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 "jit_pch.hpp"
  31 #include "builder.h"
  32 #include "common/rdtsc_buckets.h"
  33
  34 #include <cstdarg>
  35
  36 extern "C" void CallPrint(const char* fmt, ...);
  37
  38 namespace SwrJit
  39 {
  40     //////////////////////////////////////////////////////////////////////////
  41     /// @brief Convert an IEEE 754 32-bit single precision float to an
  42     ///        16 bit float with 5 exponent bits and a variable
  43     ///        number of mantissa bits.
  44     /// @param val - 32-bit float
  45     /// @todo Maybe move this outside of this file into a header?
  46     static uint16_t ConvertFloat32ToFloat16(float val)
  47     {
  48         uint32_t sign, exp, mant;
  49         uint32_t roundBits;
  50
  51         // Extract the sign, exponent, and mantissa
  52         uint32_t uf = *(uint32_t*)&val;
  53         sign = (uf & 0x80000000) >> 31;
  54         exp = (uf & 0x7F800000) >> 23;
  55         mant = uf & 0x007FFFFF;
  56
  57         // Check for out of range
  58         if (std::isnan(val))
  59         {
  60             exp = 0x1F;
  61             mant = 0x200;
  62             sign = 1;                     // set the sign bit for NANs
  63         }
  64         else if (std::isinf(val))
  65         {
  66             exp = 0x1f;
  67             mant = 0x0;
  68         }
  69         else if (exp > (0x70 + 0x1E)) // Too big to represent -> max representable value
  70         {
  71             exp = 0x1E;
  72             mant = 0x3FF;
  73         }
  74         else if ((exp <= 0x70) && (exp >= 0x66)) // It's a denorm
  75         {
  76             mant |= 0x00800000;
  77             for (; exp <= 0x70; mant >>= 1, exp++)
  78                 ;
  79             exp = 0;
  80             mant = mant >> 13;
  81         }
  82         else if (exp < 0x66) // Too small to represent -> Zero
  83         {
  84             exp = 0;
  85             mant = 0;
  86         }
  87         else
  88         {
  89             // Saves bits that will be shifted off for rounding
  90             roundBits = mant & 0x1FFFu;
  91             // convert exponent and mantissa to 16 bit format
  92             exp = exp - 0x70;
  93             mant = mant >> 13;
  94
  95             // Essentially RTZ, but round up if off by only 1 lsb
  96             if (roundBits == 0x1FFFu)
  97             {
  98                 mant++;
  99                 // check for overflow
 100                 if ((mant & 0xC00u) != 0)
 101                     exp++;
 102                 // make sure only the needed bits are used
 103                 mant &= 0x3FF;
 104             }
 105         }
 106
 107         uint32_t tmpVal = (sign << 15) | (exp << 10) | mant;
 108         return (uint16_t)tmpVal;
 109     }
 110
 111     //////////////////////////////////////////////////////////////////////////
 112     /// @brief Convert an IEEE 754 16-bit float to an 32-bit single precision
 113     ///        float
 114     /// @param val - 16-bit float
 115     /// @todo Maybe move this outside of this file into a header?
 116     static float ConvertFloat16ToFloat32(uint32_t val)
 117     {
 118         uint32_t result;
 119         if ((val & 0x7fff) == 0)
 120         {
 121             result = ((uint32_t)(val & 0x8000)) << 16;
 122         }
 123         else if ((val & 0x7c00) == 0x7c00)
 124         {
 125             result = ((val & 0x3ff) == 0) ? 0x7f800000 : 0x7fc00000;
 126             result |= ((uint32_t)val & 0x8000) << 16;
 127         }
 128         else
 129         {
 130             uint32_t sign = (val & 0x8000) << 16;
 131             uint32_t mant = (val & 0x3ff) << 13;
 132             uint32_t exp = (val >> 10) & 0x1f;
 133             if ((exp == 0) && (mant != 0)) // Adjust exponent and mantissa for denormals
 134             {
 135                 mant <<= 1;
 136                 while (mant < (0x400 << 13))
 137                 {
 138                     exp--;
 139                     mant <<= 1;
 140                 }
 141                 mant &= (0x3ff << 13);
 142             }
 143             exp = ((exp - 15 + 127) & 0xff) << 23;
 144             result = sign | exp | mant;
 145         }
 146
 147         return *(float*)&result;
 148     }
 149
 150     Constant *Builder::C(bool i)
 151     {
 152         return ConstantInt::get(IRB()->getInt1Ty(), (i ? 1 : 0));
 153     }
 154
 155     Constant *Builder::C(char i)
 156     {
 157         return ConstantInt::get(IRB()->getInt8Ty(), i);
 158     }
 159
 160     Constant *Builder::C(uint8_t i)
 161     {
 162         return ConstantInt::get(IRB()->getInt8Ty(), i);
 163     }
 164
 165     Constant *Builder::C(int i)
 166     {
 167         return ConstantInt::get(IRB()->getInt32Ty(), i);
 168     }
 169
 170     Constant *Builder::C(int64_t i)
 171     {
 172         return ConstantInt::get(IRB()->getInt64Ty(), i);
 173     }
 174
 175     Constant *Builder::C(uint16_t i)
 176     {
 177         return ConstantInt::get(mInt16Ty,i);
 178     }
 179
 180     Constant *Builder::C(uint32_t i)
 181     {
 182         return ConstantInt::get(IRB()->getInt32Ty(), i);
 183     }
 184
 185     Constant *Builder::C(float i)
 186     {
 187         return ConstantFP::get(IRB()->getFloatTy(), i);
 188     }
 189
 190     Constant *Builder::PRED(bool pred)
 191     {
 192         return ConstantInt::get(IRB()->getInt1Ty(), (pred ? 1 : 0));
 193     }
 194
 195     Value *Builder::VIMMED1(int i)
 196     {
 197         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 198     }
 199
 200     Value *Builder::VIMMED1_16(int i)
 201     {
 202         return ConstantVector::getSplat(mVWidth16, cast<ConstantInt>(C(i)));
 203     }
 204
 205     Value *Builder::VIMMED1(uint32_t i)
 206     {
 207         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 208     }
 209
 210     Value *Builder::VIMMED1_16(uint32_t i)
 211     {
 212         return ConstantVector::getSplat(mVWidth16, cast<ConstantInt>(C(i)));
 213     }
 214
 215     Value *Builder::VIMMED1(float i)
 216     {
 217         return ConstantVector::getSplat(mVWidth, cast<ConstantFP>(C(i)));
 218     }
 219
 220     Value *Builder::VIMMED1_16(float i)
 221     {
 222         return ConstantVector::getSplat(mVWidth16, cast<ConstantFP>(C(i)));
 223     }
 224
 225     Value *Builder::VIMMED1(bool i)
 226     {
 227         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 228     }
 229
 230     Value *Builder::VIMMED1_16(bool i)
 231     {
 232         return ConstantVector::getSplat(mVWidth16, cast<ConstantInt>(C(i)));
 233     }
 234
 235     Value *Builder::VUNDEF_IPTR()
 236     {
 237         return UndefValue::get(VectorType::get(mInt32PtrTy,mVWidth));
 238     }
 239
 240     Value *Builder::VUNDEF(Type* t)
 241     {
 242         return UndefValue::get(VectorType::get(t, mVWidth));
 243     }
 244
 245     Value *Builder::VUNDEF_I()
 246     {
 247         return UndefValue::get(VectorType::get(mInt32Ty, mVWidth));
 248     }
 249
 250     Value *Builder::VUNDEF_I_16()
 251     {
 252         return UndefValue::get(VectorType::get(mInt32Ty, mVWidth16));
 253     }
 254
 255     Value *Builder::VUNDEF_F()
 256     {
 257         return UndefValue::get(VectorType::get(mFP32Ty, mVWidth));
 258     }
 259
 260     Value *Builder::VUNDEF_F_16()
 261     {
 262         return UndefValue::get(VectorType::get(mFP32Ty, mVWidth16));
 263     }
 264
 265     Value *Builder::VUNDEF(Type *ty, uint32_t size)
 266     {
 267         return UndefValue::get(VectorType::get(ty, size));
 268     }
 269
 270     Value *Builder::VBROADCAST(Value *src, const llvm::Twine& name)
 271     {
 272         // check if src is already a vector
 273         if (src->getType()->isVectorTy())
 274         {
 275             return src;
 276         }
 277
 278         return VECTOR_SPLAT(mVWidth, src, name);
 279     }
 280
 281     Value *Builder::VBROADCAST_16(Value *src)
 282     {
 283         // check if src is already a vector
 284         if (src->getType()->isVectorTy())
 285         {
 286             return src;
 287         }
 288
 289         return VECTOR_SPLAT(mVWidth16, src);
 290     }
 291
 292     uint32_t Builder::IMMED(Value* v)
 293     {
 294         SWR_ASSERT(isa<ConstantInt>(v));
 295         ConstantInt *pValConst = cast<ConstantInt>(v);
 296         return pValConst->getZExtValue();
 297     }
 298
 299     int32_t Builder::S_IMMED(Value* v)
 300     {
 301         SWR_ASSERT(isa<ConstantInt>(v));
 302         ConstantInt *pValConst = cast<ConstantInt>(v);
 303         return pValConst->getSExtValue();
 304     }
 305
 306     CallInst *Builder::CALL(Value *Callee, const std::initializer_list<Value*> &argsList, const llvm::Twine& name)
 307     {
 308         std::vector<Value*> args;
 309         for (auto arg : argsList)
 310             args.push_back(arg);
 311         return CALLA(Callee, args, name);
 312     }
 313
 314     CallInst *Builder::CALL(Value *Callee, Value* arg)
 315     {
 316         std::vector<Value*> args;
 317         args.push_back(arg);
 318         return CALLA(Callee, args);
 319     }
 320
 321     CallInst *Builder::CALL2(Value *Callee, Value* arg1, Value* arg2)
 322     {
 323         std::vector<Value*> args;
 324         args.push_back(arg1);
 325         args.push_back(arg2);
 326         return CALLA(Callee, args);
 327     }
 328
 329     CallInst *Builder::CALL3(Value *Callee, Value* arg1, Value* arg2, Value* arg3)
 330     {
 331         std::vector<Value*> args;
 332         args.push_back(arg1);
 333         args.push_back(arg2);
 334         args.push_back(arg3);
 335         return CALLA(Callee, args);
 336     }
 337
 338     Value *Builder::VRCP(Value *va, const llvm::Twine& name)
 339     {
 340         return FDIV(VIMMED1(1.0f), va, name);  // 1 / a
 341     }
 342
 343     Value *Builder::VPLANEPS(Value* vA, Value* vB, Value* vC, Value* &vX, Value* &vY)
 344     {
 345         Value* vOut = FMADDPS(vA, vX, vC);
 346         vOut = FMADDPS(vB, vY, vOut);
 347         return vOut;
 348     }
 349
 350     //////////////////////////////////////////////////////////////////////////
 351     /// @brief insert a JIT call to CallPrint
 352     /// - outputs formatted string to both stdout and VS output window
 353     /// - DEBUG builds only
 354     /// Usage example:
 355     ///   PRINT("index %d = 0x%p\n",{C(lane), pIndex});
 356     ///   where C(lane) creates a constant value to print, and pIndex is the Value*
 357     ///   result from a GEP, printing out the pointer to memory
 358     /// @param printStr - constant string to print, which includes format specifiers
 359     /// @param printArgs - initializer list of Value*'s to print to std out
 360     CallInst *Builder::PRINT(const std::string &printStr,const std::initializer_list<Value*> &printArgs)
 361     {
 362         // push the arguments to CallPrint into a vector
 363         std::vector<Value*> printCallArgs;
 364         // save room for the format string.  we still need to modify it for vectors
 365         printCallArgs.resize(1);
 366
 367         // search through the format string for special processing
 368         size_t pos = 0;
 369         std::string tempStr(printStr);
 370         pos = tempStr.find('%', pos);
 371         auto v = printArgs.begin();
 372
 373         while ((pos != std::string::npos) && (v != printArgs.end()))
 374         {
 375             Value* pArg = *v;
 376             Type* pType = pArg->getType();
 377
 378             if (pType->isVectorTy())
 379             {
 380                 Type* pContainedType = pType->getContainedType(0);
 381
 382                 if (toupper(tempStr[pos + 1]) == 'X')
 383                 {
 384                     tempStr[pos] = '0';
 385                     tempStr[pos + 1] = 'x';
 386                     tempStr.insert(pos + 2, "%08X ");
 387                     pos += 7;
 388
 389                     printCallArgs.push_back(VEXTRACT(pArg, C(0)));
 390
 391                     std::string vectorFormatStr;
 392                     for (uint32_t i = 1; i < pType->getVectorNumElements(); ++i)
 393                     {
 394                         vectorFormatStr += "0x%08X ";
 395                         printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 396                     }
 397
 398                     tempStr.insert(pos, vectorFormatStr);
 399                     pos += vectorFormatStr.size();
 400                 }
 401                 else if ((tempStr[pos + 1] == 'f') && (pContainedType->isFloatTy()))
 402                 {
 403                     uint32_t i = 0;
 404                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 405                     {
 406                         tempStr.insert(pos, std::string("%f "));
 407                         pos += 3;
 408                         printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 409                     }
 410                     printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 411                 }
 412                 else if ((tempStr[pos + 1] == 'd') && (pContainedType->isIntegerTy()))
 413                 {
 414                     uint32_t i = 0;
 415                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 416                     {
 417                         tempStr.insert(pos, std::string("%d "));
 418                         pos += 3;
 419                         printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 420                     }
 421                     printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 422                 }
 423             }
 424             else
 425             {
 426                 if (toupper(tempStr[pos + 1]) == 'X')
 427                 {
 428                     tempStr[pos] = '0';
 429                     tempStr.insert(pos + 1, "x%08");
 430                     printCallArgs.push_back(pArg);
 431                     pos += 3;
 432                 }
 433                 // for %f we need to cast float Values to doubles so that they print out correctly
 434                 else if ((tempStr[pos + 1] == 'f') && (pType->isFloatTy()))
 435                 {
 436                     printCallArgs.push_back(FP_EXT(pArg, Type::getDoubleTy(JM()->mContext)));
 437                     pos++;
 438                 }
 439                 else
 440                 {
 441                     printCallArgs.push_back(pArg);
 442                 }
 443             }
 444
 445             // advance to the next arguement
 446             v++;
 447             pos = tempStr.find('%', ++pos);
 448         }
 449
 450         // create global variable constant string
 451         Constant *constString = ConstantDataArray::getString(JM()->mContext,tempStr,true);
 452         GlobalVariable *gvPtr = new GlobalVariable(constString->getType(),true,GlobalValue::InternalLinkage,constString,"printStr");
 453         JM()->mpCurrentModule->getGlobalList().push_back(gvPtr);
 454
 455         // get a pointer to the first character in the constant string array
 456         std::vector<Constant*> geplist{C(0),C(0)};
 457         Constant *strGEP = ConstantExpr::getGetElementPtr(nullptr, gvPtr,geplist,false);
 458
 459         // insert the pointer to the format string in the argument vector
 460         printCallArgs[0] = strGEP;
 461
 462         // get pointer to CallPrint function and insert decl into the module if needed
 463         std::vector<Type*> args;
 464         args.push_back(PointerType::get(mInt8Ty,0));
 465         FunctionType* callPrintTy = FunctionType::get(Type::getVoidTy(JM()->mContext),args,true);
 466         Function *callPrintFn = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("CallPrint", callPrintTy));
 467
 468         // if we haven't yet added the symbol to the symbol table
 469         if((sys::DynamicLibrary::SearchForAddressOfSymbol("CallPrint")) == nullptr)
 470         {
 471             sys::DynamicLibrary::AddSymbol("CallPrint", (void *)&CallPrint);
 472         }
 473
 474         // insert a call to CallPrint
 475         return CALLA(callPrintFn,printCallArgs);
 476     }
 477
 478     //////////////////////////////////////////////////////////////////////////
 479     /// @brief Wrapper around PRINT with initializer list.
 480     CallInst* Builder::PRINT(const std::string &printStr)
 481     {
 482         return PRINT(printStr, {});
 483     }
 484
 485     Value *Builder::EXTRACT_16(Value *x, uint32_t imm)
 486     {
 487         if (imm == 0)
 488         {
 489             return VSHUFFLE(x, UndefValue::get(x->getType()), { 0, 1, 2, 3, 4, 5, 6, 7 });
 490         }
 491         else
 492         {
 493             return VSHUFFLE(x, UndefValue::get(x->getType()), { 8, 9, 10, 11, 12, 13, 14, 15 });
 494         }
 495     }
 496
 497     Value *Builder::JOIN_16(Value *a, Value *b)
 498     {
 499         return VSHUFFLE(a, b, { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 });
 500     }
 501
 502     //////////////////////////////////////////////////////////////////////////
 503     /// @brief convert x86 <N x float> mask to llvm <N x i1> mask
 504     Value *Builder::MASK(Value *vmask)
 505     {
 506         Value *src = BITCAST(vmask, mSimdInt32Ty);
 507         return ICMP_SLT(src, VIMMED1(0));
 508     }
 509
 510     Value *Builder::MASK_16(Value *vmask)
 511     {
 512         Value *src = BITCAST(vmask, mSimd16Int32Ty);
 513         return ICMP_SLT(src, VIMMED1_16(0));
 514     }
 515
 516     //////////////////////////////////////////////////////////////////////////
 517     /// @brief convert llvm <N x i1> mask to x86 <N x i32> mask
 518     Value *Builder::VMASK(Value *mask)
 519     {
 520         return S_EXT(mask, mSimdInt32Ty);
 521     }
 522
 523     Value *Builder::VMASK_16(Value *mask)
 524     {
 525         return S_EXT(mask, mSimd16Int32Ty);
 526     }
 527
 528     //////////////////////////////////////////////////////////////////////////
 529     /// @brief Generate a VPSHUFB operation in LLVM IR.  If not
 530     /// supported on the underlying platform, emulate it
 531     /// @param a - 256bit SIMD(32x8bit) of 8bit integer values
 532     /// @param b - 256bit SIMD(32x8bit) of 8bit integer mask values
 533     /// Byte masks in lower 128 lane of b selects 8 bit values from lower
 534     /// 128bits of a, and vice versa for the upper lanes.  If the mask
 535     /// value is negative, '0' is inserted.
 536     Value *Builder::PSHUFB(Value* a, Value* b)
 537     {
 538         Value* res;
 539         // use avx2 pshufb instruction if available
 540         if(JM()->mArch.AVX2())
 541         {
 542             res = VPSHUFB(a, b);
 543         }
 544         else
 545         {
 546             Constant* cB = dyn_cast<Constant>(b);
 547             // number of 8 bit elements in b
 548             uint32_t numElms = cast<VectorType>(cB->getType())->getNumElements();
 549             // output vector
 550             Value* vShuf = UndefValue::get(VectorType::get(mInt8Ty, numElms));
 551
 552             // insert an 8 bit value from the high and low lanes of a per loop iteration
 553             numElms /= 2;
 554             for(uint32_t i = 0; i < numElms; i++)
 555             {
 556                 ConstantInt* cLow128b = cast<ConstantInt>(cB->getAggregateElement(i));
 557                 ConstantInt* cHigh128b = cast<ConstantInt>(cB->getAggregateElement(i + numElms));
 558
 559                 // extract values from constant mask
 560                 char valLow128bLane =  (char)(cLow128b->getSExtValue());
 561                 char valHigh128bLane = (char)(cHigh128b->getSExtValue());
 562
 563                 Value* insertValLow128b;
 564                 Value* insertValHigh128b;
 565
 566                 // if the mask value is negative, insert a '0' in the respective output position
 567                 // otherwise, lookup the value at mask position (bits 3..0 of the respective mask byte) in a and insert in output vector
 568                 insertValLow128b = (valLow128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valLow128bLane & 0xF)));
 569                 insertValHigh128b = (valHigh128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valHigh128bLane & 0xF) + numElms));
 570
 571                 vShuf = VINSERT(vShuf, insertValLow128b, i);
 572                 vShuf = VINSERT(vShuf, insertValHigh128b, (i + numElms));
 573             }
 574             res = vShuf;
 575         }
 576         return res;
 577     }
 578
 579     //////////////////////////////////////////////////////////////////////////
 580     /// @brief Generate a VPSHUFB operation (sign extend 8 8bit values to 32
 581     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 582     /// @param a - 128bit SIMD lane(16x8bit) of 8bit integer values.  Only
 583     /// lower 8 values are used.
 584     Value *Builder::PMOVSXBD(Value* a)
 585     {
 586         // VPMOVSXBD output type
 587         Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 588         // Extract 8 values from 128bit lane and sign extend
 589         return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 590     }
 591
 592     //////////////////////////////////////////////////////////////////////////
 593     /// @brief Generate a VPSHUFB operation (sign extend 8 16bit values to 32
 594     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 595     /// @param a - 128bit SIMD lane(8x16bit) of 16bit integer values.
 596     Value *Builder::PMOVSXWD(Value* a)
 597     {
 598         // VPMOVSXWD output type
 599         Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 600         // Extract 8 values from 128bit lane and sign extend
 601         return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 602     }
 603
 604     //////////////////////////////////////////////////////////////////////////
 605     /// @brief Generate a VCVTPH2PS operation (float16->float32 conversion)
 606     /// in LLVM IR.  If not supported on the underlying platform, emulate it
 607     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
 608     Value *Builder::CVTPH2PS(Value* a, const llvm::Twine& name)
 609     {
 610         if (JM()->mArch.F16C())
 611         {
 612             return VCVTPH2PS(a, name);
 613         }
 614         else
 615         {
 616             FunctionType* pFuncTy = FunctionType::get(mFP32Ty, mInt16Ty);
 617             Function* pCvtPh2Ps = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("ConvertFloat16ToFloat32", pFuncTy));
 618
 619             if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat16ToFloat32") == nullptr)
 620             {
 621                 sys::DynamicLibrary::AddSymbol("ConvertFloat16ToFloat32", (void *)&ConvertFloat16ToFloat32);
 622             }
 623
 624             Value* pResult = UndefValue::get(mSimdFP32Ty);
 625             for (uint32_t i = 0; i < mVWidth; ++i)
 626             {
 627                 Value* pSrc = VEXTRACT(a, C(i));
 628                 Value* pConv = CALL(pCvtPh2Ps, std::initializer_list<Value*>{pSrc});
 629                 pResult = VINSERT(pResult, pConv, C(i));
 630             }
 631
 632             pResult->setName(name);
 633             return pResult;
 634         }
 635     }
 636
 637     //////////////////////////////////////////////////////////////////////////
 638     /// @brief Generate a VCVTPS2PH operation (float32->float16 conversion)
 639     /// in LLVM IR.  If not supported on the underlying platform, emulate it
 640     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
 641     Value *Builder::CVTPS2PH(Value* a, Value* rounding)
 642     {
 643         if (JM()->mArch.F16C())
 644         {
 645             return VCVTPS2PH(a, rounding);
 646         }
 647         else
 648         {
 649             // call scalar C function for now
 650             FunctionType* pFuncTy = FunctionType::get(mInt16Ty, mFP32Ty);
 651             Function* pCvtPs2Ph = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("ConvertFloat32ToFloat16", pFuncTy));
 652
 653             if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat32ToFloat16") == nullptr)
 654             {
 655                 sys::DynamicLibrary::AddSymbol("ConvertFloat32ToFloat16", (void *)&ConvertFloat32ToFloat16);
 656             }
 657
 658             Value* pResult = UndefValue::get(mSimdInt16Ty);
 659             for (uint32_t i = 0; i < mVWidth; ++i)
 660             {
 661                 Value* pSrc = VEXTRACT(a, C(i));
 662                 Value* pConv = CALL(pCvtPs2Ph, std::initializer_list<Value*>{pSrc});
 663                 pResult = VINSERT(pResult, pConv, C(i));
 664             }
 665
 666             return pResult;
 667         }
 668     }
 669
 670     Value *Builder::PMAXSD(Value* a, Value* b)
 671     {
 672         Value* cmp = ICMP_SGT(a, b);
 673         return SELECT(cmp, a, b);
 674     }
 675
 676     Value *Builder::PMINSD(Value* a, Value* b)
 677     {
 678         Value* cmp = ICMP_SLT(a, b);
 679         return SELECT(cmp, a, b);
 680     }
 681
 682     Value *Builder::PMAXUD(Value* a, Value* b)
 683     {
 684         Value* cmp = ICMP_UGT(a, b);
 685         return SELECT(cmp, a, b);
 686     }
 687
 688     Value *Builder::PMINUD(Value* a, Value* b)
 689     {
 690         Value* cmp = ICMP_ULT(a, b);
 691         return SELECT(cmp, a, b);
 692     }
 693
 694     // Helper function to create alloca in entry block of function
 695     Value* Builder::CreateEntryAlloca(Function* pFunc, Type* pType)
 696     {
 697         auto saveIP = IRB()->saveIP();
 698         IRB()->SetInsertPoint(&pFunc->getEntryBlock(),
 699                               pFunc->getEntryBlock().begin());
 700         Value* pAlloca = ALLOCA(pType);
 701         if (saveIP.isSet()) IRB()->restoreIP(saveIP);
 702         return pAlloca;
 703     }
 704
 705     Value* Builder::CreateEntryAlloca(Function* pFunc, Type* pType, Value* pArraySize)
 706     {
 707         auto saveIP = IRB()->saveIP();
 708         IRB()->SetInsertPoint(&pFunc->getEntryBlock(),
 709             pFunc->getEntryBlock().begin());
 710         Value* pAlloca = ALLOCA(pType, pArraySize);
 711         if (saveIP.isSet()) IRB()->restoreIP(saveIP);
 712         return pAlloca;
 713     }
 714
 715     Value* Builder::VABSPS(Value* a)
 716     {
 717         Value* asInt = BITCAST(a, mSimdInt32Ty);
 718         Value* result = BITCAST(AND(asInt, VIMMED1(0x7fffffff)), mSimdFP32Ty);
 719         return result;
 720     }
 721
 722     Value *Builder::ICLAMP(Value* src, Value* low, Value* high, const llvm::Twine& name)
 723     {
 724         Value *lowCmp = ICMP_SLT(src, low);
 725         Value *ret = SELECT(lowCmp, low, src);
 726
 727         Value *highCmp = ICMP_SGT(ret, high);
 728         ret = SELECT(highCmp, high, ret, name);
 729
 730         return ret;
 731     }
 732
 733     Value *Builder::FCLAMP(Value* src, Value* low, Value* high)
 734     {
 735         Value *lowCmp = FCMP_OLT(src, low);
 736         Value *ret = SELECT(lowCmp, low, src);
 737
 738         Value *highCmp = FCMP_OGT(ret, high);
 739         ret = SELECT(highCmp, high, ret);
 740
 741         return ret;
 742     }
 743
 744     Value *Builder::FCLAMP(Value* src, float low, float high)
 745     {
 746         Value* result = VMAXPS(src, VIMMED1(low));
 747         result = VMINPS(result, VIMMED1(high));
 748
 749         return result;
 750     }
 751
 752     Value *Builder::FMADDPS(Value* a, Value* b, Value* c)
 753     {
 754         Value* vOut;
 755         // use FMADs if available
 756         if(JM()->mArch.AVX2())
 757         {
 758             vOut = VFMADDPS(a, b, c);
 759         }
 760         else
 761         {
 762             vOut = FADD(FMUL(a, b), c);
 763         }
 764         return vOut;
 765     }
 766
 767     //////////////////////////////////////////////////////////////////////////
 768     /// @brief pop count on vector mask (e.g. <8 x i1>)
 769     Value* Builder::VPOPCNT(Value* a)
 770     {
 771         Value* b = BITCAST(VMASK(a), mSimdFP32Ty);
 772         return POPCNT(VMOVMSKPS(b));
 773     }
 774
 775     //////////////////////////////////////////////////////////////////////////
 776     /// @brief C functions called by LLVM IR
 777     //////////////////////////////////////////////////////////////////////////
 778
 779     Value *Builder::VEXTRACTI128(Value* a, Constant* imm8)
 780     {
 781         bool flag = !imm8->isZeroValue();
 782         SmallVector<Constant*,8> idx;
 783         for (unsigned i = 0; i < mVWidth / 2; i++) {
 784             idx.push_back(C(flag ? i + mVWidth / 2 : i));
 785         }
 786         return VSHUFFLE(a, VUNDEF_I(), ConstantVector::get(idx));
 787     }
 788
 789     Value *Builder::VINSERTI128(Value* a, Value* b, Constant* imm8)
 790     {
 791         bool flag = !imm8->isZeroValue();
 792         SmallVector<Constant*,8> idx;
 793         for (unsigned i = 0; i < mVWidth; i++) {
 794             idx.push_back(C(i));
 795         }
 796         Value *inter = VSHUFFLE(b, VUNDEF_I(), ConstantVector::get(idx));
 797
 798         SmallVector<Constant*,8> idx2;
 799         for (unsigned i = 0; i < mVWidth / 2; i++) {
 800             idx2.push_back(C(flag ? i : i + mVWidth));
 801         }
 802         for (unsigned i = mVWidth / 2; i < mVWidth; i++) {
 803             idx2.push_back(C(flag ? i + mVWidth / 2 : i));
 804         }
 805         return VSHUFFLE(a, inter, ConstantVector::get(idx2));
 806     }
 807
 808     // rdtsc buckets macros
 809     void Builder::RDTSC_START(Value* pBucketMgr, Value* pId)
 810     {
 811         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
 812         // buckets framework when single threaded
 813         if (KNOB_SINGLE_THREADED)
 814         {
 815             std::vector<Type*> args{
 816                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
 817                 mInt32Ty                        // id
 818             };
 819
 820             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
 821             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StartBucket", pFuncTy));
 822             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StartBucket") == nullptr)
 823             {
 824                 sys::DynamicLibrary::AddSymbol("BucketManager_StartBucket", (void*)&BucketManager_StartBucket);
 825             }
 826
 827             CALL(pFunc, { pBucketMgr, pId });
 828         }
 829     }
 830
 831     void Builder::RDTSC_STOP(Value* pBucketMgr, Value* pId)
 832     {
 833         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
 834         // buckets framework when single threaded
 835         if (KNOB_SINGLE_THREADED)
 836         {
 837             std::vector<Type*> args{
 838                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
 839                 mInt32Ty                        // id
 840             };
 841
 842             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
 843             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StopBucket", pFuncTy));
 844             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StopBucket") == nullptr)
 845             {
 846                 sys::DynamicLibrary::AddSymbol("BucketManager_StopBucket", (void*)&BucketManager_StopBucket);
 847             }
 848
 849             CALL(pFunc, { pBucketMgr, pId });
 850         }
 851     }
 852
 853     uint32_t Builder::GetTypeSize(Type* pType)
 854     {
 855         if (pType->isStructTy())
 856         {
 857             uint32_t numElems = pType->getStructNumElements();
 858             Type* pElemTy = pType->getStructElementType(0);
 859             return numElems * GetTypeSize(pElemTy);
 860         }
 861
 862         if (pType->isArrayTy())
 863         {
 864             uint32_t numElems = pType->getArrayNumElements();
 865             Type* pElemTy = pType->getArrayElementType();
 866             return numElems * GetTypeSize(pElemTy);
 867         }
 868
 869         if (pType->isIntegerTy())
 870         {
 871             uint32_t bitSize = pType->getIntegerBitWidth();
 872             return bitSize / 8;
 873         }
 874
 875         if (pType->isFloatTy())
 876         {
 877             return 4;
 878         }
 879
 880         if (pType->isHalfTy())
 881         {
 882             return 2;
 883         }
 884
 885         if (pType->isDoubleTy())
 886         {
 887             return 8;
 888         }
 889
 890         SWR_ASSERT(false, "Unimplemented type.");
 891         return 0;
 892     }
 893 }