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(uint64_t i)
 186     {
 187         return ConstantInt::get(IRB()->getInt64Ty(), i);
 188     }
 189
 190     Constant *Builder::C(float i)
 191     {
 192         return ConstantFP::get(IRB()->getFloatTy(), i);
 193     }
 194
 195     Constant *Builder::PRED(bool pred)
 196     {
 197         return ConstantInt::get(IRB()->getInt1Ty(), (pred ? 1 : 0));
 198     }
 199
 200     Value *Builder::VIMMED1(int i)
 201     {
 202         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 203     }
 204
 205     Value *Builder::VIMMED1_16(int i)
 206     {
 207         return ConstantVector::getSplat(mVWidth16, cast<ConstantInt>(C(i)));
 208     }
 209
 210     Value *Builder::VIMMED1(uint32_t i)
 211     {
 212         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 213     }
 214
 215     Value *Builder::VIMMED1_16(uint32_t i)
 216     {
 217         return ConstantVector::getSplat(mVWidth16, cast<ConstantInt>(C(i)));
 218     }
 219
 220     Value *Builder::VIMMED1(float i)
 221     {
 222         return ConstantVector::getSplat(mVWidth, cast<ConstantFP>(C(i)));
 223     }
 224
 225     Value *Builder::VIMMED1_16(float i)
 226     {
 227         return ConstantVector::getSplat(mVWidth16, cast<ConstantFP>(C(i)));
 228     }
 229
 230     Value *Builder::VIMMED1(bool i)
 231     {
 232         return ConstantVector::getSplat(mVWidth, cast<ConstantInt>(C(i)));
 233     }
 234
 235     Value *Builder::VIMMED1_16(bool i)
 236     {
 237         return ConstantVector::getSplat(mVWidth16, cast<ConstantInt>(C(i)));
 238     }
 239
 240     Value *Builder::VUNDEF_IPTR()
 241     {
 242         return UndefValue::get(VectorType::get(mInt32PtrTy,mVWidth));
 243     }
 244
 245     Value *Builder::VUNDEF(Type* t)
 246     {
 247         return UndefValue::get(VectorType::get(t, mVWidth));
 248     }
 249
 250     Value *Builder::VUNDEF_I()
 251     {
 252         return UndefValue::get(VectorType::get(mInt32Ty, mVWidth));
 253     }
 254
 255     Value *Builder::VUNDEF_I_16()
 256     {
 257         return UndefValue::get(VectorType::get(mInt32Ty, mVWidth16));
 258     }
 259
 260     Value *Builder::VUNDEF_F()
 261     {
 262         return UndefValue::get(VectorType::get(mFP32Ty, mVWidth));
 263     }
 264
 265     Value *Builder::VUNDEF_F_16()
 266     {
 267         return UndefValue::get(VectorType::get(mFP32Ty, mVWidth16));
 268     }
 269
 270     Value *Builder::VUNDEF(Type *ty, uint32_t size)
 271     {
 272         return UndefValue::get(VectorType::get(ty, size));
 273     }
 274
 275     Value *Builder::VBROADCAST(Value *src, const llvm::Twine& name)
 276     {
 277         // check if src is already a vector
 278         if (src->getType()->isVectorTy())
 279         {
 280             return src;
 281         }
 282
 283         return VECTOR_SPLAT(mVWidth, src, name);
 284     }
 285
 286     Value *Builder::VBROADCAST_16(Value *src)
 287     {
 288         // check if src is already a vector
 289         if (src->getType()->isVectorTy())
 290         {
 291             return src;
 292         }
 293
 294         return VECTOR_SPLAT(mVWidth16, src);
 295     }
 296
 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     CallInst *Builder::CALL(Value *Callee, const std::initializer_list<Value*> &argsList, const llvm::Twine& name)
 312     {
 313         std::vector<Value*> args;
 314         for (auto arg : argsList)
 315             args.push_back(arg);
 316         return CALLA(Callee, args, name);
 317     }
 318
 319     CallInst *Builder::CALL(Value *Callee, Value* arg)
 320     {
 321         std::vector<Value*> args;
 322         args.push_back(arg);
 323         return CALLA(Callee, args);
 324     }
 325
 326     CallInst *Builder::CALL2(Value *Callee, Value* arg1, Value* arg2)
 327     {
 328         std::vector<Value*> args;
 329         args.push_back(arg1);
 330         args.push_back(arg2);
 331         return CALLA(Callee, args);
 332     }
 333
 334     CallInst *Builder::CALL3(Value *Callee, Value* arg1, Value* arg2, Value* arg3)
 335     {
 336         std::vector<Value*> args;
 337         args.push_back(arg1);
 338         args.push_back(arg2);
 339         args.push_back(arg3);
 340         return CALLA(Callee, args);
 341     }
 342
 343     Value *Builder::VRCP(Value *va, const llvm::Twine& name)
 344     {
 345         return FDIV(VIMMED1(1.0f), va, name);  // 1 / a
 346     }
 347
 348     Value *Builder::VPLANEPS(Value* vA, Value* vB, Value* vC, Value* &vX, Value* &vY)
 349     {
 350         Value* vOut = FMADDPS(vA, vX, vC);
 351         vOut = FMADDPS(vB, vY, vOut);
 352         return vOut;
 353     }
 354
 355     //////////////////////////////////////////////////////////////////////////
 356     /// @brief insert a JIT call to CallPrint
 357     /// - outputs formatted string to both stdout and VS output window
 358     /// - DEBUG builds only
 359     /// Usage example:
 360     ///   PRINT("index %d = 0x%p\n",{C(lane), pIndex});
 361     ///   where C(lane) creates a constant value to print, and pIndex is the Value*
 362     ///   result from a GEP, printing out the pointer to memory
 363     /// @param printStr - constant string to print, which includes format specifiers
 364     /// @param printArgs - initializer list of Value*'s to print to std out
 365     CallInst *Builder::PRINT(const std::string &printStr,const std::initializer_list<Value*> &printArgs)
 366     {
 367         // push the arguments to CallPrint into a vector
 368         std::vector<Value*> printCallArgs;
 369         // save room for the format string.  we still need to modify it for vectors
 370         printCallArgs.resize(1);
 371
 372         // search through the format string for special processing
 373         size_t pos = 0;
 374         std::string tempStr(printStr);
 375         pos = tempStr.find('%', pos);
 376         auto v = printArgs.begin();
 377
 378         while ((pos != std::string::npos) && (v != printArgs.end()))
 379         {
 380             Value* pArg = *v;
 381             Type* pType = pArg->getType();
 382
 383             if (pType->isVectorTy())
 384             {
 385                 Type* pContainedType = pType->getContainedType(0);
 386
 387                 if (toupper(tempStr[pos + 1]) == 'X')
 388                 {
 389                     tempStr[pos] = '0';
 390                     tempStr[pos + 1] = 'x';
 391                     tempStr.insert(pos + 2, "%08X ");
 392                     pos += 7;
 393
 394                     printCallArgs.push_back(VEXTRACT(pArg, C(0)));
 395
 396                     std::string vectorFormatStr;
 397                     for (uint32_t i = 1; i < pType->getVectorNumElements(); ++i)
 398                     {
 399                         vectorFormatStr += "0x%08X ";
 400                         printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 401                     }
 402
 403                     tempStr.insert(pos, vectorFormatStr);
 404                     pos += vectorFormatStr.size();
 405                 }
 406                 else if ((tempStr[pos + 1] == 'f') && (pContainedType->isFloatTy()))
 407                 {
 408                     uint32_t i = 0;
 409                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 410                     {
 411                         tempStr.insert(pos, std::string("%f "));
 412                         pos += 3;
 413                         printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 414                     }
 415                     printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 416                 }
 417                 else if ((tempStr[pos + 1] == 'd') && (pContainedType->isIntegerTy()))
 418                 {
 419                     uint32_t i = 0;
 420                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 421                     {
 422                         tempStr.insert(pos, std::string("%d "));
 423                         pos += 3;
 424                         printCallArgs.push_back(S_EXT(VEXTRACT(pArg, C(i)), Type::getInt32Ty(JM()->mContext)));
 425                     }
 426                     printCallArgs.push_back(S_EXT(VEXTRACT(pArg, C(i)), Type::getInt32Ty(JM()->mContext)));
 427                 }
 428                 else if ((tempStr[pos + 1] == 'u') && (pContainedType->isIntegerTy()))
 429                 {
 430                     uint32_t i = 0;
 431                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 432                     {
 433                         tempStr.insert(pos, std::string("%d "));
 434                         pos += 3;
 435                         printCallArgs.push_back(Z_EXT(VEXTRACT(pArg, C(i)), Type::getInt32Ty(JM()->mContext)));
 436                     }
 437                     printCallArgs.push_back(Z_EXT(VEXTRACT(pArg, C(i)), Type::getInt32Ty(JM()->mContext)));
 438                 }
 439             }
 440             else
 441             {
 442                 if (toupper(tempStr[pos + 1]) == 'X')
 443                 {
 444                     tempStr[pos] = '0';
 445                     tempStr.insert(pos + 1, "x%08");
 446                     printCallArgs.push_back(pArg);
 447                     pos += 3;
 448                 }
 449                 // for %f we need to cast float Values to doubles so that they print out correctly
 450                 else if ((tempStr[pos + 1] == 'f') && (pType->isFloatTy()))
 451                 {
 452                     printCallArgs.push_back(FP_EXT(pArg, Type::getDoubleTy(JM()->mContext)));
 453                     pos++;
 454                 }
 455                 else
 456                 {
 457                     printCallArgs.push_back(pArg);
 458                 }
 459             }
 460
 461             // advance to the next arguement
 462             v++;
 463             pos = tempStr.find('%', ++pos);
 464         }
 465
 466         // create global variable constant string
 467         Constant *constString = ConstantDataArray::getString(JM()->mContext,tempStr,true);
 468         GlobalVariable *gvPtr = new GlobalVariable(constString->getType(),true,GlobalValue::InternalLinkage,constString,"printStr");
 469         JM()->mpCurrentModule->getGlobalList().push_back(gvPtr);
 470
 471         // get a pointer to the first character in the constant string array
 472         std::vector<Constant*> geplist{C(0),C(0)};
 473         Constant *strGEP = ConstantExpr::getGetElementPtr(nullptr, gvPtr,geplist,false);
 474
 475         // insert the pointer to the format string in the argument vector
 476         printCallArgs[0] = strGEP;
 477
 478         // get pointer to CallPrint function and insert decl into the module if needed
 479         std::vector<Type*> args;
 480         args.push_back(PointerType::get(mInt8Ty,0));
 481         FunctionType* callPrintTy = FunctionType::get(Type::getVoidTy(JM()->mContext),args,true);
 482         Function *callPrintFn = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("CallPrint", callPrintTy));
 483
 484         // if we haven't yet added the symbol to the symbol table
 485         if((sys::DynamicLibrary::SearchForAddressOfSymbol("CallPrint")) == nullptr)
 486         {
 487             sys::DynamicLibrary::AddSymbol("CallPrint", (void *)&CallPrint);
 488         }
 489
 490         // insert a call to CallPrint
 491         return CALLA(callPrintFn,printCallArgs);
 492     }
 493
 494     //////////////////////////////////////////////////////////////////////////
 495     /// @brief Wrapper around PRINT with initializer list.
 496     CallInst* Builder::PRINT(const std::string &printStr)
 497     {
 498         return PRINT(printStr, {});
 499     }
 500
 501     Value *Builder::EXTRACT_16(Value *x, uint32_t imm)
 502     {
 503         if (imm == 0)
 504         {
 505             return VSHUFFLE(x, UndefValue::get(x->getType()), { 0, 1, 2, 3, 4, 5, 6, 7 });
 506         }
 507         else
 508         {
 509             return VSHUFFLE(x, UndefValue::get(x->getType()), { 8, 9, 10, 11, 12, 13, 14, 15 });
 510         }
 511     }
 512
 513     Value *Builder::JOIN_16(Value *a, Value *b)
 514     {
 515         return VSHUFFLE(a, b, { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 });
 516     }
 517
 518     //////////////////////////////////////////////////////////////////////////
 519     /// @brief convert x86 <N x float> mask to llvm <N x i1> mask
 520     Value *Builder::MASK(Value *vmask)
 521     {
 522         Value *src = BITCAST(vmask, mSimdInt32Ty);
 523         return ICMP_SLT(src, VIMMED1(0));
 524     }
 525
 526     Value *Builder::MASK_16(Value *vmask)
 527     {
 528         Value *src = BITCAST(vmask, mSimd16Int32Ty);
 529         return ICMP_SLT(src, VIMMED1_16(0));
 530     }
 531
 532     //////////////////////////////////////////////////////////////////////////
 533     /// @brief convert llvm <N x i1> mask to x86 <N x i32> mask
 534     Value *Builder::VMASK(Value *mask)
 535     {
 536         return S_EXT(mask, mSimdInt32Ty);
 537     }
 538
 539     Value *Builder::VMASK_16(Value *mask)
 540     {
 541         return S_EXT(mask, mSimd16Int32Ty);
 542     }
 543
 544     /// @brief Convert <Nxi1> llvm mask to integer
 545     Value *Builder::VMOVMSK(Value* mask)
 546     {
 547         SWR_ASSERT(mask->getType()->getVectorElementType() == mInt1Ty);
 548         uint32_t numLanes = mask->getType()->getVectorNumElements();
 549         Value* i32Result;
 550         if (numLanes == 8)
 551         {
 552             i32Result = BITCAST(mask, mInt8Ty);
 553         }
 554         else if (numLanes == 16)
 555         {
 556             i32Result = BITCAST(mask, mInt16Ty);
 557         }
 558         else
 559         {
 560             SWR_ASSERT("Unsupported vector width");
 561             i32Result = BITCAST(mask, mInt8Ty);
 562         }
 563         return Z_EXT(i32Result, mInt32Ty);
 564     }
 565
 566     //////////////////////////////////////////////////////////////////////////
 567     /// @brief Generate a VPSHUFB operation in LLVM IR.  If not
 568     /// supported on the underlying platform, emulate it
 569     /// @param a - 256bit SIMD(32x8bit) of 8bit integer values
 570     /// @param b - 256bit SIMD(32x8bit) of 8bit integer mask values
 571     /// Byte masks in lower 128 lane of b selects 8 bit values from lower
 572     /// 128bits of a, and vice versa for the upper lanes.  If the mask
 573     /// value is negative, '0' is inserted.
 574     Value *Builder::PSHUFB(Value* a, Value* b)
 575     {
 576         Value* res;
 577         // use avx2 pshufb instruction if available
 578         if(JM()->mArch.AVX2())
 579         {
 580             res = VPSHUFB(a, b);
 581         }
 582         else
 583         {
 584             Constant* cB = dyn_cast<Constant>(b);
 585             // number of 8 bit elements in b
 586             uint32_t numElms = cast<VectorType>(cB->getType())->getNumElements();
 587             // output vector
 588             Value* vShuf = UndefValue::get(VectorType::get(mInt8Ty, numElms));
 589
 590             // insert an 8 bit value from the high and low lanes of a per loop iteration
 591             numElms /= 2;
 592             for(uint32_t i = 0; i < numElms; i++)
 593             {
 594                 ConstantInt* cLow128b = cast<ConstantInt>(cB->getAggregateElement(i));
 595                 ConstantInt* cHigh128b = cast<ConstantInt>(cB->getAggregateElement(i + numElms));
 596
 597                 // extract values from constant mask
 598                 char valLow128bLane =  (char)(cLow128b->getSExtValue());
 599                 char valHigh128bLane = (char)(cHigh128b->getSExtValue());
 600
 601                 Value* insertValLow128b;
 602                 Value* insertValHigh128b;
 603
 604                 // if the mask value is negative, insert a '0' in the respective output position
 605                 // otherwise, lookup the value at mask position (bits 3..0 of the respective mask byte) in a and insert in output vector
 606                 insertValLow128b = (valLow128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valLow128bLane & 0xF)));
 607                 insertValHigh128b = (valHigh128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valHigh128bLane & 0xF) + numElms));
 608
 609                 vShuf = VINSERT(vShuf, insertValLow128b, i);
 610                 vShuf = VINSERT(vShuf, insertValHigh128b, (i + numElms));
 611             }
 612             res = vShuf;
 613         }
 614         return res;
 615     }
 616
 617     //////////////////////////////////////////////////////////////////////////
 618     /// @brief Generate a VPSHUFB operation (sign extend 8 8bit values to 32
 619     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 620     /// @param a - 128bit SIMD lane(16x8bit) of 8bit integer values.  Only
 621     /// lower 8 values are used.
 622     Value *Builder::PMOVSXBD(Value* a)
 623     {
 624         // VPMOVSXBD output type
 625         Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 626         // Extract 8 values from 128bit lane and sign extend
 627         return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 628     }
 629
 630     //////////////////////////////////////////////////////////////////////////
 631     /// @brief Generate a VPSHUFB operation (sign extend 8 16bit values to 32
 632     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 633     /// @param a - 128bit SIMD lane(8x16bit) of 16bit integer values.
 634     Value *Builder::PMOVSXWD(Value* a)
 635     {
 636         // VPMOVSXWD output type
 637         Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 638         // Extract 8 values from 128bit lane and sign extend
 639         return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 640     }
 641
 642     //////////////////////////////////////////////////////////////////////////
 643     /// @brief Generate a VCVTPH2PS operation (float16->float32 conversion)
 644     /// in LLVM IR.  If not supported on the underlying platform, emulate it
 645     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
 646     Value *Builder::CVTPH2PS(Value* a, const llvm::Twine& name)
 647     {
 648         if (JM()->mArch.F16C())
 649         {
 650             return VCVTPH2PS(a, name);
 651         }
 652         else
 653         {
 654             FunctionType* pFuncTy = FunctionType::get(mFP32Ty, mInt16Ty);
 655             Function* pCvtPh2Ps = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("ConvertFloat16ToFloat32", pFuncTy));
 656
 657             if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat16ToFloat32") == nullptr)
 658             {
 659                 sys::DynamicLibrary::AddSymbol("ConvertFloat16ToFloat32", (void *)&ConvertFloat16ToFloat32);
 660             }
 661
 662             Value* pResult = UndefValue::get(mSimdFP32Ty);
 663             for (uint32_t i = 0; i < mVWidth; ++i)
 664             {
 665                 Value* pSrc = VEXTRACT(a, C(i));
 666                 Value* pConv = CALL(pCvtPh2Ps, std::initializer_list<Value*>{pSrc});
 667                 pResult = VINSERT(pResult, pConv, C(i));
 668             }
 669
 670             pResult->setName(name);
 671             return pResult;
 672         }
 673     }
 674
 675     //////////////////////////////////////////////////////////////////////////
 676     /// @brief Generate a VCVTPS2PH operation (float32->float16 conversion)
 677     /// in LLVM IR.  If not supported on the underlying platform, emulate it
 678     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
 679     Value *Builder::CVTPS2PH(Value* a, Value* rounding)
 680     {
 681         if (JM()->mArch.F16C())
 682         {
 683             return VCVTPS2PH(a, rounding);
 684         }
 685         else
 686         {
 687             // call scalar C function for now
 688             FunctionType* pFuncTy = FunctionType::get(mInt16Ty, mFP32Ty);
 689             Function* pCvtPs2Ph = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("ConvertFloat32ToFloat16", pFuncTy));
 690
 691             if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat32ToFloat16") == nullptr)
 692             {
 693                 sys::DynamicLibrary::AddSymbol("ConvertFloat32ToFloat16", (void *)&ConvertFloat32ToFloat16);
 694             }
 695
 696             Value* pResult = UndefValue::get(mSimdInt16Ty);
 697             for (uint32_t i = 0; i < mVWidth; ++i)
 698             {
 699                 Value* pSrc = VEXTRACT(a, C(i));
 700                 Value* pConv = CALL(pCvtPs2Ph, std::initializer_list<Value*>{pSrc});
 701                 pResult = VINSERT(pResult, pConv, C(i));
 702             }
 703
 704             return pResult;
 705         }
 706     }
 707
 708     Value *Builder::PMAXSD(Value* a, Value* b)
 709     {
 710         Value* cmp = ICMP_SGT(a, b);
 711         return SELECT(cmp, a, b);
 712     }
 713
 714     Value *Builder::PMINSD(Value* a, Value* b)
 715     {
 716         Value* cmp = ICMP_SLT(a, b);
 717         return SELECT(cmp, a, b);
 718     }
 719
 720     Value *Builder::PMAXUD(Value* a, Value* b)
 721     {
 722         Value* cmp = ICMP_UGT(a, b);
 723         return SELECT(cmp, a, b);
 724     }
 725
 726     Value *Builder::PMINUD(Value* a, Value* b)
 727     {
 728         Value* cmp = ICMP_ULT(a, b);
 729         return SELECT(cmp, a, b);
 730     }
 731
 732     // Helper function to create alloca in entry block of function
 733     Value* Builder::CreateEntryAlloca(Function* pFunc, Type* pType)
 734     {
 735         auto saveIP = IRB()->saveIP();
 736         IRB()->SetInsertPoint(&pFunc->getEntryBlock(),
 737                               pFunc->getEntryBlock().begin());
 738         Value* pAlloca = ALLOCA(pType);
 739         if (saveIP.isSet()) IRB()->restoreIP(saveIP);
 740         return pAlloca;
 741     }
 742
 743     Value* Builder::CreateEntryAlloca(Function* pFunc, Type* pType, Value* pArraySize)
 744     {
 745         auto saveIP = IRB()->saveIP();
 746         IRB()->SetInsertPoint(&pFunc->getEntryBlock(),
 747             pFunc->getEntryBlock().begin());
 748         Value* pAlloca = ALLOCA(pType, pArraySize);
 749         if (saveIP.isSet()) IRB()->restoreIP(saveIP);
 750         return pAlloca;
 751     }
 752
 753     Value* Builder::VABSPS(Value* a)
 754     {
 755         Value* asInt = BITCAST(a, mSimdInt32Ty);
 756         Value* result = BITCAST(AND(asInt, VIMMED1(0x7fffffff)), mSimdFP32Ty);
 757         return result;
 758     }
 759
 760     Value *Builder::ICLAMP(Value* src, Value* low, Value* high, const llvm::Twine& name)
 761     {
 762         Value *lowCmp = ICMP_SLT(src, low);
 763         Value *ret = SELECT(lowCmp, low, src);
 764
 765         Value *highCmp = ICMP_SGT(ret, high);
 766         ret = SELECT(highCmp, high, ret, name);
 767
 768         return ret;
 769     }
 770
 771     Value *Builder::FCLAMP(Value* src, Value* low, Value* high)
 772     {
 773         Value *lowCmp = FCMP_OLT(src, low);
 774         Value *ret = SELECT(lowCmp, low, src);
 775
 776         Value *highCmp = FCMP_OGT(ret, high);
 777         ret = SELECT(highCmp, high, ret);
 778
 779         return ret;
 780     }
 781
 782     Value *Builder::FCLAMP(Value* src, float low, float high)
 783     {
 784         Value* result = VMAXPS(src, VIMMED1(low));
 785         result = VMINPS(result, VIMMED1(high));
 786
 787         return result;
 788     }
 789
 790     Value *Builder::FMADDPS(Value* a, Value* b, Value* c)
 791     {
 792         Value* vOut;
 793         // use FMADs if available
 794         if(JM()->mArch.AVX2())
 795         {
 796             vOut = VFMADDPS(a, b, c);
 797         }
 798         else
 799         {
 800             vOut = FADD(FMUL(a, b), c);
 801         }
 802         return vOut;
 803     }
 804
 805     //////////////////////////////////////////////////////////////////////////
 806     /// @brief pop count on vector mask (e.g. <8 x i1>)
 807     Value* Builder::VPOPCNT(Value* a)
 808     {
 809         return POPCNT(VMOVMSK(a));
 810     }
 811
 812     //////////////////////////////////////////////////////////////////////////
 813     /// @brief C functions called by LLVM IR
 814     //////////////////////////////////////////////////////////////////////////
 815
 816     Value *Builder::VEXTRACTI128(Value* a, Constant* imm8)
 817     {
 818         bool flag = !imm8->isZeroValue();
 819         SmallVector<Constant*,8> idx;
 820         for (unsigned i = 0; i < mVWidth / 2; i++) {
 821             idx.push_back(C(flag ? i + mVWidth / 2 : i));
 822         }
 823         return VSHUFFLE(a, VUNDEF_I(), ConstantVector::get(idx));
 824     }
 825
 826     Value *Builder::VINSERTI128(Value* a, Value* b, Constant* imm8)
 827     {
 828         bool flag = !imm8->isZeroValue();
 829         SmallVector<Constant*,8> idx;
 830         for (unsigned i = 0; i < mVWidth; i++) {
 831             idx.push_back(C(i));
 832         }
 833         Value *inter = VSHUFFLE(b, VUNDEF_I(), ConstantVector::get(idx));
 834
 835         SmallVector<Constant*,8> idx2;
 836         for (unsigned i = 0; i < mVWidth / 2; i++) {
 837             idx2.push_back(C(flag ? i : i + mVWidth));
 838         }
 839         for (unsigned i = mVWidth / 2; i < mVWidth; i++) {
 840             idx2.push_back(C(flag ? i + mVWidth / 2 : i));
 841         }
 842         return VSHUFFLE(a, inter, ConstantVector::get(idx2));
 843     }
 844
 845     // rdtsc buckets macros
 846     void Builder::RDTSC_START(Value* pBucketMgr, Value* pId)
 847     {
 848         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
 849         // buckets framework when single threaded
 850         if (KNOB_SINGLE_THREADED)
 851         {
 852             std::vector<Type*> args{
 853                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
 854                 mInt32Ty                        // id
 855             };
 856
 857             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
 858             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StartBucket", pFuncTy));
 859             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StartBucket") == nullptr)
 860             {
 861                 sys::DynamicLibrary::AddSymbol("BucketManager_StartBucket", (void*)&BucketManager_StartBucket);
 862             }
 863
 864             CALL(pFunc, { pBucketMgr, pId });
 865         }
 866     }
 867
 868     void Builder::RDTSC_STOP(Value* pBucketMgr, Value* pId)
 869     {
 870         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
 871         // buckets framework when single threaded
 872         if (KNOB_SINGLE_THREADED)
 873         {
 874             std::vector<Type*> args{
 875                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
 876                 mInt32Ty                        // id
 877             };
 878
 879             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
 880             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StopBucket", pFuncTy));
 881             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StopBucket") == nullptr)
 882             {
 883                 sys::DynamicLibrary::AddSymbol("BucketManager_StopBucket", (void*)&BucketManager_StopBucket);
 884             }
 885
 886             CALL(pFunc, { pBucketMgr, pId });
 887         }
 888     }
 889
 890     uint32_t Builder::GetTypeSize(Type* pType)
 891     {
 892         if (pType->isStructTy())
 893         {
 894             uint32_t numElems = pType->getStructNumElements();
 895             Type* pElemTy = pType->getStructElementType(0);
 896             return numElems * GetTypeSize(pElemTy);
 897         }
 898
 899         if (pType->isArrayTy())
 900         {
 901             uint32_t numElems = pType->getArrayNumElements();
 902             Type* pElemTy = pType->getArrayElementType();
 903             return numElems * GetTypeSize(pElemTy);
 904         }
 905
 906         if (pType->isIntegerTy())
 907         {
 908             uint32_t bitSize = pType->getIntegerBitWidth();
 909             return bitSize / 8;
 910         }
 911
 912         if (pType->isFloatTy())
 913         {
 914             return 4;
 915         }
 916
 917         if (pType->isHalfTy())
 918         {
 919             return 2;
 920         }
 921
 922         if (pType->isDoubleTy())
 923         {
 924             return 8;
 925         }
 926
 927         SWR_ASSERT(false, "Unimplemented type.");
 928         return 0;
 929     }
 930 }