src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp

   1 /****************************************************************************
   2 * Copyright (C) 2014-2015 Intel Corporation.   All Rights Reserved.
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * @file builder_misc.cpp
  24 *
  25 * @brief Implementation for miscellaneous builder functions
  26 *
  27 * Notes:
  28 *
  29 ******************************************************************************/
  30 #include "builder.h"
  31 #include "common/rdtsc_buckets.h"
  32
  33 #include <cstdarg>
  34
  35 namespace SwrJit
  36 {
  37     void __cdecl CallPrint(const char* fmt, ...);
  38
  39     //////////////////////////////////////////////////////////////////////////
  40     /// @brief Convert an IEEE 754 32-bit single precision float to an
  41     ///        16 bit float with 5 exponent bits and a variable
  42     ///        number of mantissa bits.
  43     /// @param val - 32-bit float
  44     /// @todo Maybe move this outside of this file into a header?
  45     static uint16_t ConvertFloat32ToFloat16(float val)
  46     {
  47         uint32_t sign, exp, mant;
  48         uint32_t roundBits;
  49
  50         // Extract the sign, exponent, and mantissa
  51         uint32_t uf = *(uint32_t*)&val;
  52         sign = (uf & 0x80000000) >> 31;
  53         exp = (uf & 0x7F800000) >> 23;
  54         mant = uf & 0x007FFFFF;
  55
  56         // Check for out of range
  57         if (std::isnan(val))
  58         {
  59             exp = 0x1F;
  60             mant = 0x200;
  61             sign = 1;                     // set the sign bit for NANs
  62         }
  63         else if (std::isinf(val))
  64         {
  65             exp = 0x1f;
  66             mant = 0x0;
  67         }
  68         else if (exp > (0x70 + 0x1E)) // Too big to represent -> max representable value
  69         {
  70             exp = 0x1E;
  71             mant = 0x3FF;
  72         }
  73         else if ((exp <= 0x70) && (exp >= 0x66)) // It's a denorm
  74         {
  75             mant |= 0x00800000;
  76             for (; exp <= 0x70; mant >>= 1, exp++)
  77                 ;
  78             exp = 0;
  79             mant = mant >> 13;
  80         }
  81         else if (exp < 0x66) // Too small to represent -> Zero
  82         {
  83             exp = 0;
  84             mant = 0;
  85         }
  86         else
  87         {
  88             // Saves bits that will be shifted off for rounding
  89             roundBits = mant & 0x1FFFu;
  90             // convert exponent and mantissa to 16 bit format
  91             exp = exp - 0x70;
  92             mant = mant >> 13;
  93
  94             // Essentially RTZ, but round up if off by only 1 lsb
  95             if (roundBits == 0x1FFFu)
  96             {
  97                 mant++;
  98                 // check for overflow
  99                 if ((mant & 0xC00u) != 0)
 100                     exp++;
 101                 // make sure only the needed bits are used
 102                 mant &= 0x3FF;
 103             }
 104         }
 105
 106         uint32_t tmpVal = (sign << 15) | (exp << 10) | mant;
 107         return (uint16_t)tmpVal;
 108     }
 109
 110     //////////////////////////////////////////////////////////////////////////
 111     /// @brief Convert an IEEE 754 16-bit float to an 32-bit single precision
 112     ///        float
 113     /// @param val - 16-bit float
 114     /// @todo Maybe move this outside of this file into a header?
 115     static float ConvertFloat16ToFloat32(uint32_t val)
 116     {
 117         uint32_t result;
 118         if ((val & 0x7fff) == 0)
 119         {
 120             result = ((uint32_t)(val & 0x8000)) << 16;
 121         }
 122         else if ((val & 0x7c00) == 0x7c00)
 123         {
 124             result = ((val & 0x3ff) == 0) ? 0x7f800000 : 0x7fc00000;
 125             result |= ((uint32_t)val & 0x8000) << 16;
 126         }
 127         else
 128         {
 129             uint32_t sign = (val & 0x8000) << 16;
 130             uint32_t mant = (val & 0x3ff) << 13;
 131             uint32_t exp = (val >> 10) & 0x1f;
 132             if ((exp == 0) && (mant != 0)) // Adjust exponent and mantissa for denormals
 133             {
 134                 mant <<= 1;
 135                 while (mant < (0x400 << 13))
 136                 {
 137                     exp--;
 138                     mant <<= 1;
 139                 }
 140                 mant &= (0x3ff << 13);
 141             }
 142             exp = ((exp - 15 + 127) & 0xff) << 23;
 143             result = sign | exp | mant;
 144         }
 145
 146         return *(float*)&result;
 147     }
 148
 149     Constant *Builder::C(bool i)
 150     {
 151         return ConstantInt::get(IRB()->getInt1Ty(), (i ? 1 : 0));
 152     }
 153
 154     Constant *Builder::C(char i)
 155     {
 156         return ConstantInt::get(IRB()->getInt8Ty(), i);
 157     }
 158
 159     Constant *Builder::C(uint8_t i)
 160     {
 161         return ConstantInt::get(IRB()->getInt8Ty(), i);
 162     }
 163
 164     Constant *Builder::C(int i)
 165     {
 166         return ConstantInt::get(IRB()->getInt32Ty(), i);
 167     }
 168
 169     Constant *Builder::C(int64_t i)
 170     {
 171         return ConstantInt::get(IRB()->getInt64Ty(), i);
 172     }
 173
 174     Constant *Builder::C(uint16_t i)
 175     {
 176         return ConstantInt::get(mInt16Ty,i);
 177     }
 178
 179     Constant *Builder::C(uint32_t i)
 180     {
 181         return ConstantInt::get(IRB()->getInt32Ty(), i);
 182     }
 183
 184     Constant *Builder::C(float i)
 185     {
 186         return ConstantFP::get(IRB()->getFloatTy(), i);
 187     }
 188
 189     Constant *Builder::PRED(bool pred)
 190     {
 191         return ConstantInt::get(IRB()->getInt1Ty(), (pred ? 1 : 0));
 192     }
 193
 194     Value *Builder::VIMMED1(int i)
 195     {
 196         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 197     }
 198
 199     Value *Builder::VIMMED1(uint32_t i)
 200     {
 201         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 202     }
 203
 204     Value *Builder::VIMMED1(float i)
 205     {
 206         return ConstantVector::getSplat(mVWidth, cast<ConstantFP>(C(i)));
 207     }
 208
 209     Value *Builder::VIMMED1(bool i)
 210     {
 211         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 212     }
 213
 214 #if USE_SIMD16_BUILDER
 215     Value *Builder::VIMMED2_1(int i)
 216     {
 217         return ConstantVector::getSplat(mVWidth2, cast<ConstantInt>(C(i)));
 218     }
 219
 220     Value *Builder::VIMMED2_1(uint32_t i)
 221     {
 222         return ConstantVector::getSplat(mVWidth2, cast<ConstantInt>(C(i)));
 223     }
 224
 225     Value *Builder::VIMMED2_1(float i)
 226     {
 227         return ConstantVector::getSplat(mVWidth2, cast<ConstantFP>(C(i)));
 228     }
 229
 230     Value *Builder::VIMMED2_1(bool i)
 231     {
 232         return ConstantVector::getSplat(mVWidth2, cast<ConstantInt>(C(i)));
 233     }
 234
 235 #endif
 236     Value *Builder::VUNDEF_IPTR()
 237     {
 238         return UndefValue::get(VectorType::get(mInt32PtrTy,mVWidth));
 239     }
 240
 241     Value *Builder::VUNDEF_I()
 242     {
 243         return UndefValue::get(VectorType::get(mInt32Ty, mVWidth));
 244     }
 245
 246     Value *Builder::VUNDEF(Type *ty, uint32_t size)
 247     {
 248         return UndefValue::get(VectorType::get(ty, size));
 249     }
 250
 251     Value *Builder::VUNDEF_F()
 252     {
 253         return UndefValue::get(VectorType::get(mFP32Ty, mVWidth));
 254     }
 255
 256 #if USE_SIMD16_BUILDER
 257     Value *Builder::VUNDEF2_F()
 258     {
 259         return UndefValue::get(VectorType::get(mFP32Ty, mVWidth2));
 260     }
 261
 262     Value *Builder::VUNDEF2_I()
 263     {
 264         return UndefValue::get(VectorType::get(mInt32Ty, mVWidth2));
 265     }
 266
 267 #endif
 268     Value *Builder::VUNDEF(Type* t)
 269     {
 270         return UndefValue::get(VectorType::get(t, mVWidth));
 271     }
 272
 273     Value *Builder::VBROADCAST(Value *src)
 274     {
 275         // check if src is already a vector
 276         if (src->getType()->isVectorTy())
 277         {
 278             return src;
 279         }
 280
 281         return VECTOR_SPLAT(mVWidth, src);
 282     }
 283
 284 #if USE_SIMD16_BUILDER
 285     Value *Builder::VBROADCAST2(Value *src)
 286     {
 287         // check if src is already a vector
 288         if (src->getType()->isVectorTy())
 289         {
 290             return src;
 291         }
 292
 293         return VECTOR_SPLAT(mVWidth2, src);
 294     }
 295
 296 #endif
 297     uint32_t Builder::IMMED(Value* v)
 298     {
 299         SWR_ASSERT(isa<ConstantInt>(v));
 300         ConstantInt *pValConst = cast<ConstantInt>(v);
 301         return pValConst->getZExtValue();
 302     }
 303
 304     int32_t Builder::S_IMMED(Value* v)
 305     {
 306         SWR_ASSERT(isa<ConstantInt>(v));
 307         ConstantInt *pValConst = cast<ConstantInt>(v);
 308         return pValConst->getSExtValue();
 309     }
 310
 311     Value *Builder::GEP(Value* ptr, const std::initializer_list<Value*> &indexList)
 312     {
 313         std::vector<Value*> indices;
 314         for (auto i : indexList)
 315             indices.push_back(i);
 316         return GEPA(ptr, indices);
 317     }
 318
 319     Value *Builder::GEP(Value* ptr, const std::initializer_list<uint32_t> &indexList)
 320     {
 321         std::vector<Value*> indices;
 322         for (auto i : indexList)
 323             indices.push_back(C(i));
 324         return GEPA(ptr, indices);
 325     }
 326
 327     Value *Builder::IN_BOUNDS_GEP(Value* ptr, const std::initializer_list<Value*> &indexList)
 328     {
 329         std::vector<Value*> indices;
 330         for (auto i : indexList)
 331             indices.push_back(i);
 332         return IN_BOUNDS_GEP(ptr, indices);
 333     }
 334
 335     Value *Builder::IN_BOUNDS_GEP(Value* ptr, const std::initializer_list<uint32_t> &indexList)
 336     {
 337         std::vector<Value*> indices;
 338         for (auto i : indexList)
 339             indices.push_back(C(i));
 340         return IN_BOUNDS_GEP(ptr, indices);
 341     }
 342
 343     LoadInst *Builder::LOAD(Value *basePtr, const std::initializer_list<uint32_t> &indices, const llvm::Twine& name)
 344     {
 345         std::vector<Value*> valIndices;
 346         for (auto i : indices)
 347             valIndices.push_back(C(i));
 348         return LOAD(GEPA(basePtr, valIndices), name);
 349     }
 350
 351     LoadInst *Builder::LOADV(Value *basePtr, const std::initializer_list<Value*> &indices, const llvm::Twine& name)
 352     {
 353         std::vector<Value*> valIndices;
 354         for (auto i : indices)
 355             valIndices.push_back(i);
 356         return LOAD(GEPA(basePtr, valIndices), name);
 357     }
 358
 359     StoreInst *Builder::STORE(Value *val, Value *basePtr, const std::initializer_list<uint32_t> &indices)
 360     {
 361         std::vector<Value*> valIndices;
 362         for (auto i : indices)
 363             valIndices.push_back(C(i));
 364         return STORE(val, GEPA(basePtr, valIndices));
 365     }
 366
 367     StoreInst *Builder::STOREV(Value *val, Value *basePtr, const std::initializer_list<Value*> &indices)
 368     {
 369         std::vector<Value*> valIndices;
 370         for (auto i : indices)
 371             valIndices.push_back(i);
 372         return STORE(val, GEPA(basePtr, valIndices));
 373     }
 374
 375     CallInst *Builder::CALL(Value *Callee, const std::initializer_list<Value*> &argsList)
 376     {
 377         std::vector<Value*> args;
 378         for (auto arg : argsList)
 379             args.push_back(arg);
 380         return CALLA(Callee, args);
 381     }
 382
 383     CallInst *Builder::CALL(Value *Callee, Value* arg)
 384     {
 385         std::vector<Value*> args;
 386         args.push_back(arg);
 387         return CALLA(Callee, args);
 388     }
 389
 390     CallInst *Builder::CALL2(Value *Callee, Value* arg1, Value* arg2)
 391     {
 392         std::vector<Value*> args;
 393         args.push_back(arg1);
 394         args.push_back(arg2);
 395         return CALLA(Callee, args);
 396     }
 397
 398     CallInst *Builder::CALL3(Value *Callee, Value* arg1, Value* arg2, Value* arg3)
 399     {
 400         std::vector<Value*> args;
 401         args.push_back(arg1);
 402         args.push_back(arg2);
 403         args.push_back(arg3);
 404         return CALLA(Callee, args);
 405     }
 406
 407     //////////////////////////////////////////////////////////////////////////
 408     Value *Builder::DEBUGTRAP()
 409     {
 410         Function *func = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::debugtrap);
 411         return CALL(func);
 412     }
 413
 414     Value *Builder::VRCP(Value *va)
 415     {
 416         return FDIV(VIMMED1(1.0f), va);  // 1 / a
 417     }
 418
 419     Value *Builder::VPLANEPS(Value* vA, Value* vB, Value* vC, Value* &vX, Value* &vY)
 420     {
 421         Value* vOut = FMADDPS(vA, vX, vC);
 422         vOut = FMADDPS(vB, vY, vOut);
 423         return vOut;
 424     }
 425
 426     //////////////////////////////////////////////////////////////////////////
 427     /// @brief Generate an i32 masked load operation in LLVM IR.  If not
 428     /// supported on the underlying platform, emulate it with float masked load
 429     /// @param src - base address pointer for the load
 430     /// @param vMask - SIMD wide mask that controls whether to access memory load 0
 431     Value *Builder::MASKLOADD(Value* src,Value* mask)
 432     {
 433         Value* vResult;
 434         // use avx2 gather instruction is available
 435         if(JM()->mArch.AVX2())
 436         {
 437             Function *func = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_avx2_maskload_d_256);
 438             vResult = CALL(func,{src,mask});
 439         }
 440         else
 441         {
 442             // maskload intrinsic expects integer mask operand in llvm >= 3.8
 443     #if (LLVM_VERSION_MAJOR > 3) || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 8)
 444             mask = BITCAST(mask,VectorType::get(mInt32Ty,mVWidth));
 445     #else
 446             mask = BITCAST(mask,VectorType::get(mFP32Ty,mVWidth));
 447     #endif
 448             Function *func = Intrinsic::getDeclaration(JM()->mpCurrentModule,Intrinsic::x86_avx_maskload_ps_256);
 449             vResult = BITCAST(CALL(func,{src,mask}), VectorType::get(mInt32Ty,mVWidth));
 450         }
 451         return vResult;
 452     }
 453
 454     //////////////////////////////////////////////////////////////////////////
 455     /// @brief insert a JIT call to CallPrint
 456     /// - outputs formatted string to both stdout and VS output window
 457     /// - DEBUG builds only
 458     /// Usage example:
 459     ///   PRINT("index %d = 0x%p\n",{C(lane), pIndex});
 460     ///   where C(lane) creates a constant value to print, and pIndex is the Value*
 461     ///   result from a GEP, printing out the pointer to memory
 462     /// @param printStr - constant string to print, which includes format specifiers
 463     /// @param printArgs - initializer list of Value*'s to print to std out
 464     CallInst *Builder::PRINT(const std::string &printStr,const std::initializer_list<Value*> &printArgs)
 465     {
 466         // push the arguments to CallPrint into a vector
 467         std::vector<Value*> printCallArgs;
 468         // save room for the format string.  we still need to modify it for vectors
 469         printCallArgs.resize(1);
 470
 471         // search through the format string for special processing
 472         size_t pos = 0;
 473         std::string tempStr(printStr);
 474         pos = tempStr.find('%', pos);
 475         auto v = printArgs.begin();
 476
 477         while ((pos != std::string::npos) && (v != printArgs.end()))
 478         {
 479             Value* pArg = *v;
 480             Type* pType = pArg->getType();
 481
 482             if (pType->isVectorTy())
 483             {
 484                 Type* pContainedType = pType->getContainedType(0);
 485
 486                 if (toupper(tempStr[pos + 1]) == 'X')
 487                 {
 488                     tempStr[pos] = '0';
 489                     tempStr[pos + 1] = 'x';
 490                     tempStr.insert(pos + 2, "%08X ");
 491                     pos += 7;
 492
 493                     printCallArgs.push_back(VEXTRACT(pArg, C(0)));
 494
 495                     std::string vectorFormatStr;
 496                     for (uint32_t i = 1; i < pType->getVectorNumElements(); ++i)
 497                     {
 498                         vectorFormatStr += "0x%08X ";
 499                         printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 500                     }
 501
 502                     tempStr.insert(pos, vectorFormatStr);
 503                     pos += vectorFormatStr.size();
 504                 }
 505                 else if ((tempStr[pos + 1] == 'f') && (pContainedType->isFloatTy()))
 506                 {
 507                     uint32_t i = 0;
 508                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 509                     {
 510                         tempStr.insert(pos, std::string("%f "));
 511                         pos += 3;
 512                         printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 513                     }
 514                     printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 515                 }
 516                 else if ((tempStr[pos + 1] == 'd') && (pContainedType->isIntegerTy()))
 517                 {
 518                     uint32_t i = 0;
 519                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 520                     {
 521                         tempStr.insert(pos, std::string("%d "));
 522                         pos += 3;
 523                         printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 524                     }
 525                     printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 526                 }
 527             }
 528             else
 529             {
 530                 if (toupper(tempStr[pos + 1]) == 'X')
 531                 {
 532                     tempStr[pos] = '0';
 533                     tempStr.insert(pos + 1, "x%08");
 534                     printCallArgs.push_back(pArg);
 535                     pos += 3;
 536                 }
 537                 // for %f we need to cast float Values to doubles so that they print out correctly
 538                 else if ((tempStr[pos + 1] == 'f') && (pType->isFloatTy()))
 539                 {
 540                     printCallArgs.push_back(FP_EXT(pArg, Type::getDoubleTy(JM()->mContext)));
 541                     pos++;
 542                 }
 543                 else
 544                 {
 545                     printCallArgs.push_back(pArg);
 546                 }
 547             }
 548
 549             // advance to the next arguement
 550             v++;
 551             pos = tempStr.find('%', ++pos);
 552         }
 553
 554         // create global variable constant string
 555         Constant *constString = ConstantDataArray::getString(JM()->mContext,tempStr,true);
 556         GlobalVariable *gvPtr = new GlobalVariable(constString->getType(),true,GlobalValue::InternalLinkage,constString,"printStr");
 557         JM()->mpCurrentModule->getGlobalList().push_back(gvPtr);
 558
 559         // get a pointer to the first character in the constant string array
 560         std::vector<Constant*> geplist{C(0),C(0)};
 561         Constant *strGEP = ConstantExpr::getGetElementPtr(nullptr, gvPtr,geplist,false);
 562
 563         // insert the pointer to the format string in the argument vector
 564         printCallArgs[0] = strGEP;
 565
 566         // get pointer to CallPrint function and insert decl into the module if needed
 567         std::vector<Type*> args;
 568         args.push_back(PointerType::get(mInt8Ty,0));
 569         FunctionType* callPrintTy = FunctionType::get(Type::getVoidTy(JM()->mContext),args,true);
 570         Function *callPrintFn = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("CallPrint", callPrintTy));
 571
 572         // if we haven't yet added the symbol to the symbol table
 573         if((sys::DynamicLibrary::SearchForAddressOfSymbol("CallPrint")) == nullptr)
 574         {
 575             sys::DynamicLibrary::AddSymbol("CallPrint", (void *)&CallPrint);
 576         }
 577
 578         // insert a call to CallPrint
 579         return CALLA(callPrintFn,printCallArgs);
 580     }
 581
 582     //////////////////////////////////////////////////////////////////////////
 583     /// @brief Wrapper around PRINT with initializer list.
 584     CallInst* Builder::PRINT(const std::string &printStr)
 585     {
 586         return PRINT(printStr, {});
 587     }
 588
 589     //////////////////////////////////////////////////////////////////////////
 590     /// @brief Generate a masked gather operation in LLVM IR.  If not
 591     /// supported on the underlying platform, emulate it with loads
 592     /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
 593     /// @param pBase - Int8* base VB address pointer value
 594     /// @param vIndices - SIMD wide value of VB byte offsets
 595     /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
 596     /// @param scale - value to scale indices by
 597     Value *Builder::GATHERPS(Value *vSrc, Value *pBase, Value *vIndices, Value *vMask, uint8_t scale)
 598     {
 599         Value *vGather;
 600
 601         // use avx2 gather instruction if available
 602         if(JM()->mArch.AVX2())
 603         {
 604             // force mask to <N x float>, required by vgather
 605             Value *mask = BITCAST(vMask, mSimdFP32Ty);
 606
 607             vGather = VGATHERPS(vSrc, pBase, vIndices, mask, C(scale));
 608         }
 609         else
 610         {
 611             Value* pStack = STACKSAVE();
 612
 613             // store vSrc on the stack.  this way we can select between a valid load address and the vSrc address
 614             Value* vSrcPtr = ALLOCA(vSrc->getType());
 615             STORE(vSrc, vSrcPtr);
 616
 617             vGather = VUNDEF_F();
 618             Value *vScaleVec = VIMMED1((uint32_t)scale);
 619             Value *vOffsets = MUL(vIndices,vScaleVec);
 620             Value *mask = MASK(vMask);
 621             for(uint32_t i = 0; i < mVWidth; ++i)
 622             {
 623                 // single component byte index
 624                 Value *offset = VEXTRACT(vOffsets,C(i));
 625                 // byte pointer to component
 626                 Value *loadAddress = GEP(pBase,offset);
 627                 loadAddress = BITCAST(loadAddress,PointerType::get(mFP32Ty,0));
 628                 // pointer to the value to load if we're masking off a component
 629                 Value *maskLoadAddress = GEP(vSrcPtr,{C(0), C(i)});
 630                 Value *selMask = VEXTRACT(mask,C(i));
 631                 // switch in a safe address to load if we're trying to access a vertex
 632                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
 633                 Value *val = LOAD(validAddress);
 634                 vGather = VINSERT(vGather,val,C(i));
 635             }
 636             STACKRESTORE(pStack);
 637         }
 638
 639         return vGather;
 640     }
 641
 642 #if USE_SIMD16_BUILDER
 643     Value *Builder::GATHERPS2(Value *vSrc, Value *pBase, Value *vIndices, Value *vMask, uint8_t scale)
 644     {
 645         Value *vGather = VUNDEF2_F();
 646
 647         // use avx512 gather instruction if available
 648         if (JM()->mArch.AVX512F())
 649         {
 650             // force mask to <N-bit Integer>, required by vgather2
 651             Value *mask = BITCAST(MASK2(vMask), mInt16Ty);
 652
 653             vGather = VGATHERPS2(vSrc, pBase, vIndices, mask, C((uint32_t)scale));
 654         }
 655         else
 656         {
 657             Value *src0 = EXTRACT2_F(vSrc, 0);
 658             Value *src1 = EXTRACT2_F(vSrc, 1);
 659
 660             Value *indices0 = EXTRACT2_I(vIndices, 0);
 661             Value *indices1 = EXTRACT2_I(vIndices, 1);
 662
 663             Value *mask0 = EXTRACT2_I(vMask, 0);
 664             Value *mask1 = EXTRACT2_I(vMask, 1);
 665
 666             Value *gather0 = GATHERPS(src0, pBase, indices0, mask0, scale);
 667             Value *gather1 = GATHERPS(src1, pBase, indices1, mask1, scale);
 668
 669             vGather = INSERT2_F(vGather, gather0, 0);
 670             vGather = INSERT2_F(vGather, gather1, 1);
 671         }
 672
 673         return vGather;
 674     }
 675
 676 #endif
 677     //////////////////////////////////////////////////////////////////////////
 678     /// @brief Generate a masked gather operation in LLVM IR.  If not
 679     /// supported on the underlying platform, emulate it with loads
 680     /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
 681     /// @param pBase - Int8* base VB address pointer value
 682     /// @param vIndices - SIMD wide value of VB byte offsets
 683     /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
 684     /// @param scale - value to scale indices by
 685     Value *Builder::GATHERDD(Value* vSrc, Value* pBase, Value* vIndices, Value* vMask, uint8_t scale)
 686     {
 687         Value* vGather;
 688
 689         // use avx2 gather instruction if available
 690         if(JM()->mArch.AVX2())
 691         {
 692             vGather = VGATHERDD(vSrc, pBase, vIndices, vMask, C(scale));
 693         }
 694         else
 695         {
 696             Value* pStack = STACKSAVE();
 697
 698             // store vSrc on the stack.  this way we can select between a valid load address and the vSrc address
 699             Value* vSrcPtr = ALLOCA(vSrc->getType());
 700             STORE(vSrc, vSrcPtr);
 701
 702             vGather = VUNDEF_I();
 703             Value *vScaleVec = VIMMED1((uint32_t)scale);
 704             Value *vOffsets = MUL(vIndices, vScaleVec);
 705             Value *mask = MASK(vMask);
 706             for(uint32_t i = 0; i < mVWidth; ++i)
 707             {
 708                 // single component byte index
 709                 Value *offset = VEXTRACT(vOffsets, C(i));
 710                 // byte pointer to component
 711                 Value *loadAddress = GEP(pBase, offset);
 712                 loadAddress = BITCAST(loadAddress, PointerType::get(mInt32Ty, 0));
 713                 // pointer to the value to load if we're masking off a component
 714                 Value *maskLoadAddress = GEP(vSrcPtr, {C(0), C(i)});
 715                 Value *selMask = VEXTRACT(mask, C(i));
 716                 // switch in a safe address to load if we're trying to access a vertex
 717                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
 718                 Value *val = LOAD(validAddress, C(0));
 719                 vGather = VINSERT(vGather, val, C(i));
 720             }
 721
 722             STACKRESTORE(pStack);
 723         }
 724         return vGather;
 725     }
 726
 727     //////////////////////////////////////////////////////////////////////////
 728     /// @brief Generate a masked gather operation in LLVM IR.  If not
 729     /// supported on the underlying platform, emulate it with loads
 730     /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
 731     /// @param pBase - Int8* base VB address pointer value
 732     /// @param vIndices - SIMD wide value of VB byte offsets
 733     /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
 734     /// @param scale - value to scale indices by
 735     Value *Builder::GATHERPD(Value* vSrc, Value* pBase, Value* vIndices, Value* vMask, uint8_t scale)
 736     {
 737         Value* vGather;
 738
 739         // use avx2 gather instruction if available
 740         if(JM()->mArch.AVX2())
 741         {
 742             vGather = VGATHERPD(vSrc, pBase, vIndices, vMask, C(scale));
 743         }
 744         else
 745         {
 746             Value* pStack = STACKSAVE();
 747
 748             // store vSrc on the stack.  this way we can select between a valid load address and the vSrc address
 749             Value* vSrcPtr = ALLOCA(vSrc->getType());
 750             STORE(vSrc, vSrcPtr);
 751
 752             vGather = UndefValue::get(VectorType::get(mDoubleTy, 4));
 753             Value *vScaleVec = VECTOR_SPLAT(4, C((uint32_t)scale));
 754             Value *vOffsets = MUL(vIndices,vScaleVec);
 755             Value *mask = MASK(vMask);
 756             for(uint32_t i = 0; i < mVWidth/2; ++i)
 757             {
 758                 // single component byte index
 759                 Value *offset = VEXTRACT(vOffsets,C(i));
 760                 // byte pointer to component
 761                 Value *loadAddress = GEP(pBase,offset);
 762                 loadAddress = BITCAST(loadAddress,PointerType::get(mDoubleTy,0));
 763                 // pointer to the value to load if we're masking off a component
 764                 Value *maskLoadAddress = GEP(vSrcPtr,{C(0), C(i)});
 765                 Value *selMask = VEXTRACT(mask,C(i));
 766                 // switch in a safe address to load if we're trying to access a vertex
 767                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
 768                 Value *val = LOAD(validAddress);
 769                 vGather = VINSERT(vGather,val,C(i));
 770             }
 771             STACKRESTORE(pStack);
 772         }
 773         return vGather;
 774     }
 775
 776 #if USE_SIMD16_BUILDER
 777     //////////////////////////////////////////////////////////////////////////
 778     /// @brief
 779     Value *Builder::EXTRACT2_F(Value *a2, uint32_t imm)
 780     {
 781         const uint32_t i0 = (imm > 0) ? mVWidth : 0;
 782
 783         Value *result = VUNDEF_F();
 784
 785         for (uint32_t i = 0; i < mVWidth; i += 1)
 786         {
 787 #if 1
 788             if (!a2->getType()->getScalarType()->isFloatTy())
 789             {
 790                 a2 = BITCAST(a2, mSimd2FP32Ty);
 791             }
 792
 793 #endif
 794             Value *temp = VEXTRACT(a2, C(i0 + i));
 795
 796             result = VINSERT(result, temp, C(i));
 797         }
 798
 799         return result;
 800     }
 801
 802     Value *Builder::EXTRACT2_I(Value *a2, uint32_t imm)
 803     {
 804         return BITCAST(EXTRACT2_F(a2, imm), mSimdInt32Ty);
 805     }
 806
 807     //////////////////////////////////////////////////////////////////////////
 808     /// @brief
 809     Value *Builder::INSERT2_F(Value *a2, Value *b, uint32_t imm)
 810     {
 811         const uint32_t i0 = (imm > 0) ? mVWidth : 0;
 812
 813         Value *result = BITCAST(a2, mSimd2FP32Ty);
 814
 815         for (uint32_t i = 0; i < mVWidth; i += 1)
 816         {
 817 #if 1
 818             if (!b->getType()->getScalarType()->isFloatTy())
 819             {
 820                 b = BITCAST(b, mSimdFP32Ty);
 821             }
 822
 823 #endif
 824             Value *temp = VEXTRACT(b, C(i));
 825
 826             result = VINSERT(result, temp, C(i0 + i));
 827         }
 828
 829         return result;
 830     }
 831
 832     Value *Builder::INSERT2_I(Value *a2, Value *b, uint32_t imm)
 833     {
 834         return BITCAST(INSERT2_F(a2, b, imm), mSimd2Int32Ty);
 835     }
 836
 837 #endif
 838     //////////////////////////////////////////////////////////////////////////
 839     /// @brief convert x86 <N x float> mask to llvm <N x i1> mask
 840     Value *Builder::MASK(Value *vmask)
 841     {
 842         Value *src = BITCAST(vmask, mSimdInt32Ty);
 843         return ICMP_SLT(src, VIMMED1(0));
 844     }
 845
 846 #if USE_SIMD16_BUILDER
 847     Value *Builder::MASK2(Value *vmask)
 848     {
 849         Value *src = BITCAST(vmask, mSimd2Int32Ty);
 850         return ICMP_SLT(src, VIMMED2_1(0));
 851     }
 852
 853 #endif
 854     //////////////////////////////////////////////////////////////////////////
 855     /// @brief convert llvm <N x i1> mask to x86 <N x i32> mask
 856     Value *Builder::VMASK(Value *mask)
 857     {
 858         return S_EXT(mask, mSimdInt32Ty);
 859     }
 860
 861 #if USE_SIMD16_BUILDER
 862     Value *Builder::VMASK2(Value *mask)
 863     {
 864         return S_EXT(mask, mSimd2Int32Ty);
 865     }
 866
 867 #endif
 868     //////////////////////////////////////////////////////////////////////////
 869     /// @brief Generate a VPSHUFB operation in LLVM IR.  If not
 870     /// supported on the underlying platform, emulate it
 871     /// @param a - 256bit SIMD(32x8bit) of 8bit integer values
 872     /// @param b - 256bit SIMD(32x8bit) of 8bit integer mask values
 873     /// Byte masks in lower 128 lane of b selects 8 bit values from lower
 874     /// 128bits of a, and vice versa for the upper lanes.  If the mask
 875     /// value is negative, '0' is inserted.
 876     Value *Builder::PSHUFB(Value* a, Value* b)
 877     {
 878         Value* res;
 879         // use avx2 pshufb instruction if available
 880         if(JM()->mArch.AVX2())
 881         {
 882             res = VPSHUFB(a, b);
 883         }
 884         else
 885         {
 886             Constant* cB = dyn_cast<Constant>(b);
 887             // number of 8 bit elements in b
 888             uint32_t numElms = cast<VectorType>(cB->getType())->getNumElements();
 889             // output vector
 890             Value* vShuf = UndefValue::get(VectorType::get(mInt8Ty, numElms));
 891
 892             // insert an 8 bit value from the high and low lanes of a per loop iteration
 893             numElms /= 2;
 894             for(uint32_t i = 0; i < numElms; i++)
 895             {
 896                 ConstantInt* cLow128b = cast<ConstantInt>(cB->getAggregateElement(i));
 897                 ConstantInt* cHigh128b = cast<ConstantInt>(cB->getAggregateElement(i + numElms));
 898
 899                 // extract values from constant mask
 900                 char valLow128bLane =  (char)(cLow128b->getSExtValue());
 901                 char valHigh128bLane = (char)(cHigh128b->getSExtValue());
 902
 903                 Value* insertValLow128b;
 904                 Value* insertValHigh128b;
 905
 906                 // if the mask value is negative, insert a '0' in the respective output position
 907                 // otherwise, lookup the value at mask position (bits 3..0 of the respective mask byte) in a and insert in output vector
 908                 insertValLow128b = (valLow128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valLow128bLane & 0xF)));
 909                 insertValHigh128b = (valHigh128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valHigh128bLane & 0xF) + numElms));
 910
 911                 vShuf = VINSERT(vShuf, insertValLow128b, i);
 912                 vShuf = VINSERT(vShuf, insertValHigh128b, (i + numElms));
 913             }
 914             res = vShuf;
 915         }
 916         return res;
 917     }
 918
 919     //////////////////////////////////////////////////////////////////////////
 920     /// @brief Generate a VPSHUFB operation (sign extend 8 8bit values to 32
 921     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 922     /// @param a - 128bit SIMD lane(16x8bit) of 8bit integer values.  Only
 923     /// lower 8 values are used.
 924     Value *Builder::PMOVSXBD(Value* a)
 925     {
 926         // VPMOVSXBD output type
 927         Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 928         // Extract 8 values from 128bit lane and sign extend
 929         return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 930     }
 931
 932     //////////////////////////////////////////////////////////////////////////
 933     /// @brief Generate a VPSHUFB operation (sign extend 8 16bit values to 32
 934     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 935     /// @param a - 128bit SIMD lane(8x16bit) of 16bit integer values.
 936     Value *Builder::PMOVSXWD(Value* a)
 937     {
 938         // VPMOVSXWD output type
 939         Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 940         // Extract 8 values from 128bit lane and sign extend
 941         return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 942     }
 943
 944     //////////////////////////////////////////////////////////////////////////
 945     /// @brief Generate a VPERMD operation (shuffle 32 bit integer values
 946     /// across 128 bit lanes) in LLVM IR.  If not supported on the underlying
 947     /// platform, emulate it
 948     /// @param a - 256bit SIMD lane(8x32bit) of integer values.
 949     /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
 950     Value *Builder::PERMD(Value* a, Value* idx)
 951     {
 952         Value* res;
 953         // use avx2 permute instruction if available
 954         if(JM()->mArch.AVX2())
 955         {
 956             res = VPERMD(a, idx);
 957         }
 958         else
 959         {
 960             if (isa<Constant>(idx))
 961             {
 962                 res = VSHUFFLE(a, a, idx);
 963             }
 964             else
 965             {
 966                 res = VUNDEF_I();
 967                 for (uint32_t l = 0; l < JM()->mVWidth; ++l)
 968                 {
 969                     Value* pIndex = VEXTRACT(idx, C(l));
 970                     Value* pVal = VEXTRACT(a, pIndex);
 971                     res = VINSERT(res, pVal, C(l));
 972                 }
 973             }
 974         }
 975         return res;
 976     }
 977
 978     //////////////////////////////////////////////////////////////////////////
 979     /// @brief Generate a VPERMPS operation (shuffle 32 bit float values
 980     /// across 128 bit lanes) in LLVM IR.  If not supported on the underlying
 981     /// platform, emulate it
 982     /// @param a - 256bit SIMD lane(8x32bit) of float values.
 983     /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
 984     Value *Builder::PERMPS(Value* a, Value* idx)
 985     {
 986         Value* res;
 987         // use avx2 permute instruction if available
 988         if (JM()->mArch.AVX2())
 989         {
 990             // llvm 3.6.0 swapped the order of the args to vpermd
 991             res = VPERMPS(idx, a);
 992         }
 993         else
 994         {
 995             if (isa<Constant>(idx))
 996             {
 997                 res = VSHUFFLE(a, a, idx);
 998             }
 999             else
1000             {
1001                 res = VUNDEF_F();
1002                 for (uint32_t l = 0; l < JM()->mVWidth; ++l)
1003                 {
1004                     Value* pIndex = VEXTRACT(idx, C(l));
1005                     Value* pVal = VEXTRACT(a, pIndex);
1006                     res = VINSERT(res, pVal, C(l));
1007                 }
1008             }
1009         }
1010
1011         return res;
1012     }
1013
1014     //////////////////////////////////////////////////////////////////////////
1015     /// @brief Generate a VCVTPH2PS operation (float16->float32 conversion)
1016     /// in LLVM IR.  If not supported on the underlying platform, emulate it
1017     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
1018     Value *Builder::CVTPH2PS(Value* a)
1019     {
1020         if (JM()->mArch.F16C())
1021         {
1022             return VCVTPH2PS(a);
1023         }
1024         else
1025         {
1026             FunctionType* pFuncTy = FunctionType::get(mFP32Ty, mInt16Ty);
1027             Function* pCvtPh2Ps = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("ConvertFloat16ToFloat32", pFuncTy));
1028
1029             if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat16ToFloat32") == nullptr)
1030             {
1031                 sys::DynamicLibrary::AddSymbol("ConvertFloat16ToFloat32", (void *)&ConvertFloat16ToFloat32);
1032             }
1033
1034             Value* pResult = UndefValue::get(mSimdFP32Ty);
1035             for (uint32_t i = 0; i < mVWidth; ++i)
1036             {
1037                 Value* pSrc = VEXTRACT(a, C(i));
1038                 Value* pConv = CALL(pCvtPh2Ps, std::initializer_list<Value*>{pSrc});
1039                 pResult = VINSERT(pResult, pConv, C(i));
1040             }
1041
1042             return pResult;
1043         }
1044     }
1045
1046     //////////////////////////////////////////////////////////////////////////
1047     /// @brief Generate a VCVTPS2PH operation (float32->float16 conversion)
1048     /// in LLVM IR.  If not supported on the underlying platform, emulate it
1049     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
1050     Value *Builder::CVTPS2PH(Value* a, Value* rounding)
1051     {
1052         if (JM()->mArch.F16C())
1053         {
1054             return VCVTPS2PH(a, rounding);
1055         }
1056         else
1057         {
1058             // call scalar C function for now
1059             FunctionType* pFuncTy = FunctionType::get(mInt16Ty, mFP32Ty);
1060             Function* pCvtPs2Ph = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("ConvertFloat32ToFloat16", pFuncTy));
1061
1062             if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat32ToFloat16") == nullptr)
1063             {
1064                 sys::DynamicLibrary::AddSymbol("ConvertFloat32ToFloat16", (void *)&ConvertFloat32ToFloat16);
1065             }
1066
1067             Value* pResult = UndefValue::get(mSimdInt16Ty);
1068             for (uint32_t i = 0; i < mVWidth; ++i)
1069             {
1070                 Value* pSrc = VEXTRACT(a, C(i));
1071                 Value* pConv = CALL(pCvtPs2Ph, std::initializer_list<Value*>{pSrc});
1072                 pResult = VINSERT(pResult, pConv, C(i));
1073             }
1074
1075             return pResult;
1076         }
1077     }
1078
1079     Value *Builder::PMAXSD(Value* a, Value* b)
1080     {
1081         Value* cmp = ICMP_SGT(a, b);
1082         return SELECT(cmp, a, b);
1083     }
1084
1085     Value *Builder::PMINSD(Value* a, Value* b)
1086     {
1087         Value* cmp = ICMP_SLT(a, b);
1088         return SELECT(cmp, a, b);
1089     }
1090
1091     void Builder::Gather4(const SWR_FORMAT format, Value* pSrcBase, Value* byteOffsets,
1092                           Value* mask, Value* vGatherComponents[], bool bPackedOutput)
1093     {
1094         const SWR_FORMAT_INFO &info = GetFormatInfo(format);
1095         if(info.type[0] == SWR_TYPE_FLOAT && info.bpc[0] == 32)
1096         {
1097             // ensure our mask is the correct type
1098             mask = BITCAST(mask, mSimdFP32Ty);
1099             GATHER4PS(info, pSrcBase, byteOffsets, mask, vGatherComponents, bPackedOutput);
1100         }
1101         else
1102         {
1103             // ensure our mask is the correct type
1104             mask = BITCAST(mask, mSimdInt32Ty);
1105             GATHER4DD(info, pSrcBase, byteOffsets, mask, vGatherComponents, bPackedOutput);
1106         }
1107     }
1108
1109     void Builder::GATHER4PS(const SWR_FORMAT_INFO &info, Value* pSrcBase, Value* byteOffsets,
1110                             Value* vMask, Value* vGatherComponents[], bool bPackedOutput)
1111     {
1112         switch(info.bpp / info.numComps)
1113         {
1114             case 16:
1115             {
1116                     Value* vGatherResult[2];
1117
1118                     // TODO: vGatherMaskedVal
1119                     Value* vGatherMaskedVal = VIMMED1((float)0);
1120
1121                     // always have at least one component out of x or y to fetch
1122
1123                     vGatherResult[0] = GATHERPS(vGatherMaskedVal, pSrcBase, byteOffsets, vMask);
1124                     // e.g. result of first 8x32bit integer gather for 16bit components
1125                     // 256i - 0    1    2    3    4    5    6    7
1126                     //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1127                     //
1128
1129                     // if we have at least one component out of x or y to fetch
1130                     if(info.numComps > 2)
1131                     {
1132                         // offset base to the next components(zw) in the vertex to gather
1133                         pSrcBase = GEP(pSrcBase, C((char)4));
1134
1135                         vGatherResult[1] =  GATHERPS(vGatherMaskedVal, pSrcBase, byteOffsets, vMask);
1136                         // e.g. result of second 8x32bit integer gather for 16bit components
1137                         // 256i - 0    1    2    3    4    5    6    7
1138                         //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1139                         //
1140                     }
1141                     else
1142                     {
1143                         vGatherResult[1] =  vGatherMaskedVal;
1144                     }
1145
1146                     // Shuffle gathered components into place, each row is a component
1147                     Shuffle16bpcGather4(info, vGatherResult, vGatherComponents, bPackedOutput);
1148             }
1149                 break;
1150             case 32:
1151             {
1152                 // apply defaults
1153                 for (uint32_t i = 0; i < 4; ++i)
1154                 {
1155                     vGatherComponents[i] = VIMMED1(*(float*)&info.defaults[i]);
1156                 }
1157
1158                 for(uint32_t i = 0; i < info.numComps; i++)
1159                 {
1160                     uint32_t swizzleIndex = info.swizzle[i];
1161
1162                     // Gather a SIMD of components
1163                     vGatherComponents[swizzleIndex] = GATHERPS(vGatherComponents[swizzleIndex], pSrcBase, byteOffsets, vMask);
1164
1165                     // offset base to the next component to gather
1166                     pSrcBase = GEP(pSrcBase, C((char)4));
1167                 }
1168             }
1169                 break;
1170             default:
1171                 SWR_INVALID("Invalid float format");
1172                 break;
1173         }
1174     }
1175
1176     void Builder::GATHER4DD(const SWR_FORMAT_INFO &info, Value* pSrcBase, Value* byteOffsets,
1177                             Value* vMask, Value* vGatherComponents[], bool bPackedOutput)
1178     {
1179         switch (info.bpp / info.numComps)
1180         {
1181             case 8:
1182             {
1183                 Value* vGatherMaskedVal = VIMMED1((int32_t)0);
1184                 Value* vGatherResult = GATHERDD(vGatherMaskedVal, pSrcBase, byteOffsets, vMask);
1185                 // e.g. result of an 8x32bit integer gather for 8bit components
1186                 // 256i - 0    1    2    3    4    5    6    7
1187                 //        xyzw xyzw xyzw xyzw xyzw xyzw xyzw xyzw
1188
1189                 Shuffle8bpcGather4(info, vGatherResult, vGatherComponents, bPackedOutput);
1190             }
1191                 break;
1192             case 16:
1193             {
1194                 Value* vGatherResult[2];
1195
1196                 // TODO: vGatherMaskedVal
1197                 Value* vGatherMaskedVal = VIMMED1((int32_t)0);
1198
1199                 // always have at least one component out of x or y to fetch
1200
1201                 vGatherResult[0] = GATHERDD(vGatherMaskedVal, pSrcBase, byteOffsets, vMask);
1202                 // e.g. result of first 8x32bit integer gather for 16bit components
1203                 // 256i - 0    1    2    3    4    5    6    7
1204                 //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1205                 //
1206
1207                 // if we have at least one component out of x or y to fetch
1208                 if(info.numComps > 2)
1209                 {
1210                     // offset base to the next components(zw) in the vertex to gather
1211                     pSrcBase = GEP(pSrcBase, C((char)4));
1212
1213                     vGatherResult[1] = GATHERDD(vGatherMaskedVal, pSrcBase, byteOffsets, vMask);
1214                     // e.g. result of second 8x32bit integer gather for 16bit components
1215                     // 256i - 0    1    2    3    4    5    6    7
1216                     //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1217                     //
1218                 }
1219                 else
1220                 {
1221                     vGatherResult[1] = vGatherMaskedVal;
1222                 }
1223
1224                 // Shuffle gathered components into place, each row is a component
1225                 Shuffle16bpcGather4(info, vGatherResult, vGatherComponents, bPackedOutput);
1226
1227             }
1228                 break;
1229             case 32:
1230             {
1231                 // apply defaults
1232                 for (uint32_t i = 0; i < 4; ++i)
1233                 {
1234                     vGatherComponents[i] = VIMMED1((int)info.defaults[i]);
1235                 }
1236
1237                 for(uint32_t i = 0; i < info.numComps; i++)
1238                 {
1239                     uint32_t swizzleIndex = info.swizzle[i];
1240
1241                     // Gather a SIMD of components
1242                     vGatherComponents[swizzleIndex] = GATHERDD(vGatherComponents[swizzleIndex], pSrcBase, byteOffsets, vMask);
1243
1244                     // offset base to the next component to gather
1245                     pSrcBase = GEP(pSrcBase, C((char)4));
1246                 }
1247             }
1248                 break;
1249             default:
1250                 SWR_INVALID("unsupported format");
1251             break;
1252         }
1253     }
1254
1255     void Builder::Shuffle16bpcGather4(const SWR_FORMAT_INFO &info, Value* vGatherInput[2], Value* vGatherOutput[4], bool bPackedOutput)
1256     {
1257         // cast types
1258         Type* vGatherTy = VectorType::get(IntegerType::getInt32Ty(JM()->mContext), mVWidth);
1259         Type* v32x8Ty = VectorType::get(mInt8Ty, mVWidth * 4); // vwidth is units of 32 bits
1260
1261         // input could either be float or int vector; do shuffle work in int
1262         vGatherInput[0] = BITCAST(vGatherInput[0], mSimdInt32Ty);
1263         vGatherInput[1] = BITCAST(vGatherInput[1], mSimdInt32Ty);
1264
1265         if(bPackedOutput)
1266         {
1267             Type* v128bitTy = VectorType::get(IntegerType::getIntNTy(JM()->mContext, 128), mVWidth / 4); // vwidth is units of 32 bits
1268
1269             // shuffle mask
1270             Value* vConstMask = C<char>({0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15,
1271                                          0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15});
1272             Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherInput[0], v32x8Ty), vConstMask), vGatherTy);
1273             // after pshufb: group components together in each 128bit lane
1274             // 256i - 0    1    2    3    4    5    6    7
1275             //        xxxx xxxx yyyy yyyy xxxx xxxx yyyy yyyy
1276
1277             Value* vi128XY = BITCAST(PERMD(vShufResult, C<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy);
1278             // after PERMD: move and pack xy components into each 128bit lane
1279             // 256i - 0    1    2    3    4    5    6    7
1280             //        xxxx xxxx xxxx xxxx yyyy yyyy yyyy yyyy
1281
1282             // do the same for zw components
1283             Value* vi128ZW = nullptr;
1284             if(info.numComps > 2)
1285             {
1286                 Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherInput[1], v32x8Ty), vConstMask), vGatherTy);
1287                 vi128ZW = BITCAST(PERMD(vShufResult, C<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy);
1288             }
1289
1290             for(uint32_t i = 0; i < 4; i++)
1291             {
1292                 uint32_t swizzleIndex = info.swizzle[i];
1293                 // todo: fixed for packed
1294                 Value* vGatherMaskedVal = VIMMED1((int32_t)(info.defaults[i]));
1295                 if(i >= info.numComps)
1296                 {
1297                     // set the default component val
1298                     vGatherOutput[swizzleIndex] = vGatherMaskedVal;
1299                     continue;
1300                 }
1301
1302                 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1303                 uint32_t lane = ((i == 0) || (i == 2)) ? 0 : 1;
1304                 // if x or y, use vi128XY permute result, else use vi128ZW
1305                 Value* selectedPermute = (i < 2) ? vi128XY : vi128ZW;
1306
1307                 // extract packed component 128 bit lanes
1308                 vGatherOutput[swizzleIndex] = VEXTRACT(selectedPermute, C(lane));
1309             }
1310
1311         }
1312         else
1313         {
1314             // pshufb masks for each component
1315             Value* vConstMask[2];
1316             // x/z shuffle mask
1317             vConstMask[0] = C<char>({0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1,
1318                                      0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1, });
1319
1320             // y/w shuffle mask
1321             vConstMask[1] = C<char>({2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1,
1322                                      2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1});
1323
1324
1325             // shuffle enabled components into lower word of each 32bit lane, 0 extending to 32 bits
1326             // apply defaults
1327             for (uint32_t i = 0; i < 4; ++i)
1328             {
1329                 vGatherOutput[i] = VIMMED1((int32_t)info.defaults[i]);
1330             }
1331
1332             for(uint32_t i = 0; i < info.numComps; i++)
1333             {
1334                 uint32_t swizzleIndex = info.swizzle[i];
1335
1336                 // select correct constMask for x/z or y/w pshufb
1337                 uint32_t selectedMask = ((i == 0) || (i == 2)) ? 0 : 1;
1338                 // if x or y, use vi128XY permute result, else use vi128ZW
1339                 uint32_t selectedGather = (i < 2) ? 0 : 1;
1340
1341                 vGatherOutput[swizzleIndex] = BITCAST(PSHUFB(BITCAST(vGatherInput[selectedGather], v32x8Ty), vConstMask[selectedMask]), vGatherTy);
1342                 // after pshufb mask for x channel; z uses the same shuffle from the second gather
1343                 // 256i - 0    1    2    3    4    5    6    7
1344                 //        xx00 xx00 xx00 xx00 xx00 xx00 xx00 xx00
1345             }
1346         }
1347     }
1348
1349     void Builder::Shuffle8bpcGather4(const SWR_FORMAT_INFO &info, Value* vGatherInput, Value* vGatherOutput[], bool bPackedOutput)
1350     {
1351         // cast types
1352         Type* vGatherTy = VectorType::get(IntegerType::getInt32Ty(JM()->mContext), mVWidth);
1353         Type* v32x8Ty =  VectorType::get(mInt8Ty, mVWidth * 4 ); // vwidth is units of 32 bits
1354
1355         if(bPackedOutput)
1356         {
1357             Type* v128Ty = VectorType::get(IntegerType::getIntNTy(JM()->mContext, 128), mVWidth / 4); // vwidth is units of 32 bits
1358             // shuffle mask
1359             Value* vConstMask = C<char>({0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
1360                                          0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15});
1361             Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherInput, v32x8Ty), vConstMask), vGatherTy);
1362             // after pshufb: group components together in each 128bit lane
1363             // 256i - 0    1    2    3    4    5    6    7
1364             //        xxxx yyyy zzzz wwww xxxx yyyy zzzz wwww
1365
1366             Value* vi128XY = BITCAST(PERMD(vShufResult, C<int32_t>({0, 4, 0, 0, 1, 5, 0, 0})), v128Ty);
1367             // after PERMD: move and pack xy and zw components in low 64 bits of each 128bit lane
1368             // 256i - 0    1    2    3    4    5    6    7
1369             //        xxxx xxxx dcdc dcdc yyyy yyyy dcdc dcdc (dc - don't care)
1370
1371             // do the same for zw components
1372             Value* vi128ZW = nullptr;
1373             if(info.numComps > 2)
1374             {
1375                 vi128ZW = BITCAST(PERMD(vShufResult, C<int32_t>({2, 6, 0, 0, 3, 7, 0, 0})), v128Ty);
1376             }
1377
1378             // sign extend all enabled components. If we have a fill vVertexElements, output to current simdvertex
1379             for(uint32_t i = 0; i < 4; i++)
1380             {
1381                 uint32_t swizzleIndex = info.swizzle[i];
1382                 // todo: fix for packed
1383                 Value* vGatherMaskedVal = VIMMED1((int32_t)(info.defaults[i]));
1384                 if(i >= info.numComps)
1385                 {
1386                     // set the default component val
1387                     vGatherOutput[swizzleIndex] = vGatherMaskedVal;
1388                     continue;
1389                 }
1390
1391                 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1392                 uint32_t lane = ((i == 0) || (i == 2)) ? 0 : 1;
1393                 // if x or y, use vi128XY permute result, else use vi128ZW
1394                 Value* selectedPermute = (i < 2) ? vi128XY : vi128ZW;
1395
1396                 // sign extend
1397                 vGatherOutput[swizzleIndex] = VEXTRACT(selectedPermute, C(lane));
1398             }
1399         }
1400         // else zero extend
1401         else{
1402             // shuffle enabled components into lower byte of each 32bit lane, 0 extending to 32 bits
1403             // apply defaults
1404             for (uint32_t i = 0; i < 4; ++i)
1405             {
1406                 vGatherOutput[i] = VIMMED1((int32_t)info.defaults[i]);
1407             }
1408
1409             for(uint32_t i = 0; i < info.numComps; i++){
1410                 uint32_t swizzleIndex = info.swizzle[i];
1411
1412                 // pshufb masks for each component
1413                 Value* vConstMask;
1414                 switch(i)
1415                 {
1416                     case 0:
1417                         // x shuffle mask
1418                         vConstMask = C<char>({0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1,
1419                                               0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1});
1420                         break;
1421                     case 1:
1422                         // y shuffle mask
1423                         vConstMask = C<char>({1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1,
1424                                               1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1});
1425                         break;
1426                     case 2:
1427                         // z shuffle mask
1428                         vConstMask = C<char>({2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1,
1429                                               2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1});
1430                         break;
1431                     case 3:
1432                         // w shuffle mask
1433                         vConstMask = C<char>({3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1,
1434                                               3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1});
1435                         break;
1436                     default:
1437                         vConstMask = nullptr;
1438                         break;
1439                 }
1440
1441                     vGatherOutput[swizzleIndex] = BITCAST(PSHUFB(BITCAST(vGatherInput, v32x8Ty), vConstMask), vGatherTy);
1442                     // after pshufb for x channel
1443                     // 256i - 0    1    2    3    4    5    6    7
1444                     //        x000 x000 x000 x000 x000 x000 x000 x000
1445             }
1446         }
1447     }
1448
1449     // Helper function to create alloca in entry block of function
1450     Value* Builder::CreateEntryAlloca(Function* pFunc, Type* pType)
1451     {
1452         auto saveIP = IRB()->saveIP();
1453         IRB()->SetInsertPoint(&pFunc->getEntryBlock(),
1454                               pFunc->getEntryBlock().begin());
1455         Value* pAlloca = ALLOCA(pType);
1456         if (saveIP.isSet()) IRB()->restoreIP(saveIP);
1457         return pAlloca;
1458     }
1459
1460     Value* Builder::CreateEntryAlloca(Function* pFunc, Type* pType, Value* pArraySize)
1461     {
1462         auto saveIP = IRB()->saveIP();
1463         IRB()->SetInsertPoint(&pFunc->getEntryBlock(),
1464             pFunc->getEntryBlock().begin());
1465         Value* pAlloca = ALLOCA(pType, pArraySize);
1466         if (saveIP.isSet()) IRB()->restoreIP(saveIP);
1467         return pAlloca;
1468     }
1469
1470     //////////////////////////////////////////////////////////////////////////
1471     /// @brief emulates a scatter operation.
1472     /// @param pDst - pointer to destination
1473     /// @param vSrc - vector of src data to scatter
1474     /// @param vOffsets - vector of byte offsets from pDst
1475     /// @param vMask - mask of valid lanes
1476     void Builder::SCATTERPS(Value* pDst, Value* vSrc, Value* vOffsets, Value* vMask)
1477     {
1478         /* Scatter algorithm
1479
1480            while(Index = BitScanForward(mask))
1481                 srcElem = srcVector[Index]
1482                 offsetElem = offsetVector[Index]
1483                 *(pDst + offsetElem) = srcElem
1484                 Update mask (&= ~(1<<Index)
1485
1486         */
1487
1488         BasicBlock* pCurBB = IRB()->GetInsertBlock();
1489         Function* pFunc = pCurBB->getParent();
1490         Type* pSrcTy = vSrc->getType()->getVectorElementType();
1491
1492         // Store vectors on stack
1493         if (pScatterStackSrc == nullptr)
1494         {
1495             // Save off stack allocations and reuse per scatter. Significantly reduces stack
1496             // requirements for shaders with a lot of scatters.
1497             pScatterStackSrc = CreateEntryAlloca(pFunc, mSimdInt64Ty);
1498             pScatterStackOffsets = CreateEntryAlloca(pFunc, mSimdInt32Ty);
1499         }
1500
1501         Value* pSrcArrayPtr = BITCAST(pScatterStackSrc, PointerType::get(vSrc->getType(), 0));
1502         Value* pOffsetsArrayPtr = pScatterStackOffsets;
1503         STORE(vSrc, pSrcArrayPtr);
1504         STORE(vOffsets, pOffsetsArrayPtr);
1505
1506         // Cast to pointers for random access
1507         pSrcArrayPtr = POINTER_CAST(pSrcArrayPtr, PointerType::get(pSrcTy, 0));
1508         pOffsetsArrayPtr = POINTER_CAST(pOffsetsArrayPtr, PointerType::get(mInt32Ty, 0));
1509
1510         Value* pMask = VMOVMSKPS(BITCAST(vMask, mSimdFP32Ty));
1511
1512         // Get cttz function
1513         Function* pfnCttz = Intrinsic::getDeclaration(mpJitMgr->mpCurrentModule, Intrinsic::cttz, { mInt32Ty });
1514
1515         // Setup loop basic block
1516         BasicBlock* pLoop = BasicBlock::Create(mpJitMgr->mContext, "Scatter Loop", pFunc);
1517
1518         // compute first set bit
1519         Value* pIndex = CALL(pfnCttz, { pMask, C(false) });
1520
1521         Value* pIsUndef = ICMP_EQ(pIndex, C(32));
1522
1523         // Split current block
1524         BasicBlock* pPostLoop = pCurBB->splitBasicBlock(cast<Instruction>(pIsUndef)->getNextNode());
1525
1526         // Remove unconditional jump created by splitBasicBlock
1527         pCurBB->getTerminator()->eraseFromParent();
1528
1529         // Add terminator to end of original block
1530         IRB()->SetInsertPoint(pCurBB);
1531
1532         // Add conditional branch
1533         COND_BR(pIsUndef, pPostLoop, pLoop);
1534
1535         // Add loop basic block contents
1536         IRB()->SetInsertPoint(pLoop);
1537         PHINode* pIndexPhi = PHI(mInt32Ty, 2);
1538         PHINode* pMaskPhi = PHI(mInt32Ty, 2);
1539
1540         pIndexPhi->addIncoming(pIndex, pCurBB);
1541         pMaskPhi->addIncoming(pMask, pCurBB);
1542
1543         // Extract elements for this index
1544         Value* pSrcElem = LOADV(pSrcArrayPtr, { pIndexPhi });
1545         Value* pOffsetElem = LOADV(pOffsetsArrayPtr, { pIndexPhi });
1546
1547         // GEP to this offset in dst
1548         Value* pCurDst = GEP(pDst, pOffsetElem);
1549         pCurDst = POINTER_CAST(pCurDst, PointerType::get(pSrcTy, 0));
1550         STORE(pSrcElem, pCurDst);
1551
1552         // Update the mask
1553         Value* pNewMask = AND(pMaskPhi, NOT(SHL(C(1), pIndexPhi)));
1554
1555         // Terminator
1556         Value* pNewIndex = CALL(pfnCttz, { pNewMask, C(false) });
1557
1558         pIsUndef = ICMP_EQ(pNewIndex, C(32));
1559         COND_BR(pIsUndef, pPostLoop, pLoop);
1560
1561         // Update phi edges
1562         pIndexPhi->addIncoming(pNewIndex, pLoop);
1563         pMaskPhi->addIncoming(pNewMask, pLoop);
1564
1565         // Move builder to beginning of post loop
1566         IRB()->SetInsertPoint(pPostLoop, pPostLoop->begin());
1567     }
1568
1569     Value* Builder::VABSPS(Value* a)
1570     {
1571         Value* asInt = BITCAST(a, mSimdInt32Ty);
1572         Value* result = BITCAST(AND(asInt, VIMMED1(0x7fffffff)), mSimdFP32Ty);
1573         return result;
1574     }
1575
1576     Value *Builder::ICLAMP(Value* src, Value* low, Value* high)
1577     {
1578         Value *lowCmp = ICMP_SLT(src, low);
1579         Value *ret = SELECT(lowCmp, low, src);
1580
1581         Value *highCmp = ICMP_SGT(ret, high);
1582         ret = SELECT(highCmp, high, ret);
1583
1584         return ret;
1585     }
1586
1587     Value *Builder::FCLAMP(Value* src, Value* low, Value* high)
1588     {
1589         Value *lowCmp = FCMP_OLT(src, low);
1590         Value *ret = SELECT(lowCmp, low, src);
1591
1592         Value *highCmp = FCMP_OGT(ret, high);
1593         ret = SELECT(highCmp, high, ret);
1594
1595         return ret;
1596     }
1597
1598     Value *Builder::FCLAMP(Value* src, float low, float high)
1599     {
1600         Value* result = VMAXPS(src, VIMMED1(low));
1601         result = VMINPS(result, VIMMED1(high));
1602
1603         return result;
1604     }
1605
1606     //////////////////////////////////////////////////////////////////////////
1607     /// @brief save/restore stack, providing ability to push/pop the stack and
1608     ///        reduce overall stack requirements for temporary stack use
1609     Value* Builder::STACKSAVE()
1610     {
1611         Function* pfnStackSave = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::stacksave);
1612         return CALLA(pfnStackSave);
1613     }
1614
1615     void Builder::STACKRESTORE(Value* pSaved)
1616     {
1617         Function* pfnStackRestore = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::stackrestore);
1618         CALL(pfnStackRestore, std::initializer_list<Value*>{pSaved});
1619     }
1620
1621     Value *Builder::FMADDPS(Value* a, Value* b, Value* c)
1622     {
1623         Value* vOut;
1624         // use FMADs if available
1625         if(JM()->mArch.AVX2())
1626         {
1627             vOut = VFMADDPS(a, b, c);
1628         }
1629         else
1630         {
1631             vOut = FADD(FMUL(a, b), c);
1632         }
1633         return vOut;
1634     }
1635
1636     Value* Builder::POPCNT(Value* a)
1637     {
1638         Function* pCtPop = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::ctpop, { a->getType() });
1639         return CALL(pCtPop, std::initializer_list<Value*>{a});
1640     }
1641
1642     //////////////////////////////////////////////////////////////////////////
1643     /// @brief C functions called by LLVM IR
1644     //////////////////////////////////////////////////////////////////////////
1645
1646     //////////////////////////////////////////////////////////////////////////
1647     /// @brief called in JIT code, inserted by PRINT
1648     /// output to both stdout and visual studio debug console
1649     void __cdecl CallPrint(const char* fmt, ...)
1650     {
1651         va_list args;
1652         va_start(args, fmt);
1653         vprintf(fmt, args);
1654
1655     #if defined( _WIN32 )
1656         char strBuf[1024];
1657         vsnprintf_s(strBuf, _TRUNCATE, fmt, args);
1658         OutputDebugStringA(strBuf);
1659     #endif
1660
1661         va_end(args);
1662     }
1663
1664     Value *Builder::VEXTRACTI128(Value* a, Constant* imm8)
1665     {
1666         bool flag = !imm8->isZeroValue();
1667         SmallVector<Constant*,8> idx;
1668         for (unsigned i = 0; i < mVWidth / 2; i++) {
1669             idx.push_back(C(flag ? i + mVWidth / 2 : i));
1670         }
1671         return VSHUFFLE(a, VUNDEF_I(), ConstantVector::get(idx));
1672     }
1673
1674     Value *Builder::VINSERTI128(Value* a, Value* b, Constant* imm8)
1675     {
1676         bool flag = !imm8->isZeroValue();
1677         SmallVector<Constant*,8> idx;
1678         for (unsigned i = 0; i < mVWidth; i++) {
1679             idx.push_back(C(i));
1680         }
1681         Value *inter = VSHUFFLE(b, VUNDEF_I(), ConstantVector::get(idx));
1682
1683         SmallVector<Constant*,8> idx2;
1684         for (unsigned i = 0; i < mVWidth / 2; i++) {
1685             idx2.push_back(C(flag ? i : i + mVWidth));
1686         }
1687         for (unsigned i = mVWidth / 2; i < mVWidth; i++) {
1688             idx2.push_back(C(flag ? i + mVWidth / 2 : i));
1689         }
1690         return VSHUFFLE(a, inter, ConstantVector::get(idx2));
1691     }
1692
1693     // rdtsc buckets macros
1694     void Builder::RDTSC_START(Value* pBucketMgr, Value* pId)
1695     {
1696         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1697         // buckets framework when single threaded
1698         if (KNOB_SINGLE_THREADED)
1699         {
1700             std::vector<Type*> args{
1701                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
1702                 mInt32Ty                        // id
1703             };
1704
1705             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
1706             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StartBucket", pFuncTy));
1707             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StartBucket") == nullptr)
1708             {
1709                 sys::DynamicLibrary::AddSymbol("BucketManager_StartBucket", (void*)&BucketManager_StartBucket);
1710             }
1711
1712             CALL(pFunc, { pBucketMgr, pId });
1713         }
1714     }
1715
1716     void Builder::RDTSC_STOP(Value* pBucketMgr, Value* pId)
1717     {
1718         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1719         // buckets framework when single threaded
1720         if (KNOB_SINGLE_THREADED)
1721         {
1722             std::vector<Type*> args{
1723                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
1724                 mInt32Ty                        // id
1725             };
1726
1727             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
1728             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StopBucket", pFuncTy));
1729             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StopBucket") == nullptr)
1730             {
1731                 sys::DynamicLibrary::AddSymbol("BucketManager_StopBucket", (void*)&BucketManager_StopBucket);
1732             }
1733
1734             CALL(pFunc, { pBucketMgr, pId });
1735         }
1736     }
1737
1738
1739     uint32_t Builder::GetTypeSize(Type* pType)
1740     {
1741         if (pType->isStructTy())
1742         {
1743             uint32_t numElems = pType->getStructNumElements();
1744             Type* pElemTy = pType->getStructElementType(0);
1745             return numElems * GetTypeSize(pElemTy);
1746         }
1747
1748         if (pType->isArrayTy())
1749         {
1750             uint32_t numElems = pType->getArrayNumElements();
1751             Type* pElemTy = pType->getArrayElementType();
1752             return numElems * GetTypeSize(pElemTy);
1753         }
1754
1755         if (pType->isIntegerTy())
1756         {
1757             uint32_t bitSize = pType->getIntegerBitWidth();
1758             return bitSize / 8;
1759         }
1760
1761         if (pType->isFloatTy())
1762         {
1763             return 4;
1764         }
1765
1766         if (pType->isHalfTy())
1767         {
1768             return 2;
1769         }
1770
1771         if (pType->isDoubleTy())
1772         {
1773             return 8;
1774         }
1775
1776         SWR_ASSERT(false, "Unimplemented type.");
1777         return 0;
1778     }
1779 }