2129                                 argumentInfo.bufferIndex += bufferOffset;
2130                                 argumentInfo.clampedIndex -= bufferOffset * blockRegisters;
2131                         }
2132
2133                         int regIndex = registerIndex(arg);
2134                         for(int i = regIndex + argumentInfo.clampedIndex; i >= regIndex; --i)
2135                         {
2136                                 it = blockDefinition.find(i);
2137                                 if(it != itEnd)
2138                                 {
2139                                         argumentInfo.clampedIndex -= (i - regIndex);
2140                                         break;
2141                                 }
2142                         }
2143                         ASSERT(it != itEnd);
2144
2145                         argumentInfo.typedMemberInfo = it->second;
2146
2147                         int registerCount = argumentInfo.typedMemberInfo.type.totalRegisterCount();
2148                         argumentInfo.clampedIndex = (argumentInfo.clampedIndex >= registerCount) ? registerCount - 1 : argumentInfo.clampedIndex;
2149                 }
2150                 else
2151                 {
2152                         argumentInfo.clampedIndex = (index >= arg->totalRegisterCount()) ? arg->totalRegisterCount() - 1 : index;
2153                 }
2154
2155                 return argumentInfo;
2156         }
2157
2158         void OutputASM::argument(sw::Shader::SourceParameter &parameter, TIntermNode *argument, int index)
2159         {
2160                 if(argument)
2161                 {
2162                         TIntermTyped *arg = argument->getAsTyped();
2163                         Temporary unpackedUniform(this);
2164
2165                         const TType& srcType = arg->getType();
2166                         TInterfaceBlock* srcBlock = srcType.getInterfaceBlock();
2167                         if(srcBlock && (srcType.getQualifier() == EvqUniform))
2168                         {
2169                                 const ArgumentInfo argumentInfo = getArgumentInfo(arg, index);
2170                                 const TType &memberType = argumentInfo.typedMemberInfo.type;
2171
2172                                 if(memberType.getBasicType() == EbtBool)
2173                                 {
2174                                         int arraySize = (memberType.isArray() ? memberType.getArraySize() : 1);
2175                                         ASSERT(argumentInfo.clampedIndex < arraySize);
2176
2177                                         // Convert the packed bool, which is currently an int, to a true bool
2178                                         Instruction *instruction = new Instruction(sw::Shader::OPCODE_I2B);
2179                                         instruction->dst.type = sw::Shader::PARAMETER_TEMP;
2180                                         instruction->dst.index = registerIndex(&unpackedUniform);
2181                                         instruction->src[0].type = sw::Shader::PARAMETER_CONST;
2182                                         instruction->src[0].bufferIndex = argumentInfo.bufferIndex;
2183                                         instruction->src[0].index = argumentInfo.typedMemberInfo.offset + argumentInfo.clampedIndex * argumentInfo.typedMemberInfo.arrayStride;
2184
2185                                         shader->append(instruction);
2186
2187                                         arg = &unpackedUniform;
2188                                         index = 0;
2189                                 }
2190                                 else if((srcBlock->matrixPacking() == EmpRowMajor) && memberType.isMatrix())
2191                                 {
2192                                         int numCols = memberType.getNominalSize();
2193                                         int numRows = memberType.getSecondarySize();
2194                                         int arraySize = (memberType.isArray() ? memberType.getArraySize() : 1);
2195
2196                                         ASSERT(argumentInfo.clampedIndex < (numCols * arraySize));
2197
2198                                         unsigned int dstIndex = registerIndex(&unpackedUniform);
2199                                         unsigned int srcSwizzle = (argumentInfo.clampedIndex % numCols) * 0x55;
2200                                         int arrayIndex = argumentInfo.clampedIndex / numCols;
2201                                         int matrixStartOffset = argumentInfo.typedMemberInfo.offset + arrayIndex * argumentInfo.typedMemberInfo.arrayStride;
2202
2203                                         for(int j = 0; j < numRows; ++j)
2204                                         {
2205                                                 // Transpose the row major matrix
2206                                                 Instruction *instruction = new Instruction(sw::Shader::OPCODE_MOV);
2207                                                 instruction->dst.type = sw::Shader::PARAMETER_TEMP;
2208                                                 instruction->dst.index = dstIndex;
2209                                                 instruction->dst.mask = 1 << j;
2210                                                 instruction->src[0].type = sw::Shader::PARAMETER_CONST;
2211                                                 instruction->src[0].bufferIndex = argumentInfo.bufferIndex;
2212                                                 instruction->src[0].index = matrixStartOffset + j * argumentInfo.typedMemberInfo.matrixStride;
2213                                                 instruction->src[0].swizzle = srcSwizzle;
2214
2215                                                 shader->append(instruction);
2216                                         }
2217
2218                                         arg = &unpackedUniform;
2219                                         index = 0;
2220                                 }
2221                         }
2222
2223                         const ArgumentInfo argumentInfo = getArgumentInfo(arg, index);
2224                         const TType &type = argumentInfo.typedMemberInfo.type;
2225
2226                         int size = registerSize(type, argumentInfo.clampedIndex);
2227
2228                         parameter.type = registerType(arg);
2229                         parameter.bufferIndex = argumentInfo.bufferIndex;
2230
2231                         if(arg->getAsConstantUnion() && arg->getAsConstantUnion()->getUnionArrayPointer())
2232                         {
2233                                 int component = componentCount(type, argumentInfo.clampedIndex);
2234                                 ConstantUnion *constants = arg->getAsConstantUnion()->getUnionArrayPointer();
2235
2236                                 for(int i = 0; i < 4; i++)
2237                                 {
2238                                         if(size == 1)   // Replicate
2239                                         {
2240                                                 parameter.value[i] = constants[component + 0].getAsFloat();
2241                                         }
2242                                         else if(i < size)
2243                                         {
2244                                                 parameter.value[i] = constants[component + i].getAsFloat();
2245                                         }
2246                                         else
2247                                         {
2248                                                 parameter.value[i] = 0.0f;
2249                                         }
2250                                 }
2251                         }
2252                         else
2253                         {
2254                                 parameter.index = registerIndex(arg) + argumentInfo.clampedIndex;
2255
2256                                 if(parameter.bufferIndex != -1)
2257                                 {
2258                                         int stride = (argumentInfo.typedMemberInfo.matrixStride > 0) ? argumentInfo.typedMemberInfo.matrixStride : argumentInfo.typedMemberInfo.arrayStride;
2259                                         parameter.index = argumentInfo.typedMemberInfo.offset + argumentInfo.clampedIndex * stride;
2260                                 }
2261                         }
2262
2263                         if(!IsSampler(arg->getBasicType()))
2264                         {
2265                                 parameter.swizzle = readSwizzle(arg, size);
2266                         }
2267                 }
2268         }
2269
2270         void OutputASM::copy(TIntermTyped *dst, TIntermNode *src, int offset)
2271         {
2272                 for(int index = 0; index < dst->totalRegisterCount(); index++)
2273                 {
2274                         Instruction *mov = emit(sw::Shader::OPCODE_MOV, dst, index, src, offset + index);
2275                         mov->dst.mask = writeMask(dst, index);
2276                 }
2277         }
2278
2279         int swizzleElement(int swizzle, int index)
2280         {
2281                 return (swizzle >> (index * 2)) & 0x03;
2282         }
2283
2284         int swizzleSwizzle(int leftSwizzle, int rightSwizzle)
2285         {
2286                 return (swizzleElement(leftSwizzle, swizzleElement(rightSwizzle, 0)) << 0) |
2287                        (swizzleElement(leftSwizzle, swizzleElement(rightSwizzle, 1)) << 2) |
2288                        (swizzleElement(leftSwizzle, swizzleElement(rightSwizzle, 2)) << 4) |
2289                        (swizzleElement(leftSwizzle, swizzleElement(rightSwizzle, 3)) << 6);
2290         }
2291
2292         void OutputASM::assignLvalue(TIntermTyped *dst, TIntermTyped *src)
2293         {
2294                 if(src &&
2295                         ((src->isVector() && (!dst->isVector() || (src->getNominalSize() != dst->getNominalSize()))) ||
2296                          (src->isMatrix() && (!dst->isMatrix() || (src->getNominalSize() != dst->getNominalSize()) || (src->getSecondarySize() != dst->getSecondarySize())))))
2297                 {
2298                         return mContext.error(src->getLine(), "Result type should match the l-value type in compound assignment", src->isVector() ? "vector" : "matrix");
2299                 }
2300
2301                 TIntermBinary *binary = dst->getAsBinaryNode();
2302
2303                 if(binary && binary->getOp() == EOpIndexIndirect && binary->getLeft()->isVector() && dst->isScalar())
2304                 {
2305                         Instruction *insert = new Instruction(sw::Shader::OPCODE_INSERT);
2306
2307                         Temporary address(this);
2308                         lvalue(insert->dst, address, dst);
2309
2310                         insert->src[0].type = insert->dst.type;
2311                         insert->src[0].index = insert->dst.index;
2312                         insert->src[0].rel = insert->dst.rel;
2313                         argument(insert->src[1], src);
2314                         argument(insert->src[2], binary->getRight());
2315
2316                         shader->append(insert);
2317                 }
2318                 else
2319                 {
2320                         for(int offset = 0; offset < dst->totalRegisterCount(); offset++)
2321                         {
2322                                 Instruction *mov = new Instruction(sw::Shader::OPCODE_MOV);
2323
2324                                 Temporary address(this);
2325                                 int swizzle = lvalue(mov->dst, address, dst);
2326                                 mov->dst.index += offset;
2327
2328                                 if(offset > 0)
2329                                 {
2330                                         mov->dst.mask = writeMask(dst, offset);
2331                                 }
2332
2333                                 argument(mov->src[0], src, offset);
2334                                 mov->src[0].swizzle = swizzleSwizzle(mov->src[0].swizzle, swizzle);
2335
2336                                 shader->append(mov);
2337                         }
2338                 }
2339         }
2340
2341         int OutputASM::lvalue(sw::Shader::DestinationParameter &dst, Temporary &address, TIntermTyped *node)
2342         {
2343                 TIntermTyped *result = node;
2344                 TIntermBinary *binary = node->getAsBinaryNode();
2345                 TIntermSymbol *symbol = node->getAsSymbolNode();
2346
2347                 if(binary)
2348                 {
2349                         TIntermTyped *left = binary->getLeft();
2350                         TIntermTyped *right = binary->getRight();
2351
2352                         int leftSwizzle = lvalue(dst, address, left);   // Resolve the l-value of the left side
2353
2354                         switch(binary->getOp())
2355                         {
2356                         case EOpIndexDirect:
2357                                 {
2358                                         int rightIndex = right->getAsConstantUnion()->getIConst(0);
2359
2360                                         if(left->isRegister())
2361                                         {
2362                                                 int leftMask = dst.mask;
2363
2364                                                 dst.mask = 1;
2365                                                 while((leftMask & dst.mask) == 0)
2366                                                 {
2367                                                         dst.mask = dst.mask << 1;
2368                                                 }
2369
2370                                                 int element = swizzleElement(leftSwizzle, rightIndex);
2371                                                 dst.mask = 1 << element;
2372
2373                                                 return element;
2374                                         }
2375                                         else if(left->isArray() || left->isMatrix())
2376                                         {
2377                                                 dst.index += rightIndex * result->totalRegisterCount();
2378                                                 return 0xE4;
2379                                         }
2380                                         else UNREACHABLE(0);
2381                                 }
2382                                 break;
2383                         case EOpIndexIndirect:
2384                                 {
2385                                         if(left->isRegister())
2386                                         {
2387                                                 // Requires INSERT instruction (handled by calling function)
2388                                         }
2389                                         else if(left->isArray() || left->isMatrix())
2390                                         {
2391                                                 int scale = result->totalRegisterCount();
2392
2393                                                 if(dst.rel.type == sw::Shader::PARAMETER_VOID)   // Use the index register as the relative address directly
2394                                                 {
2395                                                         if(left->totalRegisterCount() > 1)
2396                                                         {
2397                                                                 sw::Shader::SourceParameter relativeRegister;
2398                                                                 argument(relativeRegister, right);
2399
2400                                                                 dst.rel.index = relativeRegister.index;
2401                                                                 dst.rel.type = relativeRegister.type;
2402                                                                 dst.rel.scale = scale;
2403                                                                 dst.rel.deterministic = !(vertexShader && left->getQualifier() == EvqUniform);
2404                                                         }
2405                                                 }
2406                                                 else if(dst.rel.index != registerIndex(&address))   // Move the previous index register to the address register
2407                                                 {
2408                                                         if(scale == 1)
2409                                                         {
2410                                                                 Constant oldScale((int)dst.rel.scale);
2411                                                                 Instruction *mad = emit(sw::Shader::OPCODE_IMAD, &address, &address, &oldScale, right);
2412                                                                 mad->src[0].index = dst.rel.index;
2413                                                                 mad->src[0].type = dst.rel.type;
2414                                                         }
2415                                                         else
2416                                                         {
2417                                                                 Constant oldScale((int)dst.rel.scale);
2418                                                                 Instruction *mul = emit(sw::Shader::OPCODE_IMUL, &address, &address, &oldScale);
2419                                                                 mul->src[0].index = dst.rel.index;
2420                                                                 mul->src[0].type = dst.rel.type;
2421
2422                                                                 Constant newScale(scale);
2423                                                                 emit(sw::Shader::OPCODE_IMAD, &address, right, &newScale, &address);
2424                                                         }
2425
2426                                                         dst.rel.type = sw::Shader::PARAMETER_TEMP;
2427                                                         dst.rel.index = registerIndex(&address);
2428                                                         dst.rel.scale = 1;
2429                                                 }
2430                                                 else   // Just add the new index to the address register
2431                                                 {
2432                                                         if(scale == 1)
2433                                                         {
2434                                                                 emit(sw::Shader::OPCODE_IADD, &address, &address, right);
2435                                                         }
2436                                                         else
2437                                                         {
2438                                                                 Constant newScale(scale);
2439                                                                 emit(sw::Shader::OPCODE_IMAD, &address, right, &newScale, &address);
2440                                                         }
2441                                                 }
2442                                         }
2443                                         else UNREACHABLE(0);
2444                                 }
2445                                 break;
2446                         case EOpIndexDirectStruct:
2447                         case EOpIndexDirectInterfaceBlock:
2448                                 {
2449                                         const TFieldList& fields = (binary->getOp() == EOpIndexDirectStruct) ?
2450                                                                    left->getType().getStruct()->fields() :
2451                                                                    left->getType().getInterfaceBlock()->fields();
2452                                         int index = right->getAsConstantUnion()->getIConst(0);
2453                                         int fieldOffset = 0;
2454
2455                                         for(int i = 0; i < index; i++)
2456                                         {
2457                                                 fieldOffset += fields[i]->type()->totalRegisterCount();
2458                                         }
2459
2460                                         dst.type = registerType(left);
2461                                         dst.index += fieldOffset;
2462                                         dst.mask = writeMask(right);
2463
2464                                         return 0xE4;
2465                                 }
2466                                 break;
2467                         case EOpVectorSwizzle:
2468                                 {
2469                                         ASSERT(left->isRegister());
2470
2471                                         int leftMask = dst.mask;
2472
2473                                         int swizzle = 0;
2474                                         int rightMask = 0;
2475
2476                                         TIntermSequence &sequence = right->getAsAggregate()->getSequence();
2477
2478                                         for(unsigned int i = 0; i < sequence.size(); i++)
2479                                         {
2480                                                 int index = sequence[i]->getAsConstantUnion()->getIConst(0);
2481
2482                                                 int element = swizzleElement(leftSwizzle, index);
2483                                                 rightMask = rightMask | (1 << element);
2484                                                 swizzle = swizzle | swizzleElement(leftSwizzle, i) << (element * 2);
2485                                         }
2486
2487                                         dst.mask = leftMask & rightMask;
2488
2489                                         return swizzle;
2490                                 }
2491                                 break;
2492                         default:
2493                                 UNREACHABLE(binary->getOp());   // Not an l-value operator
2494                                 break;
2495                         }
2496                 }
2497                 else if(symbol)
2498                 {
2499                         dst.type = registerType(symbol);
2500                         dst.index = registerIndex(symbol);
2501                         dst.mask = writeMask(symbol);
2502                         return 0xE4;
2503                 }
2504
2505                 return 0xE4;
2506         }
2507
2508         sw::Shader::ParameterType OutputASM::registerType(TIntermTyped *operand)
2509         {
2510                 if(isSamplerRegister(operand))
2511                 {
2512                         return sw::Shader::PARAMETER_SAMPLER;
2513                 }
2514
2515                 const TQualifier qualifier = operand->getQualifier();
2516                 if((EvqFragColor == qualifier) || (EvqFragData == qualifier))
2517                 {
2518                         if(((EvqFragData == qualifier) && (EvqFragColor == outputQualifier)) ||
2519                            ((EvqFragColor == qualifier) && (EvqFragData == outputQualifier)))
2520                         {
2521                                 mContext.error(operand->getLine(), "static assignment to both gl_FragData and gl_FragColor", "");
2522                         }
2523                         outputQualifier = qualifier;
2524                 }
2525
2526                 if(qualifier == EvqConstExpr && (!operand->getAsConstantUnion() || !operand->getAsConstantUnion()->getUnionArrayPointer()))
2527                 {
2528                         return sw::Shader::PARAMETER_TEMP;
2529                 }
2530
2531                 switch(qualifier)
2532                 {
2533                 case EvqTemporary:           return sw::Shader::PARAMETER_TEMP;
2534                 case EvqGlobal:              return sw::Shader::PARAMETER_TEMP;
2535                 case EvqConstExpr:           return sw::Shader::PARAMETER_FLOAT4LITERAL;   // All converted to float
2536                 case EvqAttribute:           return sw::Shader::PARAMETER_INPUT;
2537                 case EvqVaryingIn:           return sw::Shader::PARAMETER_INPUT;
2538                 case EvqVaryingOut:          return sw::Shader::PARAMETER_OUTPUT;
2539                 case EvqVertexIn:            return sw::Shader::PARAMETER_INPUT;
2540                 case EvqFragmentOut:         return sw::Shader::PARAMETER_COLOROUT;
2541                 case EvqVertexOut:           return sw::Shader::PARAMETER_OUTPUT;
2542                 case EvqFragmentIn:          return sw::Shader::PARAMETER_INPUT;
2543                 case EvqInvariantVaryingIn:  return sw::Shader::PARAMETER_INPUT;    // FIXME: Guarantee invariance at the backend
2544                 case EvqInvariantVaryingOut: return sw::Shader::PARAMETER_OUTPUT;   // FIXME: Guarantee invariance at the backend
2545                 case EvqSmooth:              return sw::Shader::PARAMETER_OUTPUT;
2546                 case EvqFlat:                return sw::Shader::PARAMETER_OUTPUT;
2547                 case EvqCentroidOut:         return sw::Shader::PARAMETER_OUTPUT;
2548                 case EvqSmoothIn:            return sw::Shader::PARAMETER_INPUT;
2549                 case EvqFlatIn:              return sw::Shader::PARAMETER_INPUT;
2550                 case EvqCentroidIn:          return sw::Shader::PARAMETER_INPUT;
2551                 case EvqUniform:             return sw::Shader::PARAMETER_CONST;
2552                 case EvqIn:                  return sw::Shader::PARAMETER_TEMP;
2553                 case EvqOut:                 return sw::Shader::PARAMETER_TEMP;
2554                 case EvqInOut:               return sw::Shader::PARAMETER_TEMP;
2555                 case EvqConstReadOnly:       return sw::Shader::PARAMETER_TEMP;
2556                 case EvqPosition:            return sw::Shader::PARAMETER_OUTPUT;
2557                 case EvqPointSize:           return sw::Shader::PARAMETER_OUTPUT;
2558                 case EvqInstanceID:          return sw::Shader::PARAMETER_MISCTYPE;
2559                 case EvqFragCoord:           return sw::Shader::PARAMETER_MISCTYPE;
2560                 case EvqFrontFacing:         return sw::Shader::PARAMETER_MISCTYPE;
2561                 case EvqPointCoord:          return sw::Shader::PARAMETER_INPUT;
2562                 case EvqFragColor:           return sw::Shader::PARAMETER_COLOROUT;
2563                 case EvqFragData:            return sw::Shader::PARAMETER_COLOROUT;
2564                 case EvqFragDepth:           return sw::Shader::PARAMETER_DEPTHOUT;
2565                 default: UNREACHABLE(qualifier);
2566                 }
2567
2568                 return sw::Shader::PARAMETER_VOID;
2569         }
2570
2571         unsigned int OutputASM::registerIndex(TIntermTyped *operand)
2572         {
2573                 if(isSamplerRegister(operand))
2574                 {
2575                         return samplerRegister(operand);
2576                 }
2577
2578                 switch(operand->getQualifier())
2579                 {
2580                 case EvqTemporary:           return temporaryRegister(operand);
2581                 case EvqGlobal:              return temporaryRegister(operand);
2582                 case EvqConstExpr:           return temporaryRegister(operand);   // Unevaluated constant expression
2583                 case EvqAttribute:           return attributeRegister(operand);
2584                 case EvqVaryingIn:           return varyingRegister(operand);
2585                 case EvqVaryingOut:          return varyingRegister(operand);
2586                 case EvqVertexIn:            return attributeRegister(operand);
2587                 case EvqFragmentOut:         return fragmentOutputRegister(operand);
2588                 case EvqVertexOut:           return varyingRegister(operand);
2589                 case EvqFragmentIn:          return varyingRegister(operand);
2590                 case EvqInvariantVaryingIn:  return varyingRegister(operand);
2591                 case EvqInvariantVaryingOut: return varyingRegister(operand);
2592                 case EvqSmooth:              return varyingRegister(operand);
2593                 case EvqFlat:                return varyingRegister(operand);
2594                 case EvqCentroidOut:         return varyingRegister(operand);
2595                 case EvqSmoothIn:            return varyingRegister(operand);
2596                 case EvqFlatIn:              return varyingRegister(operand);
2597                 case EvqCentroidIn:          return varyingRegister(operand);
2598                 case EvqUniform:             return uniformRegister(operand);
2599                 case EvqIn:                  return temporaryRegister(operand);
2600                 case EvqOut:                 return temporaryRegister(operand);
2601                 case EvqInOut:               return temporaryRegister(operand);
2602                 case EvqConstReadOnly:       return temporaryRegister(operand);
2603                 case EvqPosition:            return varyingRegister(operand);
2604                 case EvqPointSize:           return varyingRegister(operand);
2605                 case EvqInstanceID:          vertexShader->instanceIdDeclared = true; return 0;
2606                 case EvqFragCoord:           pixelShader->vPosDeclared = true;  return 0;
2607                 case EvqFrontFacing:         pixelShader->vFaceDeclared = true; return 1;
2608                 case EvqPointCoord:          return varyingRegister(operand);
2609                 case EvqFragColor:           return 0;
2610                 case EvqFragData:            return fragmentOutputRegister(operand);
2611                 case EvqFragDepth:           return 0;
2612                 default: UNREACHABLE(operand->getQualifier());
2613                 }
2614
2615                 return 0;
2616         }
2617
2618         int OutputASM::writeMask(TIntermTyped *destination, int index)
2619         {
2620                 if(destination->getQualifier() == EvqPointSize)
2621                 {
2622                         return 0x2;   // Point size stored in the y component
2623                 }
2624
2625                 return 0xF >> (4 - registerSize(destination->getType(), index));
2626         }
2627
2628         int OutputASM::readSwizzle(TIntermTyped *argument, int size)
2629         {
2630                 if(argument->getQualifier() == EvqPointSize)
2631                 {
2632                         return 0x55;   // Point size stored in the y component
2633                 }
2634
2635                 static const unsigned char swizzleSize[5] = {0x00, 0x00, 0x54, 0xA4, 0xE4};   // (void), xxxx, xyyy, xyzz, xyzw
2636
2637                 return swizzleSize[size];
2638         }
2639
2640         // Conservatively checks whether an expression is fast to compute and has no side effects
2641         bool OutputASM::trivial(TIntermTyped *expression, int budget)
2642         {
2643                 if(!expression->isRegister())
2644                 {
2645                         return false;
2646                 }
2647
2648                 return cost(expression, budget) >= 0;
2649         }
2650
2651         // Returns the remaining computing budget (if < 0 the expression is too expensive or has side effects)
2652         int OutputASM::cost(TIntermNode *expression, int budget)
2653         {
2654                 if(budget < 0)
2655                 {
2656                         return budget;
2657                 }
2658
2659                 if(expression->getAsSymbolNode())
2660                 {
2661                         return budget;
2662                 }
2663                 else if(expression->getAsConstantUnion())
2664                 {
2665                         return budget;
2666                 }
2667                 else if(expression->getAsBinaryNode())
2668                 {
2669                         TIntermBinary *binary = expression->getAsBinaryNode();
2670
2671                         switch(binary->getOp())
2672                         {
2673                         case EOpVectorSwizzle:
2674                         case EOpIndexDirect:
2675                         case EOpIndexDirectStruct:
2676                         case EOpIndexDirectInterfaceBlock:
2677                                 return cost(binary->getLeft(), budget - 0);
2678                         case EOpAdd:
2679                         case EOpSub:
2680                         case EOpMul:
2681                                 return cost(binary->getLeft(), cost(binary->getRight(), budget - 1));
2682                         default:
2683                                 return -1;
2684                         }
2685                 }
2686                 else if(expression->getAsUnaryNode())
2687                 {
2688                         TIntermUnary *unary = expression->getAsUnaryNode();
2689
2690                         switch(unary->getOp())
2691                         {
2692                         case EOpAbs:
2693                         case EOpNegative:
2694                                 return cost(unary->getOperand(), budget - 1);
2695                         default:
2696                                 return -1;
2697                         }
2698                 }
2699                 else if(expression->getAsSelectionNode())
2700                 {
2701                         TIntermSelection *selection = expression->getAsSelectionNode();
2702
2703                         if(selection->usesTernaryOperator())
2704                         {
2705                                 TIntermTyped *condition = selection->getCondition();
2706                                 TIntermNode *trueBlock = selection->getTrueBlock();
2707                                 TIntermNode *falseBlock = selection->getFalseBlock();
2708                                 TIntermConstantUnion *constantCondition = condition->getAsConstantUnion();
2709
2710                                 if(constantCondition)
2711                                 {
2712                                         bool trueCondition = constantCondition->getUnionArrayPointer()->getBConst();
2713
2714                                         if(trueCondition)
2715                                         {
2716                                                 return cost(trueBlock, budget - 0);
2717                                         }
2718                                         else
2719                                         {
2720                                                 return cost(falseBlock, budget - 0);
2721                                         }
2722                                 }
2723                                 else
2724                                 {
2725                                         return cost(trueBlock, cost(falseBlock, budget - 2));
2726                                 }
2727                         }
2728                 }
2729
2730                 return -1;
2731         }
2732
2733         const Function *OutputASM::findFunction(const TString &name)
2734         {
2735                 for(unsigned int f = 0; f < functionArray.size(); f++)
2736                 {
2737                         if(functionArray[f].name == name)
2738                         {
2739                                 return &functionArray[f];
2740                         }
2741                 }
2742
2743                 return 0;
2744         }
2745
2746         int OutputASM::temporaryRegister(TIntermTyped *temporary)
2747         {
2748                 return allocate(temporaries, temporary);
2749         }
2750
2751         int OutputASM::varyingRegister(TIntermTyped *varying)
2752         {
2753                 int var = lookup(varyings, varying);
2754
2755                 if(var == -1)
2756                 {
2757                         var = allocate(varyings, varying);
2758                         int componentCount = varying->registerSize();
2759                         int registerCount = varying->totalRegisterCount();
2760
2761                         if(pixelShader)
2762                         {
2763                                 if((var + registerCount) > sw::PixelShader::MAX_INPUT_VARYINGS)
2764                                 {
2765                                         mContext.error(varying->getLine(), "Varyings packing failed: Too many varyings", "fragment shader");
2766                                         return 0;
2767                                 }
2768
2769                                 if(varying->getQualifier() == EvqPointCoord)
2770                                 {
2771                                         ASSERT(varying->isRegister());
2772                                         if(componentCount >= 1) pixelShader->semantic[var][0] = sw::Shader::Semantic(sw::Shader::USAGE_TEXCOORD, var);
2773                                         if(componentCount >= 2) pixelShader->semantic[var][1] = sw::Shader::Semantic(sw::Shader::USAGE_TEXCOORD, var);
2774                                         if(componentCount >= 3) pixelShader->semantic[var][2] = sw::Shader::Semantic(sw::Shader::USAGE_TEXCOORD, var);
2775                                         if(componentCount >= 4) pixelShader->semantic[var][3] = sw::Shader::Semantic(sw::Shader::USAGE_TEXCOORD, var);
2776                                 }
2777                                 else
2778                                 {
2779                                         for(int i = 0; i < varying->totalRegisterCount(); i++)
2780                                         {
2781                                                 if(componentCount >= 1) pixelShader->semantic[var + i][0] = sw::Shader::Semantic(sw::Shader::USAGE_COLOR, var + i);
2782                                                 if(componentCount >= 2) pixelShader->semantic[var + i][1] = sw::Shader::Semantic(sw::Shader::USAGE_COLOR, var + i);
2783                                                 if(componentCount >= 3) pixelShader->semantic[var + i][2] = sw::Shader::Semantic(sw::Shader::USAGE_COLOR, var + i);
2784                                                 if(componentCount >= 4) pixelShader->semantic[var + i][3] = sw::Shader::Semantic(sw::Shader::USAGE_COLOR, var + i);
2785                                         }
2786                                 }
2787                         }
2788                         else if(vertexShader)
2789                         {
2790                                 if((var + registerCount) > sw::MAX_VERTEX_OUTPUTS)
2791                                 {
2792                                         mContext.error(varying->getLine(), "Varyings packing failed: Too many varyings", "vertex shader");
2793                                         return 0;
2794                                 }
2795
2796                                 if(varying->getQualifier() == EvqPosition)
2797                                 {
2798                                         ASSERT(varying->isRegister());
2799                                         vertexShader->output[var][0] = sw::Shader::Semantic(sw::Shader::USAGE_POSITION, 0);
2800                                         vertexShader->output[var][1] = sw::Shader::Semantic(sw::Shader::USAGE_POSITION, 0);
2801                                         vertexShader->output[var][2] = sw::Shader::Semantic(sw::Shader::USAGE_POSITION, 0);
2802                                         vertexShader->output[var][3] = sw::Shader::Semantic(sw::Shader::USAGE_POSITION, 0);
2803                                         vertexShader->positionRegister = var;
2804                                 }
2805                                 else if(varying->getQualifier() == EvqPointSize)
2806                                 {
2807                                         ASSERT(varying->isRegister());
2808                                         vertexShader->output[var][0] = sw::Shader::Semantic(sw::Shader::USAGE_PSIZE, 0);
2809                                         vertexShader->output[var][1] = sw::Shader::Semantic(sw::Shader::USAGE_PSIZE, 0);
2810                                         vertexShader->output[var][2] = sw::Shader::Semantic(sw::Shader::USAGE_PSIZE, 0);
2811                                         vertexShader->output[var][3] = sw::Shader::Semantic(sw::Shader::USAGE_PSIZE, 0);
2812                                         vertexShader->pointSizeRegister = var;
2813                                 }
2814                                 else
2815                                 {
2816                                         // Semantic indexes for user varyings will be assigned during program link to match the pixel shader
2817                                 }
2818                         }
2819                         else UNREACHABLE(0);
2820
2821                         declareVarying(varying, var);
2822                 }
2823
2824                 return var;
2825         }
2826
2827         void OutputASM::declareVarying(TIntermTyped *varying, int reg)
2828         {
2829                 if(varying->getQualifier() != EvqPointCoord)   // gl_PointCoord does not need linking
2830                 {
2831                         const TType &type = varying->getType();
2832                         const char *name = varying->getAsSymbolNode()->getSymbol().c_str();
2833                         VaryingList &activeVaryings = shaderObject->varyings;
2834
2835                         // Check if this varying has been declared before without having a register assigned
2836                         for(VaryingList::iterator v = activeVaryings.begin(); v != activeVaryings.end(); v++)
2837                         {
2838                                 if(v->name == name)
2839                                 {
2840                                         if(reg >= 0)
2841                                         {
2842                                                 ASSERT(v->reg < 0 || v->reg == reg);
2843                                                 v->reg = reg;
2844                                         }
2845
2846                                         return;
2847                                 }
2848                         }
2849
2850                         activeVaryings.push_back(glsl::Varying(glVariableType(type), name, varying->getArraySize(), reg, 0));
2851                 }
2852         }
2853
2854         int OutputASM::uniformRegister(TIntermTyped *uniform)
2855         {
2856                 const TType &type = uniform->getType();
2857                 ASSERT(!IsSampler(type.getBasicType()));
2858                 TInterfaceBlock *block = type.getAsInterfaceBlock();
2859                 TIntermSymbol *symbol = uniform->getAsSymbolNode();
2860                 ASSERT(symbol || block);
2861
2862                 if(symbol || block)
2863                 {
2864                         TInterfaceBlock* parentBlock = type.getInterfaceBlock();
2865                         bool isBlockMember = (!block && parentBlock);
2866                         int index = isBlockMember ? lookup(uniforms, parentBlock) : lookup(uniforms, uniform);
2867
2868                         if(index == -1 || isBlockMember)
2869                         {
2870                                 if(index == -1)
2871                                 {
2872                                         index = allocate(uniforms, uniform);
2873                                 }
2874
2875                                 // Verify if the current uniform is a member of an already declared block
2876                                 const TString &name = symbol ? symbol->getSymbol() : block->name();
2877                                 int blockMemberIndex = blockMemberLookup(type, name, index);
2878                                 if(blockMemberIndex == -1)
2879                                 {
2880                                         declareUniform(type, name, index);
2881                                 }
2882                                 else
2883                                 {
2884                                         index = blockMemberIndex;
2885                                 }
2886                         }
2887
2888                         return index;
2889                 }
2890
2891                 return 0;
2892         }
2893
2894         int OutputASM::attributeRegister(TIntermTyped *attribute)
2895         {
2896                 ASSERT(!attribute->isArray());
2897
2898                 int index = lookup(attributes, attribute);
2899
2900                 if(index == -1)
2901                 {
2902                         TIntermSymbol *symbol = attribute->getAsSymbolNode();
2903                         ASSERT(symbol);
2904
2905                         if(symbol)
2906                         {
2907                                 index = allocate(attributes, attribute);
2908                                 const TType &type = attribute->getType();
2909                                 int registerCount = attribute->totalRegisterCount();
2910
2911                                 if(vertexShader && (index + registerCount) <= sw::VertexShader::MAX_INPUT_ATTRIBUTES)
2912                                 {
2913                                         for(int i = 0; i < registerCount; i++)
2914                                         {
2915                                                 vertexShader->input[index + i] = sw::Shader::Semantic(sw::Shader::USAGE_TEXCOORD, index + i);
2916                                         }
2917                                 }
2918
2919                                 ActiveAttributes &activeAttributes = shaderObject->activeAttributes;
2920
2921                                 const char *name = symbol->getSymbol().c_str();
2922                                 activeAttributes.push_back(Attribute(glVariableType(type), name, type.getArraySize(), type.getLayoutQualifier().location, index));
2923                         }
2924                 }
2925
2926                 return index;
2927         }
2928
2929         int OutputASM::fragmentOutputRegister(TIntermTyped *fragmentOutput)
2930         {
2931                 return allocate(fragmentOutputs, fragmentOutput);
2932         }
2933
2934         int OutputASM::samplerRegister(TIntermTyped *sampler)
2935         {
2936                 const TType &type = sampler->getType();
2937                 ASSERT(IsSampler(type.getBasicType()) || type.isStruct());   // Structures can contain samplers
2938
2939                 TIntermSymbol *symbol = sampler->getAsSymbolNode();
2940                 TIntermBinary *binary = sampler->getAsBinaryNode();
2941
2942                 if(symbol && type.getQualifier() == EvqUniform)
2943                 {
2944                         return samplerRegister(symbol);
2945                 }
2946                 else if(binary)
2947                 {
2948                         TIntermTyped *left = binary->getLeft();
2949                         TIntermTyped *right = binary->getRight();
2950                         const TType &leftType = left->getType();
2951                         int index = right->getAsConstantUnion() ? right->getAsConstantUnion()->getIConst(0) : 0;
2952                         int offset = 0;
2953
2954                         switch(binary->getOp())
2955                         {
2956                         case EOpIndexDirect:
2957                                 ASSERT(left->isArray());
2958                                 offset = index * leftType.elementRegisterCount();
2959                                 break;
2960                         case EOpIndexDirectStruct:
2961                                 ASSERT(leftType.isStruct());
2962                                 {
2963                                         const TFieldList &fields = leftType.getStruct()->fields();
2964
2965                                         for(int i = 0; i < index; i++)
2966                                         {
2967                                                 offset += fields[i]->type()->totalRegisterCount();
2968                                         }
2969                                 }
2970                                 break;
2971                         case EOpIndexIndirect:               // Indirect indexing produces a temporary, not a sampler register
2972                                 return -1;
2973                         case EOpIndexDirectInterfaceBlock:   // Interface blocks can't contain samplers
2974                         default:
2975                                 UNREACHABLE(binary->getOp());
2976                                 return -1;
2977                         }
2978
2979                         int base = samplerRegister(left);
2980
2981                         if(base < 0)
2982                         {
2983                                 return -1;
2984                         }
2985
2986                         return base + offset;
2987                 }
2988
2989                 UNREACHABLE(0);
2990                 return -1;   // Not a sampler register
2991         }
2992
2993         int OutputASM::samplerRegister(TIntermSymbol *sampler)
2994         {
2995                 const TType &type = sampler->getType();
2996                 ASSERT(IsSampler(type.getBasicType()) || type.isStruct());   // Structures can contain samplers
2997
2998                 int index = lookup(samplers, sampler);
2999
3000                 if(index == -1)
3001                 {
3002                         index = allocate(samplers, sampler);
3003
3004                         if(sampler->getQualifier() == EvqUniform)
3005                         {
3006                                 const char *name = sampler->getSymbol().c_str();
3007                                 declareUniform(type, name, index);
3008                         }
3009                 }
3010
3011                 return index;
3012         }
3013
3014         bool OutputASM::isSamplerRegister(TIntermTyped *operand)
3015         {
3016                 return operand && IsSampler(operand->getBasicType()) && samplerRegister(operand) >= 0;
3017         }
3018
3019         int OutputASM::lookup(VariableArray &list, TIntermTyped *variable)
3020         {
3021                 for(unsigned int i = 0; i < list.size(); i++)
3022                 {
3023                         if(list[i] == variable)
3024                         {
3025                                 return i;   // Pointer match
3026                         }
3027                 }
3028
3029                 TIntermSymbol *varSymbol = variable->getAsSymbolNode();
3030                 TInterfaceBlock *varBlock = variable->getType().getAsInterfaceBlock();
3031
3032                 if(varBlock)
3033                 {
3034                         for(unsigned int i = 0; i < list.size(); i++)
3035                         {
3036                                 if(list[i])
3037                                 {
3038                                         TInterfaceBlock *listBlock = list[i]->getType().getAsInterfaceBlock();
3039
3040                                         if(listBlock)
3041                                         {
3042                                                 if(listBlock->name() == varBlock->name())
3043                                                 {
3044                                                         ASSERT(listBlock->arraySize() == varBlock->arraySize());
3045                                                         ASSERT(listBlock->fields() == varBlock->fields());
3046                                                         ASSERT(listBlock->blockStorage() == varBlock->blockStorage());
3047                                                         ASSERT(listBlock->matrixPacking() == varBlock->matrixPacking());
3048
3049                                                         return i;
3050                                                 }
3051                                         }
3052                                 }
3053                         }
3054                 }
3055                 else if(varSymbol)
3056                 {
3057                         for(unsigned int i = 0; i < list.size(); i++)
3058                         {
3059                                 if(list[i])
3060                                 {
3061                                         TIntermSymbol *listSymbol = list[i]->getAsSymbolNode();
3062
3063                                         if(listSymbol)
3064                                         {
3065                                                 if(listSymbol->getId() == varSymbol->getId())
3066                                                 {
3067                                                         ASSERT(listSymbol->getSymbol() == varSymbol->getSymbol());
3068                                                         ASSERT(listSymbol->getType() == varSymbol->getType());
3069                                                         ASSERT(listSymbol->getQualifier() == varSymbol->getQualifier());
3070
3071                                                         return i;
3072                                                 }
3073                                         }
3074                                 }
3075                         }
3076                 }
3077
3078                 return -1;
3079         }
3080
3081         int OutputASM::lookup(VariableArray &list, TInterfaceBlock *block)
3082         {
3083                 for(unsigned int i = 0; i < list.size(); i++)
3084                 {
3085                         if(list[i] && (list[i]->getType().getInterfaceBlock() == block))
3086                         {
3087                                 return i;   // Pointer match
3088                         }
3089                 }
3090                 return -1;
3091         }
3092
3093         int OutputASM::allocate(VariableArray &list, TIntermTyped *variable)
3094         {
3095                 int index = lookup(list, variable);
3096
3097                 if(index == -1)
3098                 {
3099                         unsigned int registerCount = variable->blockRegisterCount();
3100
3101                         for(unsigned int i = 0; i < list.size(); i++)
3102                         {
3103                                 if(list[i] == 0)
3104                                 {
3105                                         unsigned int j = 1;
3106                                         for( ; j < registerCount && (i + j) < list.size(); j++)
3107                                         {
3108                                                 if(list[i + j] != 0)
3109                                                 {
3110                                                         break;
3111                                                 }
3112                                         }
3113
3114                                         if(j == registerCount)   // Found free slots
3115                                         {
3116                                                 for(unsigned int j = 0; j < registerCount; j++)
3117                                                 {
3118                                                         list[i + j] = variable;
3119                                                 }
3120
3121                                                 return i;
3122                                         }
3123                                 }
3124                         }
3125
3126                         index = list.size();
3127
3128                         for(unsigned int i = 0; i < registerCount; i++)
3129                         {
3130                                 list.push_back(variable);
3131                         }
3132                 }
3133
3134                 return index;
3135         }
3136
3137         void OutputASM::free(VariableArray &list, TIntermTyped *variable)
3138         {
3139                 int index = lookup(list, variable);
3140
3141                 if(index >= 0)
3142                 {
3143                         list[index] = 0;
3144                 }
3145         }
3146
3147         int OutputASM::blockMemberLookup(const TType &type, const TString &name, int registerIndex)
3148         {
3149                 const TInterfaceBlock *block = type.getInterfaceBlock();
3150
3151                 if(block)
3152                 {
3153                         ActiveUniformBlocks &activeUniformBlocks = shaderObject->activeUniformBlocks;
3154                         const TFieldList& fields = block->fields();
3155                         const TString &blockName = block->name();
3156                         int fieldRegisterIndex = registerIndex;
3157
3158                         if(!type.isInterfaceBlock())
3159                         {
3160                                 // This is a uniform that's part of a block, let's see if the block is already defined
3161                                 for(size_t i = 0; i < activeUniformBlocks.size(); ++i)
3162                                 {
3163                                         if(activeUniformBlocks[i].name == blockName.c_str())
3164                                         {
3165                                                 // The block is already defined, find the register for the current uniform and return it
3166                                                 for(size_t j = 0; j < fields.size(); j++)
3167                                                 {
3168                                                         const TString &fieldName = fields[j]->name();
3169                                                         if(fieldName == name)
3170                                                         {
3171                                                                 return fieldRegisterIndex;
3172                                                         }
3173
3174                                                         fieldRegisterIndex += fields[j]->type()->totalRegisterCount();
3175                                                 }
3176
3177                                                 ASSERT(false);
3178                                                 return fieldRegisterIndex;
3179                                         }
3180                                 }
3181                         }
3182                 }
3183
3184                 return -1;
3185         }
3186
3187         void OutputASM::declareUniform(const TType &type, const TString &name, int registerIndex, int blockId, BlockLayoutEncoder* encoder)
3188         {
3189                 const TStructure *structure = type.getStruct();
3190                 const TInterfaceBlock *block = (type.isInterfaceBlock() || (blockId == -1)) ? type.getInterfaceBlock() : nullptr;
3191
3192                 if(!structure && !block)
3193                 {
3194                         ActiveUniforms &activeUniforms = shaderObject->activeUniforms;
3195                         const BlockMemberInfo blockInfo = encoder ? encoder->encodeType(type) : BlockMemberInfo::getDefaultBlockInfo();
3196                         if(blockId >= 0)
3197                         {
3198                                 blockDefinitions[blockId][registerIndex] = TypedMemberInfo(blockInfo, type);
3199                                 shaderObject->activeUniformBlocks[blockId].fields.push_back(activeUniforms.size());
3200                         }
3201                         int fieldRegisterIndex = encoder ? shaderObject->activeUniformBlocks[blockId].registerIndex + BlockLayoutEncoder::getBlockRegister(blockInfo) : registerIndex;
3202                         activeUniforms.push_back(Uniform(glVariableType(type), glVariablePrecision(type), name.c_str(), type.getArraySize(),
3203                                                          fieldRegisterIndex, blockId, blockInfo));
3204                         if(IsSampler(type.getBasicType()))
3205                         {
3206                                 for(int i = 0; i < type.totalRegisterCount(); i++)
3207                                 {
3208                                         shader->declareSampler(fieldRegisterIndex + i);
3209                                 }
3210                         }
3211                 }
3212                 else if(block)
3213                 {
3214                         ActiveUniformBlocks &activeUniformBlocks = shaderObject->activeUniformBlocks;
3215                         const TFieldList& fields = block->fields();
3216                         const TString &blockName = block->name();
3217                         int fieldRegisterIndex = registerIndex;
3218                         bool isUniformBlockMember = !type.isInterfaceBlock() && (blockId == -1);
3219
3220                         blockId = activeUniformBlocks.size();
3221                         bool isRowMajor = block->matrixPacking() == EmpRowMajor;
3222                         activeUniformBlocks.push_back(UniformBlock(blockName.c_str(), 0, block->arraySize(),
3223                                                                    block->blockStorage(), isRowMajor, registerIndex, blockId));
3224                         blockDefinitions.push_back(BlockDefinitionIndexMap());
3225
3226                         Std140BlockEncoder currentBlockEncoder(isRowMajor);
3227                         currentBlockEncoder.enterAggregateType();
3228                         for(size_t i = 0; i < fields.size(); i++)
3229                         {
3230                                 const TType &fieldType = *(fields[i]->type());
3231                                 const TString &fieldName = fields[i]->name();
3232                                 if(isUniformBlockMember && (fieldName == name))
3233                                 {
3234                                         registerIndex = fieldRegisterIndex;
3235                                 }
3236
3237                                 const TString uniformName = block->hasInstanceName() ? blockName + "." + fieldName : fieldName;
3238
3239                                 declareUniform(fieldType, uniformName, fieldRegisterIndex, blockId, &currentBlockEncoder);
3240                                 fieldRegisterIndex += fieldType.totalRegisterCount();
3241                         }
3242                         currentBlockEncoder.exitAggregateType();
3243                         activeUniformBlocks[blockId].dataSize = currentBlockEncoder.getBlockSize();
3244                 }
3245                 else
3246                 {
3247                         int fieldRegisterIndex = registerIndex;
3248
3249                         const TFieldList& fields = structure->fields();
3250                         if(type.isArray() && (structure || type.isInterfaceBlock()))
3251                         {
3252                                 for(int i = 0; i < type.getArraySize(); i++)
3253                                 {
3254                                         if(encoder)
3255                                         {
3256                                                 encoder->enterAggregateType();
3257                                         }
3258                                         for(size_t j = 0; j < fields.size(); j++)
3259                                         {
3260                                                 const TType &fieldType = *(fields[j]->type());
3261                                                 const TString &fieldName = fields[j]->name();
3262                                                 const TString uniformName = name + "[" + str(i) + "]." + fieldName;
3263
3264                                                 declareUniform(fieldType, uniformName, fieldRegisterIndex, blockId, encoder);
3265                                                 fieldRegisterIndex += fieldType.totalRegisterCount();
3266                                         }
3267                                         if(encoder)
3268                                         {
3269                                                 encoder->exitAggregateType();
3270                                         }
3271                                 }
3272                         }
3273                         else
3274                         {
3275                                 if(encoder)
3276                                 {
3277                                         encoder->enterAggregateType();
3278                                 }
3279                                 for(size_t i = 0; i < fields.size(); i++)
3280                                 {
3281                                         const TType &fieldType = *(fields[i]->type());
3282                                         const TString &fieldName = fields[i]->name();
3283                                         const TString uniformName = name + "." + fieldName;
3284
3285                                         declareUniform(fieldType, uniformName, fieldRegisterIndex, blockId, encoder);
3286                                         fieldRegisterIndex += fieldType.totalRegisterCount();
3287                                 }
3288                                 if(encoder)
3289                                 {
3290                                         encoder->exitAggregateType();
3291                                 }
3292                         }
3293                 }
3294         }
3295
3296         GLenum OutputASM::glVariableType(const TType &type)
3297         {
3298                 switch(type.getBasicType())
3299                 {
3300                 case EbtFloat:
3301                         if(type.isScalar())
3302                         {
3303                                 return GL_FLOAT;
3304                         }
3305                         else if(type.isVector())
3306                         {
3307                                 switch(type.getNominalSize())
3308                                 {
3309                                 case 2: return GL_FLOAT_VEC2;
3310                                 case 3: return GL_FLOAT_VEC3;
3311                                 case 4: return GL_FLOAT_VEC4;
3312                                 default: UNREACHABLE(type.getNominalSize());
3313                                 }
3314                         }
3315                         else if(type.isMatrix())
3316                         {
3317                                 switch(type.getNominalSize())
3318                                 {
3319                                 case 2:
3320                                         switch(type.getSecondarySize())
3321                                         {
3322                                         case 2: return GL_FLOAT_MAT2;
3323                                         case 3: return GL_FLOAT_MAT2x3;
3324                                         case 4: return GL_FLOAT_MAT2x4;
3325                                         default: UNREACHABLE(type.getSecondarySize());
3326                                         }
3327                                 case 3:
3328                                         switch(type.getSecondarySize())
3329                                         {
3330                                         case 2: return GL_FLOAT_MAT3x2;
3331                                         case 3: return GL_FLOAT_MAT3;
3332                                         case 4: return GL_FLOAT_MAT3x4;
3333                                         default: UNREACHABLE(type.getSecondarySize());
3334                                         }
3335                                 case 4:
3336                                         switch(type.getSecondarySize())
3337                                         {
3338                                         case 2: return GL_FLOAT_MAT4x2;
3339                                         case 3: return GL_FLOAT_MAT4x3;
3340                                         case 4: return GL_FLOAT_MAT4;
3341                                         default: UNREACHABLE(type.getSecondarySize());
3342                                         }
3343                                 default: UNREACHABLE(type.getNominalSize());
3344                                 }
3345                         }
3346                         else UNREACHABLE(0);
3347                         break;
3348                 case EbtInt:
3349                         if(type.isScalar())
3350                         {
3351                                 return GL_INT;
3352                         }
3353                         else if(type.isVector())
3354                         {
3355                                 switch(type.getNominalSize())
3356                                 {
3357                                 case 2: return GL_INT_VEC2;
3358                                 case 3: return GL_INT_VEC3;
3359                                 case 4: return GL_INT_VEC4;
3360                                 default: UNREACHABLE(type.getNominalSize());
3361                                 }
3362                         }
3363                         else UNREACHABLE(0);
3364                         break;
3365                 case EbtUInt:
3366                         if(type.isScalar())
3367                         {
3368                                 return GL_UNSIGNED_INT;
3369                         }
3370                         else if(type.isVector())
3371                         {
3372                                 switch(type.getNominalSize())
3373                                 {
3374                                 case 2: return GL_UNSIGNED_INT_VEC2;
3375                                 case 3: return GL_UNSIGNED_INT_VEC3;
3376                                 case 4: return GL_UNSIGNED_INT_VEC4;
3377                                 default: UNREACHABLE(type.getNominalSize());
3378                                 }
3379                         }
3380                         else UNREACHABLE(0);
3381                         break;
3382                 case EbtBool:
3383                         if(type.isScalar())
3384                         {
3385                                 return GL_BOOL;
3386                         }
3387                         else if(type.isVector())
3388                         {
3389                                 switch(type.getNominalSize())
3390                                 {
3391                                 case 2: return GL_BOOL_VEC2;
3392                                 case 3: return GL_BOOL_VEC3;
3393                                 case 4: return GL_BOOL_VEC4;
3394                                 default: UNREACHABLE(type.getNominalSize());
3395                                 }
3396                         }
3397                         else UNREACHABLE(0);
3398                         break;
3399                 case EbtSampler2D:
3400                         return GL_SAMPLER_2D;
3401                 case EbtISampler2D:
3402                         return GL_INT_SAMPLER_2D;
3403                 case EbtUSampler2D:
3404                         return GL_UNSIGNED_INT_SAMPLER_2D;
3405                 case EbtSamplerCube:
3406                         return GL_SAMPLER_CUBE;
3407                 case EbtISamplerCube:
3408                         return GL_INT_SAMPLER_CUBE;
3409                 case EbtUSamplerCube:
3410                         return GL_UNSIGNED_INT_SAMPLER_CUBE;
3411                 case EbtSamplerExternalOES:
3412                         return GL_SAMPLER_EXTERNAL_OES;
3413                 case EbtSampler3D:
3414                         return GL_SAMPLER_3D_OES;
3415                 case EbtISampler3D:
3416                         return GL_INT_SAMPLER_3D;
3417                 case EbtUSampler3D:
3418                         return GL_UNSIGNED_INT_SAMPLER_3D;
3419                 case EbtSampler2DArray:
3420                         return GL_SAMPLER_2D_ARRAY;
3421                 case EbtISampler2DArray:
3422                         return GL_INT_SAMPLER_2D_ARRAY;
3423                 case EbtUSampler2DArray:
3424                         return GL_UNSIGNED_INT_SAMPLER_2D_ARRAY;
3425                 case EbtSampler2DShadow:
3426                         return GL_SAMPLER_2D_SHADOW;
3427                 case EbtSamplerCubeShadow:
3428                         return GL_SAMPLER_CUBE_SHADOW;
3429                 case EbtSampler2DArrayShadow:
3430                         return GL_SAMPLER_2D_ARRAY_SHADOW;
3431                 default:
3432                         UNREACHABLE(type.getBasicType());
3433                         break;
3434                 }
3435
3436                 return GL_NONE;
3437         }
3438
3439         GLenum OutputASM::glVariablePrecision(const TType &type)
3440         {
3441                 if(type.getBasicType() == EbtFloat)
3442                 {
3443                         switch(type.getPrecision())
3444                         {
3445                         case EbpHigh:   return GL_HIGH_FLOAT;
3446                         case EbpMedium: return GL_MEDIUM_FLOAT;
3447                         case EbpLow:    return GL_LOW_FLOAT;
3448                         case EbpUndefined:
3449                                 // Should be defined as the default precision by the parser
3450                         default: UNREACHABLE(type.getPrecision());
3451                         }
3452                 }
3453                 else if(type.getBasicType() == EbtInt)
3454                 {
3455                         switch(type.getPrecision())
3456                         {
3457                         case EbpHigh:   return GL_HIGH_INT;
3458                         case EbpMedium: return GL_MEDIUM_INT;
3459                         case EbpLow:    return GL_LOW_INT;
3460                         case EbpUndefined:
3461                                 // Should be defined as the default precision by the parser
3462                         default: UNREACHABLE(type.getPrecision());
3463                         }
3464                 }
3465
3466                 // Other types (boolean, sampler) don't have a precision
3467                 return GL_NONE;
3468         }
3469
3470         int OutputASM::dim(TIntermNode *v)
3471         {
3472                 TIntermTyped *vector = v->getAsTyped();
3473                 ASSERT(vector && vector->isRegister());
3474                 return vector->getNominalSize();
3475         }
3476
3477         int OutputASM::dim2(TIntermNode *m)
3478         {
3479                 TIntermTyped *matrix = m->getAsTyped();
3480                 ASSERT(matrix && matrix->isMatrix() && !matrix->isArray());
3481                 return matrix->getSecondarySize();
3482         }
3483
3484         // Returns ~0u if no loop count could be determined
3485         unsigned int OutputASM::loopCount(TIntermLoop *node)
3486         {
3487                 // Parse loops of the form:
3488                 // for(int index = initial; index [comparator] limit; index += increment)
3489                 TIntermSymbol *index = 0;
3490                 TOperator comparator = EOpNull;
3491                 int initial = 0;
3492                 int limit = 0;
3493                 int increment = 0;
3494
3495                 // Parse index name and intial value
3496                 if(node->getInit())
3497                 {
3498                         TIntermAggregate *init = node->getInit()->getAsAggregate();
3499
3500                         if(init)
3501                         {
3502                                 TIntermSequence &sequence = init->getSequence();
3503                                 TIntermTyped *variable = sequence[0]->getAsTyped();
3504
3505                                 if(variable && variable->getQualifier() == EvqTemporary)
3506                                 {
3507                                         TIntermBinary *assign = variable->getAsBinaryNode();
3508
3509                                         if(assign->getOp() == EOpInitialize)
3510                                         {
3511                                                 TIntermSymbol *symbol = assign->getLeft()->getAsSymbolNode();
3512                                                 TIntermConstantUnion *constant = assign->getRight()->getAsConstantUnion();
3513
3514                                                 if(symbol && constant)
3515                                                 {
3516                                                         if(constant->getBasicType() == EbtInt && constant->getNominalSize() == 1)
3517                                                         {
3518                                                                 index = symbol;
3519                                                                 initial = constant->getUnionArrayPointer()[0].getIConst();
3520                                                         }
3521                                                 }
3522                                         }
3523                                 }
3524                         }
3525                 }
3526
3527                 // Parse comparator and limit value
3528                 if(index && node->getCondition())
3529                 {
3530                         TIntermBinary *test = node->getCondition()->getAsBinaryNode();
3531
3532                         if(test && test->getLeft()->getAsSymbolNode()->getId() == index->getId())
3533                         {
3534                                 TIntermConstantUnion *constant = test->getRight()->getAsConstantUnion();
3535
3536                                 if(constant)
3537                                 {
3538                                         if(constant->getBasicType() == EbtInt && constant->getNominalSize() == 1)
3539                                         {
3540                                                 comparator = test->getOp();
3541                                                 limit = constant->getUnionArrayPointer()[0].getIConst();
3542                                         }
3543                                 }
3544                         }
3545                 }
3546
3547                 // Parse increment
3548                 if(index && comparator != EOpNull && node->getExpression())
3549                 {
3550                         TIntermBinary *binaryTerminal = node->getExpression()->getAsBinaryNode();
3551                         TIntermUnary *unaryTerminal = node->getExpression()->getAsUnaryNode();
3552
3553                         if(binaryTerminal)
3554                         {
3555                                 TOperator op = binaryTerminal->getOp();
3556                                 TIntermConstantUnion *constant = binaryTerminal->getRight()->getAsConstantUnion();
3557
3558                                 if(constant)
3559                                 {
3560                                         if(constant->getBasicType() == EbtInt && constant->getNominalSize() == 1)
3561                                         {
3562                                                 int value = constant->getUnionArrayPointer()[0].getIConst();
3563
3564                                                 switch(op)
3565                                                 {
3566                                                 case EOpAddAssign: increment = value;  break;
3567                                                 case EOpSubAssign: increment = -value; break;
3568                                                 default: UNIMPLEMENTED();
3569                                                 }
3570                                         }
3571                                 }
3572                         }
3573                         else if(unaryTerminal)
3574                         {
3575                                 TOperator op = unaryTerminal->getOp();
3576
3577                                 switch(op)
3578                                 {
3579                                 case EOpPostIncrement: increment = 1;  break;
3580                                 case EOpPostDecrement: increment = -1; break;
3581                                 case EOpPreIncrement:  increment = 1;  break;
3582                                 case EOpPreDecrement:  increment = -1; break;
3583                                 default: UNIMPLEMENTED();
3584                                 }
3585                         }
3586                 }
3587
3588                 if(index && comparator != EOpNull && increment != 0)
3589                 {
3590                         if(comparator == EOpLessThanEqual)
3591                         {
3592                                 comparator = EOpLessThan;
3593                                 limit += 1;
3594                         }
3595
3596                         if(comparator == EOpLessThan)
3597                         {
3598                                 int iterations = (limit - initial) / increment;
3599
3600                                 if(iterations <= 0)
3601                                 {
3602                                         iterations = 0;
3603                                 }
3604
3605                                 return iterations;
3606                         }
3607                         else UNIMPLEMENTED();   // Falls through
3608                 }
3609
3610                 return ~0u;
3611         }
3612
3613         bool LoopUnrollable::traverse(TIntermNode *node)
3614         {
3615                 loopDepth = 0;
3616                 loopUnrollable = true;
3617
3618                 node->traverse(this);
3619
3620                 return loopUnrollable;
3621         }
3622
3623         bool LoopUnrollable::visitLoop(Visit visit, TIntermLoop *loop)
3624         {
3625                 if(visit == PreVisit)
3626                 {
3627                         loopDepth++;
3628                 }
3629                 else if(visit == PostVisit)
3630                 {
3631                         loopDepth++;
3632                 }
3633
3634                 return true;
3635         }
3636
3637         bool LoopUnrollable::visitBranch(Visit visit, TIntermBranch *node)
3638         {
3639                 if(!loopUnrollable)
3640                 {
3641                         return false;
3642                 }
3643
3644                 if(!loopDepth)
3645                 {
3646                         return true;
3647                 }
3648
3649                 switch(node->getFlowOp())
3650                 {
3651                 case EOpKill:
3652                 case EOpReturn:
3653                         break;
3654                 case EOpBreak:
3655                 case EOpContinue:
3656                         loopUnrollable = false;
3657                         break;
3658                 default: UNREACHABLE(node->getFlowOp());
3659                 }
3660
3661                 return loopUnrollable;
3662         }
3663
3664         bool LoopUnrollable::visitAggregate(Visit visit, TIntermAggregate *node)
3665         {
3666                 return loopUnrollable;
3667         }
3668 }