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 Convert32To16Float(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 ConvertSmallFloatTo32(UINT val)
 116     {
 117         UINT 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     Value *Builder::VUNDEF_IPTR()
 215     {
 216         return UndefValue::get(VectorType::get(mInt32PtrTy,mVWidth));
 217     }
 218
 219     Value *Builder::VUNDEF_I()
 220     {
 221         return UndefValue::get(VectorType::get(mInt32Ty, mVWidth));
 222     }
 223
 224     Value *Builder::VUNDEF(Type *ty, uint32_t size)
 225     {
 226         return UndefValue::get(VectorType::get(ty, size));
 227     }
 228
 229     Value *Builder::VUNDEF_F()
 230     {
 231         return UndefValue::get(VectorType::get(mFP32Ty, mVWidth));
 232     }
 233
 234     Value *Builder::VUNDEF(Type* t)
 235     {
 236         return UndefValue::get(VectorType::get(t, mVWidth));
 237     }
 238
 239     #if HAVE_LLVM == 0x306
 240     Value *Builder::VINSERT(Value *vec, Value *val, uint64_t index)
 241     {
 242         return VINSERT(vec, val, C((int64_t)index));
 243     }
 244     #endif
 245
 246     Value *Builder::VBROADCAST(Value *src)
 247     {
 248         // check if src is already a vector
 249         if (src->getType()->isVectorTy())
 250         {
 251             return src;
 252         }
 253
 254         return VECTOR_SPLAT(mVWidth, src);
 255     }
 256
 257     uint32_t Builder::IMMED(Value* v)
 258     {
 259         SWR_ASSERT(isa<ConstantInt>(v));
 260         ConstantInt *pValConst = cast<ConstantInt>(v);
 261         return pValConst->getZExtValue();
 262     }
 263
 264     int32_t Builder::S_IMMED(Value* v)
 265     {
 266         SWR_ASSERT(isa<ConstantInt>(v));
 267         ConstantInt *pValConst = cast<ConstantInt>(v);
 268         return pValConst->getSExtValue();
 269     }
 270
 271     Value *Builder::GEP(Value* ptr, const std::initializer_list<Value*> &indexList)
 272     {
 273         std::vector<Value*> indices;
 274         for (auto i : indexList)
 275             indices.push_back(i);
 276         return GEPA(ptr, indices);
 277     }
 278
 279     Value *Builder::GEP(Value* ptr, const std::initializer_list<uint32_t> &indexList)
 280     {
 281         std::vector<Value*> indices;
 282         for (auto i : indexList)
 283             indices.push_back(C(i));
 284         return GEPA(ptr, indices);
 285     }
 286
 287     LoadInst *Builder::LOAD(Value *basePtr, const std::initializer_list<uint32_t> &indices, const llvm::Twine& name)
 288     {
 289         std::vector<Value*> valIndices;
 290         for (auto i : indices)
 291             valIndices.push_back(C(i));
 292         return LOAD(GEPA(basePtr, valIndices), name);
 293     }
 294
 295     LoadInst *Builder::LOADV(Value *basePtr, const std::initializer_list<Value*> &indices, const llvm::Twine& name)
 296     {
 297         std::vector<Value*> valIndices;
 298         for (auto i : indices)
 299             valIndices.push_back(i);
 300         return LOAD(GEPA(basePtr, valIndices), name);
 301     }
 302
 303     StoreInst *Builder::STORE(Value *val, Value *basePtr, const std::initializer_list<uint32_t> &indices)
 304     {
 305         std::vector<Value*> valIndices;
 306         for (auto i : indices)
 307             valIndices.push_back(C(i));
 308         return STORE(val, GEPA(basePtr, valIndices));
 309     }
 310
 311     StoreInst *Builder::STOREV(Value *val, Value *basePtr, const std::initializer_list<Value*> &indices)
 312     {
 313         std::vector<Value*> valIndices;
 314         for (auto i : indices)
 315             valIndices.push_back(i);
 316         return STORE(val, GEPA(basePtr, valIndices));
 317     }
 318
 319     CallInst *Builder::CALL(Value *Callee, const std::initializer_list<Value*> &argsList)
 320     {
 321         std::vector<Value*> args;
 322         for (auto arg : argsList)
 323             args.push_back(arg);
 324         return CALLA(Callee, args);
 325     }
 326
 327     #if HAVE_LLVM > 0x306
 328     CallInst *Builder::CALL(Value *Callee, Value* arg)
 329     {
 330         std::vector<Value*> args;
 331         args.push_back(arg);
 332         return CALLA(Callee, args);
 333     }
 334
 335     CallInst *Builder::CALL2(Value *Callee, Value* arg1, Value* arg2)
 336     {
 337         std::vector<Value*> args;
 338         args.push_back(arg1);
 339         args.push_back(arg2);
 340         return CALLA(Callee, args);
 341     }
 342
 343     CallInst *Builder::CALL3(Value *Callee, Value* arg1, Value* arg2, Value* arg3)
 344     {
 345         std::vector<Value*> args;
 346         args.push_back(arg1);
 347         args.push_back(arg2);
 348         args.push_back(arg3);
 349         return CALLA(Callee, args);
 350     }
 351     #endif
 352
 353     Value *Builder::VRCP(Value *va)
 354     {
 355         return FDIV(VIMMED1(1.0f), va);  // 1 / a
 356     }
 357
 358     Value *Builder::VPLANEPS(Value* vA, Value* vB, Value* vC, Value* &vX, Value* &vY)
 359     {
 360         Value* vOut = FMADDPS(vA, vX, vC);
 361         vOut = FMADDPS(vB, vY, vOut);
 362         return vOut;
 363     }
 364
 365     //////////////////////////////////////////////////////////////////////////
 366     /// @brief Generate an i32 masked load operation in LLVM IR.  If not
 367     /// supported on the underlying platform, emulate it with float masked load
 368     /// @param src - base address pointer for the load
 369     /// @param vMask - SIMD wide mask that controls whether to access memory load 0
 370     Value *Builder::MASKLOADD(Value* src,Value* mask)
 371     {
 372         Value* vResult;
 373         // use avx2 gather instruction is available
 374         if(JM()->mArch.AVX2())
 375         {
 376             Function *func = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_avx2_maskload_d_256);
 377             vResult = CALL(func,{src,mask});
 378         }
 379         else
 380         {
 381             // maskload intrinsic expects integer mask operand in llvm >= 3.8
 382     #if (LLVM_VERSION_MAJOR > 3) || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 8)
 383             mask = BITCAST(mask,VectorType::get(mInt32Ty,mVWidth));
 384     #else
 385             mask = BITCAST(mask,VectorType::get(mFP32Ty,mVWidth));
 386     #endif
 387             Function *func = Intrinsic::getDeclaration(JM()->mpCurrentModule,Intrinsic::x86_avx_maskload_ps_256);
 388             vResult = BITCAST(CALL(func,{src,mask}), VectorType::get(mInt32Ty,mVWidth));
 389         }
 390         return vResult;
 391     }
 392
 393     //////////////////////////////////////////////////////////////////////////
 394     /// @brief insert a JIT call to CallPrint
 395     /// - outputs formatted string to both stdout and VS output window
 396     /// - DEBUG builds only
 397     /// Usage example:
 398     ///   PRINT("index %d = 0x%p\n",{C(lane), pIndex});
 399     ///   where C(lane) creates a constant value to print, and pIndex is the Value*
 400     ///   result from a GEP, printing out the pointer to memory
 401     /// @param printStr - constant string to print, which includes format specifiers
 402     /// @param printArgs - initializer list of Value*'s to print to std out
 403     CallInst *Builder::PRINT(const std::string &printStr,const std::initializer_list<Value*> &printArgs)
 404     {
 405         // push the arguments to CallPrint into a vector
 406         std::vector<Value*> printCallArgs;
 407         // save room for the format string.  we still need to modify it for vectors
 408         printCallArgs.resize(1);
 409
 410         // search through the format string for special processing
 411         size_t pos = 0;
 412         std::string tempStr(printStr);
 413         pos = tempStr.find('%', pos);
 414         auto v = printArgs.begin();
 415
 416         while ((pos != std::string::npos) && (v != printArgs.end()))
 417         {
 418             Value* pArg = *v;
 419             Type* pType = pArg->getType();
 420
 421             if (pType->isVectorTy())
 422             {
 423                 Type* pContainedType = pType->getContainedType(0);
 424
 425                 if (toupper(tempStr[pos + 1]) == 'X')
 426                 {
 427                     tempStr[pos] = '0';
 428                     tempStr[pos + 1] = 'x';
 429                     tempStr.insert(pos + 2, "%08X ");
 430                     pos += 7;
 431
 432                     printCallArgs.push_back(VEXTRACT(pArg, C(0)));
 433
 434                     std::string vectorFormatStr;
 435                     for (uint32_t i = 1; i < pType->getVectorNumElements(); ++i)
 436                     {
 437                         vectorFormatStr += "0x%08X ";
 438                         printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 439                     }
 440
 441                     tempStr.insert(pos, vectorFormatStr);
 442                     pos += vectorFormatStr.size();
 443                 }
 444                 else if ((tempStr[pos + 1] == 'f') && (pContainedType->isFloatTy()))
 445                 {
 446                     uint32_t i = 0;
 447                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 448                     {
 449                         tempStr.insert(pos, std::string("%f "));
 450                         pos += 3;
 451                         printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 452                     }
 453                     printCallArgs.push_back(FP_EXT(VEXTRACT(pArg, C(i)), Type::getDoubleTy(JM()->mContext)));
 454                 }
 455                 else if ((tempStr[pos + 1] == 'd') && (pContainedType->isIntegerTy()))
 456                 {
 457                     uint32_t i = 0;
 458                     for (; i < (pArg->getType()->getVectorNumElements()) - 1; i++)
 459                     {
 460                         tempStr.insert(pos, std::string("%d "));
 461                         pos += 3;
 462                         printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 463                     }
 464                     printCallArgs.push_back(VEXTRACT(pArg, C(i)));
 465                 }
 466             }
 467             else
 468             {
 469                 if (toupper(tempStr[pos + 1]) == 'X')
 470                 {
 471                     tempStr[pos] = '0';
 472                     tempStr.insert(pos + 1, "x%08");
 473                     printCallArgs.push_back(pArg);
 474                     pos += 3;
 475                 }
 476                 // for %f we need to cast float Values to doubles so that they print out correctly
 477                 else if ((tempStr[pos + 1] == 'f') && (pType->isFloatTy()))
 478                 {
 479                     printCallArgs.push_back(FP_EXT(pArg, Type::getDoubleTy(JM()->mContext)));
 480                     pos++;
 481                 }
 482                 else
 483                 {
 484                     printCallArgs.push_back(pArg);
 485                 }
 486             }
 487
 488             // advance to the next arguement
 489             v++;
 490             pos = tempStr.find('%', ++pos);
 491         }
 492
 493         // create global variable constant string
 494         Constant *constString = ConstantDataArray::getString(JM()->mContext,tempStr,true);
 495         GlobalVariable *gvPtr = new GlobalVariable(constString->getType(),true,GlobalValue::InternalLinkage,constString,"printStr");
 496         JM()->mpCurrentModule->getGlobalList().push_back(gvPtr);
 497
 498         // get a pointer to the first character in the constant string array
 499         std::vector<Constant*> geplist{C(0),C(0)};
 500     #if HAVE_LLVM == 0x306
 501         Constant *strGEP = ConstantExpr::getGetElementPtr(gvPtr,geplist,false);
 502     #else
 503         Constant *strGEP = ConstantExpr::getGetElementPtr(nullptr, gvPtr,geplist,false);
 504     #endif
 505
 506         // insert the pointer to the format string in the argument vector
 507         printCallArgs[0] = strGEP;
 508
 509         // get pointer to CallPrint function and insert decl into the module if needed
 510         std::vector<Type*> args;
 511         args.push_back(PointerType::get(mInt8Ty,0));
 512         FunctionType* callPrintTy = FunctionType::get(Type::getVoidTy(JM()->mContext),args,true);
 513         Function *callPrintFn = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("CallPrint", callPrintTy));
 514
 515         // if we haven't yet added the symbol to the symbol table
 516         if((sys::DynamicLibrary::SearchForAddressOfSymbol("CallPrint")) == nullptr)
 517         {
 518             sys::DynamicLibrary::AddSymbol("CallPrint", (void *)&CallPrint);
 519         }
 520
 521         // insert a call to CallPrint
 522         return CALLA(callPrintFn,printCallArgs);
 523     }
 524
 525     //////////////////////////////////////////////////////////////////////////
 526     /// @brief Wrapper around PRINT with initializer list.
 527     CallInst* Builder::PRINT(const std::string &printStr)
 528     {
 529         return PRINT(printStr, {});
 530     }
 531
 532     //////////////////////////////////////////////////////////////////////////
 533     /// @brief Generate a masked gather operation in LLVM IR.  If not
 534     /// supported on the underlying platform, emulate it with loads
 535     /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
 536     /// @param pBase - Int8* base VB address pointer value
 537     /// @param vIndices - SIMD wide value of VB byte offsets
 538     /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
 539     /// @param scale - value to scale indices by
 540     Value *Builder::GATHERPS(Value* vSrc, Value* pBase, Value* vIndices, Value* vMask, Value* scale)
 541     {
 542         Value* vGather;
 543
 544         // use avx2 gather instruction if available
 545         if(JM()->mArch.AVX2())
 546         {
 547             // force mask to <N x float>, required by vgather
 548             vMask = BITCAST(vMask, mSimdFP32Ty);
 549             vGather = VGATHERPS(vSrc,pBase,vIndices,vMask,scale);
 550         }
 551         else
 552         {
 553             Value* pStack = STACKSAVE();
 554
 555             // store vSrc on the stack.  this way we can select between a valid load address and the vSrc address
 556             Value* vSrcPtr = ALLOCA(vSrc->getType());
 557             STORE(vSrc, vSrcPtr);
 558
 559             vGather = VUNDEF_F();
 560             Value *vScaleVec = VBROADCAST(Z_EXT(scale,mInt32Ty));
 561             Value *vOffsets = MUL(vIndices,vScaleVec);
 562             Value *mask = MASK(vMask);
 563             for(uint32_t i = 0; i < mVWidth; ++i)
 564             {
 565                 // single component byte index
 566                 Value *offset = VEXTRACT(vOffsets,C(i));
 567                 // byte pointer to component
 568                 Value *loadAddress = GEP(pBase,offset);
 569                 loadAddress = BITCAST(loadAddress,PointerType::get(mFP32Ty,0));
 570                 // pointer to the value to load if we're masking off a component
 571                 Value *maskLoadAddress = GEP(vSrcPtr,{C(0), C(i)});
 572                 Value *selMask = VEXTRACT(mask,C(i));
 573                 // switch in a safe address to load if we're trying to access a vertex
 574                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
 575                 Value *val = LOAD(validAddress);
 576                 vGather = VINSERT(vGather,val,C(i));
 577             }
 578             STACKRESTORE(pStack);
 579         }
 580
 581         return vGather;
 582     }
 583
 584     //////////////////////////////////////////////////////////////////////////
 585     /// @brief Generate a masked gather operation in LLVM IR.  If not
 586     /// supported on the underlying platform, emulate it with loads
 587     /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
 588     /// @param pBase - Int8* base VB address pointer value
 589     /// @param vIndices - SIMD wide value of VB byte offsets
 590     /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
 591     /// @param scale - value to scale indices by
 592     Value *Builder::GATHERDD(Value* vSrc, Value* pBase, Value* vIndices, Value* vMask, Value* scale)
 593     {
 594         Value* vGather;
 595
 596         // use avx2 gather instruction if available
 597         if(JM()->mArch.AVX2())
 598         {
 599             vGather = VGATHERDD(vSrc, pBase, vIndices, vMask, scale);
 600         }
 601         else
 602         {
 603             Value* pStack = STACKSAVE();
 604
 605             // store vSrc on the stack.  this way we can select between a valid load address and the vSrc address
 606             Value* vSrcPtr = ALLOCA(vSrc->getType());
 607             STORE(vSrc, vSrcPtr);
 608
 609             vGather = VUNDEF_I();
 610             Value *vScaleVec = VBROADCAST(Z_EXT(scale, mInt32Ty));
 611             Value *vOffsets = MUL(vIndices, vScaleVec);
 612             Value *mask = MASK(vMask);
 613             for(uint32_t i = 0; i < mVWidth; ++i)
 614             {
 615                 // single component byte index
 616                 Value *offset = VEXTRACT(vOffsets, C(i));
 617                 // byte pointer to component
 618                 Value *loadAddress = GEP(pBase, offset);
 619                 loadAddress = BITCAST(loadAddress, PointerType::get(mInt32Ty, 0));
 620                 // pointer to the value to load if we're masking off a component
 621                 Value *maskLoadAddress = GEP(vSrcPtr, {C(0), C(i)});
 622                 Value *selMask = VEXTRACT(mask, C(i));
 623                 // switch in a safe address to load if we're trying to access a vertex
 624                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
 625                 Value *val = LOAD(validAddress, C(0));
 626                 vGather = VINSERT(vGather, val, C(i));
 627             }
 628
 629             STACKRESTORE(pStack);
 630         }
 631         return vGather;
 632     }
 633
 634     //////////////////////////////////////////////////////////////////////////
 635     /// @brief convert x86 <N x float> mask to llvm <N x i1> mask
 636     Value* Builder::MASK(Value* vmask)
 637     {
 638         Value* src = BITCAST(vmask, mSimdInt32Ty);
 639         return ICMP_SLT(src, VIMMED1(0));
 640     }
 641
 642     //////////////////////////////////////////////////////////////////////////
 643     /// @brief convert llvm <N x i1> mask to x86 <N x i32> mask
 644     Value* Builder::VMASK(Value* mask)
 645     {
 646         return S_EXT(mask, mSimdInt32Ty);
 647     }
 648
 649     //////////////////////////////////////////////////////////////////////////
 650     /// @brief Generate a VPSHUFB operation in LLVM IR.  If not
 651     /// supported on the underlying platform, emulate it
 652     /// @param a - 256bit SIMD(32x8bit) of 8bit integer values
 653     /// @param b - 256bit SIMD(32x8bit) of 8bit integer mask values
 654     /// Byte masks in lower 128 lane of b selects 8 bit values from lower
 655     /// 128bits of a, and vice versa for the upper lanes.  If the mask
 656     /// value is negative, '0' is inserted.
 657     Value *Builder::PSHUFB(Value* a, Value* b)
 658     {
 659         Value* res;
 660         // use avx2 pshufb instruction if available
 661         if(JM()->mArch.AVX2())
 662         {
 663             res = VPSHUFB(a, b);
 664         }
 665         else
 666         {
 667             Constant* cB = dyn_cast<Constant>(b);
 668             // number of 8 bit elements in b
 669             uint32_t numElms = cast<VectorType>(cB->getType())->getNumElements();
 670             // output vector
 671             Value* vShuf = UndefValue::get(VectorType::get(mInt8Ty, numElms));
 672
 673             // insert an 8 bit value from the high and low lanes of a per loop iteration
 674             numElms /= 2;
 675             for(uint32_t i = 0; i < numElms; i++)
 676             {
 677                 ConstantInt* cLow128b = cast<ConstantInt>(cB->getAggregateElement(i));
 678                 ConstantInt* cHigh128b = cast<ConstantInt>(cB->getAggregateElement(i + numElms));
 679
 680                 // extract values from constant mask
 681                 char valLow128bLane =  (char)(cLow128b->getSExtValue());
 682                 char valHigh128bLane = (char)(cHigh128b->getSExtValue());
 683
 684                 Value* insertValLow128b;
 685                 Value* insertValHigh128b;
 686
 687                 // if the mask value is negative, insert a '0' in the respective output position
 688                 // otherwise, lookup the value at mask position (bits 3..0 of the respective mask byte) in a and insert in output vector
 689                 insertValLow128b = (valLow128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valLow128bLane & 0xF)));
 690                 insertValHigh128b = (valHigh128bLane < 0) ? C((char)0) : VEXTRACT(a, C((valHigh128bLane & 0xF) + numElms));
 691
 692                 vShuf = VINSERT(vShuf, insertValLow128b, i);
 693                 vShuf = VINSERT(vShuf, insertValHigh128b, (i + numElms));
 694             }
 695             res = vShuf;
 696         }
 697         return res;
 698     }
 699
 700     //////////////////////////////////////////////////////////////////////////
 701     /// @brief Generate a VPSHUFB operation (sign extend 8 8bit values to 32
 702     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 703     /// @param a - 128bit SIMD lane(16x8bit) of 8bit integer values.  Only
 704     /// lower 8 values are used.
 705     Value *Builder::PMOVSXBD(Value* a)
 706     {
 707         // llvm-3.9 removed the pmovsxbd intrinsic
 708     #if HAVE_LLVM < 0x309
 709         // use avx2 byte sign extend instruction if available
 710         if(JM()->mArch.AVX2())
 711         {
 712             Function *pmovsxbd = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_avx2_pmovsxbd);
 713             return CALL(pmovsxbd, std::initializer_list<Value*>{a});
 714         }
 715         else
 716     #endif
 717         {
 718             // VPMOVSXBD output type
 719             Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 720             // Extract 8 values from 128bit lane and sign extend
 721             return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 722         }
 723     }
 724
 725     //////////////////////////////////////////////////////////////////////////
 726     /// @brief Generate a VPSHUFB operation (sign extend 8 16bit values to 32
 727     /// bits)in LLVM IR.  If not supported on the underlying platform, emulate it
 728     /// @param a - 128bit SIMD lane(8x16bit) of 16bit integer values.
 729     Value *Builder::PMOVSXWD(Value* a)
 730     {
 731         // llvm-3.9 removed the pmovsxwd intrinsic
 732     #if HAVE_LLVM < 0x309
 733         // use avx2 word sign extend if available
 734         if(JM()->mArch.AVX2())
 735         {
 736             Function *pmovsxwd = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_avx2_pmovsxwd);
 737             return CALL(pmovsxwd, std::initializer_list<Value*>{a});
 738         }
 739         else
 740     #endif
 741         {
 742             // VPMOVSXWD output type
 743             Type* v8x32Ty = VectorType::get(mInt32Ty, 8);
 744             // Extract 8 values from 128bit lane and sign extend
 745             return S_EXT(VSHUFFLE(a, a, C<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty);
 746         }
 747     }
 748
 749     //////////////////////////////////////////////////////////////////////////
 750     /// @brief Generate a VPERMD operation (shuffle 32 bit integer values
 751     /// across 128 bit lanes) in LLVM IR.  If not supported on the underlying
 752     /// platform, emulate it
 753     /// @param a - 256bit SIMD lane(8x32bit) of integer values.
 754     /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
 755     Value *Builder::PERMD(Value* a, Value* idx)
 756     {
 757         Value* res;
 758         // use avx2 permute instruction if available
 759         if(JM()->mArch.AVX2())
 760         {
 761             res = VPERMD(a, idx);
 762         }
 763         else
 764         {
 765             if (isa<Constant>(idx))
 766             {
 767                 res = VSHUFFLE(a, a, idx);
 768             }
 769             else
 770             {
 771                 res = VUNDEF_I();
 772                 for (uint32_t l = 0; l < JM()->mVWidth; ++l)
 773                 {
 774                     Value* pIndex = VEXTRACT(idx, C(l));
 775                     Value* pVal = VEXTRACT(a, pIndex);
 776                     res = VINSERT(res, pVal, C(l));
 777                 }
 778             }
 779         }
 780         return res;
 781     }
 782
 783     //////////////////////////////////////////////////////////////////////////
 784     /// @brief Generate a VPERMPS operation (shuffle 32 bit float values
 785     /// across 128 bit lanes) in LLVM IR.  If not supported on the underlying
 786     /// platform, emulate it
 787     /// @param a - 256bit SIMD lane(8x32bit) of float values.
 788     /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
 789     Value *Builder::PERMPS(Value* a, Value* idx)
 790     {
 791         Value* res;
 792         // use avx2 permute instruction if available
 793         if (JM()->mArch.AVX2())
 794         {
 795             // llvm 3.6.0 swapped the order of the args to vpermd
 796             res = VPERMPS(idx, a);
 797         }
 798         else
 799         {
 800             if (isa<Constant>(idx))
 801             {
 802                 res = VSHUFFLE(a, a, idx);
 803             }
 804             else
 805             {
 806                 res = VUNDEF_F();
 807                 for (uint32_t l = 0; l < JM()->mVWidth; ++l)
 808                 {
 809                     Value* pIndex = VEXTRACT(idx, C(l));
 810                     Value* pVal = VEXTRACT(a, pIndex);
 811                     res = VINSERT(res, pVal, C(l));
 812                 }
 813             }
 814         }
 815
 816         return res;
 817     }
 818
 819     //////////////////////////////////////////////////////////////////////////
 820     /// @brief Generate a VCVTPH2PS operation (float16->float32 conversion)
 821     /// in LLVM IR.  If not supported on the underlying platform, emulate it
 822     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
 823     Value *Builder::CVTPH2PS(Value* a)
 824     {
 825         if (JM()->mArch.F16C())
 826         {
 827             return VCVTPH2PS(a);
 828         }
 829         else
 830         {
 831             FunctionType* pFuncTy = FunctionType::get(mFP32Ty, mInt16Ty);
 832             Function* pCvtPh2Ps = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("ConvertSmallFloatTo32", pFuncTy));
 833
 834             if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertSmallFloatTo32") == nullptr)
 835             {
 836                 sys::DynamicLibrary::AddSymbol("ConvertSmallFloatTo32", (void *)&ConvertSmallFloatTo32);
 837             }
 838
 839             Value* pResult = UndefValue::get(mSimdFP32Ty);
 840             for (uint32_t i = 0; i < mVWidth; ++i)
 841             {
 842                 Value* pSrc = VEXTRACT(a, C(i));
 843                 Value* pConv = CALL(pCvtPh2Ps, std::initializer_list<Value*>{pSrc});
 844                 pResult = VINSERT(pResult, pConv, C(i));
 845             }
 846
 847             return pResult;
 848         }
 849     }
 850
 851     //////////////////////////////////////////////////////////////////////////
 852     /// @brief Generate a VCVTPS2PH operation (float32->float16 conversion)
 853     /// in LLVM IR.  If not supported on the underlying platform, emulate it
 854     /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
 855     Value *Builder::CVTPS2PH(Value* a, Value* rounding)
 856     {
 857         if (JM()->mArch.F16C())
 858         {
 859             return VCVTPS2PH(a, rounding);
 860         }
 861         else
 862         {
 863             // call scalar C function for now
 864             FunctionType* pFuncTy = FunctionType::get(mInt16Ty, mFP32Ty);
 865             Function* pCvtPs2Ph = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("Convert32To16Float", pFuncTy));
 866
 867             if (sys::DynamicLibrary::SearchForAddressOfSymbol("Convert32To16Float") == nullptr)
 868             {
 869                 sys::DynamicLibrary::AddSymbol("Convert32To16Float", (void *)&Convert32To16Float);
 870             }
 871
 872             Value* pResult = UndefValue::get(mSimdInt16Ty);
 873             for (uint32_t i = 0; i < mVWidth; ++i)
 874             {
 875                 Value* pSrc = VEXTRACT(a, C(i));
 876                 Value* pConv = CALL(pCvtPs2Ph, std::initializer_list<Value*>{pSrc});
 877                 pResult = VINSERT(pResult, pConv, C(i));
 878             }
 879
 880             return pResult;
 881         }
 882     }
 883
 884     Value *Builder::PMAXSD(Value* a, Value* b)
 885     {
 886         // llvm-3.9 removed the pmax intrinsics
 887     #if HAVE_LLVM >= 0x309
 888         Value* cmp = ICMP_SGT(a, b);
 889         return SELECT(cmp, a, b);
 890     #else
 891         if (JM()->mArch.AVX2())
 892         {
 893             Function* pmaxsd = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_avx2_pmaxs_d);
 894             return CALL(pmaxsd, {a, b});
 895         }
 896         else
 897         {
 898             // use 4-wide sse max intrinsic on lower/upper halves of 8-wide sources
 899             Function* pmaxsd = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_sse41_pmaxsd);
 900
 901             // low 128
 902             Value* aLo = VEXTRACTI128(a, C((uint8_t)0));
 903             Value* bLo = VEXTRACTI128(b, C((uint8_t)0));
 904             Value* resLo = CALL(pmaxsd, {aLo, bLo});
 905
 906             // high 128
 907             Value* aHi = VEXTRACTI128(a, C((uint8_t)1));
 908             Value* bHi = VEXTRACTI128(b, C((uint8_t)1));
 909             Value* resHi = CALL(pmaxsd, {aHi, bHi});
 910
 911             // combine
 912             Value* result = VINSERTI128(VUNDEF_I(), resLo, C((uint8_t)0));
 913             result = VINSERTI128(result, resHi, C((uint8_t)1));
 914
 915             return result;
 916         }
 917     #endif
 918     }
 919
 920     Value *Builder::PMINSD(Value* a, Value* b)
 921     {
 922         // llvm-3.9 removed the pmin intrinsics
 923     #if HAVE_LLVM >= 0x309
 924         Value* cmp = ICMP_SLT(a, b);
 925         return SELECT(cmp, a, b);
 926     #else
 927         if (JM()->mArch.AVX2())
 928         {
 929             Function* pminsd = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_avx2_pmins_d);
 930             return CALL(pminsd, {a, b});
 931         }
 932         else
 933         {
 934             // use 4-wide sse max intrinsic on lower/upper halves of 8-wide sources
 935             Function* pminsd = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_sse41_pminsd);
 936
 937             // low 128
 938             Value* aLo = VEXTRACTI128(a, C((uint8_t)0));
 939             Value* bLo = VEXTRACTI128(b, C((uint8_t)0));
 940             Value* resLo = CALL(pminsd, {aLo, bLo});
 941
 942             // high 128
 943             Value* aHi = VEXTRACTI128(a, C((uint8_t)1));
 944             Value* bHi = VEXTRACTI128(b, C((uint8_t)1));
 945             Value* resHi = CALL(pminsd, {aHi, bHi});
 946
 947             // combine
 948             Value* result = VINSERTI128(VUNDEF_I(), resLo, C((uint8_t)0));
 949             result = VINSERTI128(result, resHi, C((uint8_t)1));
 950
 951             return result;
 952         }
 953     #endif
 954     }
 955
 956     void Builder::Gather4(const SWR_FORMAT format, Value* pSrcBase, Value* byteOffsets,
 957                           Value* mask, Value* vGatherComponents[], bool bPackedOutput)
 958     {
 959         const SWR_FORMAT_INFO &info = GetFormatInfo(format);
 960         if(info.type[0] == SWR_TYPE_FLOAT && info.bpc[0] == 32)
 961         {
 962             // ensure our mask is the correct type
 963             mask = BITCAST(mask, mSimdFP32Ty);
 964             GATHER4PS(info, pSrcBase, byteOffsets, mask, vGatherComponents, bPackedOutput);
 965         }
 966         else
 967         {
 968             // ensure our mask is the correct type
 969             mask = BITCAST(mask, mSimdInt32Ty);
 970             GATHER4DD(info, pSrcBase, byteOffsets, mask, vGatherComponents, bPackedOutput);
 971         }
 972     }
 973
 974     void Builder::GATHER4PS(const SWR_FORMAT_INFO &info, Value* pSrcBase, Value* byteOffsets,
 975                             Value* mask, Value* vGatherComponents[], bool bPackedOutput)
 976     {
 977         switch(info.bpp / info.numComps)
 978         {
 979             case 16:
 980             {
 981                     Value* vGatherResult[2];
 982                     Value *vMask;
 983
 984                     // TODO: vGatherMaskedVal
 985                     Value* vGatherMaskedVal = VIMMED1((float)0);
 986
 987                     // always have at least one component out of x or y to fetch
 988
 989                     // save mask as it is zero'd out after each gather
 990                     vMask = mask;
 991
 992                     vGatherResult[0] = GATHERPS(vGatherMaskedVal, pSrcBase, byteOffsets, vMask, C((char)1));
 993                     // e.g. result of first 8x32bit integer gather for 16bit components
 994                     // 256i - 0    1    2    3    4    5    6    7
 995                     //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
 996                     //
 997
 998                     // if we have at least one component out of x or y to fetch
 999                     if(info.numComps > 2)
1000                     {
1001                         // offset base to the next components(zw) in the vertex to gather
1002                         pSrcBase = GEP(pSrcBase, C((char)4));
1003                         vMask = mask;
1004
1005                         vGatherResult[1] =  GATHERPS(vGatherMaskedVal, pSrcBase, byteOffsets, vMask, C((char)1));
1006                         // e.g. result of second 8x32bit integer gather for 16bit components
1007                         // 256i - 0    1    2    3    4    5    6    7
1008                         //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1009                         //
1010                     }
1011                     else
1012                     {
1013                         vGatherResult[1] =  vGatherMaskedVal;
1014                     }
1015
1016                     // Shuffle gathered components into place, each row is a component
1017                     Shuffle16bpcGather4(info, vGatherResult, vGatherComponents, bPackedOutput);
1018             }
1019                 break;
1020             case 32:
1021             {
1022                 // apply defaults
1023                 for (uint32_t i = 0; i < 4; ++i)
1024                 {
1025                     vGatherComponents[i] = VIMMED1(*(float*)&info.defaults[i]);
1026                 }
1027
1028                 for(uint32_t i = 0; i < info.numComps; i++)
1029                 {
1030                     uint32_t swizzleIndex = info.swizzle[i];
1031
1032                     // save mask as it is zero'd out after each gather
1033                     Value *vMask = mask;
1034
1035                     // Gather a SIMD of components
1036                     vGatherComponents[swizzleIndex] = GATHERPS(vGatherComponents[swizzleIndex], pSrcBase, byteOffsets, vMask, C((char)1));
1037
1038                     // offset base to the next component to gather
1039                     pSrcBase = GEP(pSrcBase, C((char)4));
1040                 }
1041             }
1042                 break;
1043             default:
1044                 SWR_ASSERT(0, "Invalid float format");
1045                 break;
1046         }
1047     }
1048
1049     void Builder::GATHER4DD(const SWR_FORMAT_INFO &info, Value* pSrcBase, Value* byteOffsets,
1050                             Value* mask, Value* vGatherComponents[], bool bPackedOutput)
1051     {
1052         switch (info.bpp / info.numComps)
1053         {
1054             case 8:
1055             {
1056                 Value* vGatherMaskedVal = VIMMED1((int32_t)0);
1057                 Value* vGatherResult = GATHERDD(vGatherMaskedVal, pSrcBase, byteOffsets, mask, C((char)1));
1058                 // e.g. result of an 8x32bit integer gather for 8bit components
1059                 // 256i - 0    1    2    3    4    5    6    7
1060                 //        xyzw xyzw xyzw xyzw xyzw xyzw xyzw xyzw
1061
1062                 Shuffle8bpcGather4(info, vGatherResult, vGatherComponents, bPackedOutput);
1063             }
1064                 break;
1065             case 16:
1066             {
1067                 Value* vGatherResult[2];
1068                 Value *vMask;
1069
1070                 // TODO: vGatherMaskedVal
1071                 Value* vGatherMaskedVal = VIMMED1((int32_t)0);
1072
1073                 // always have at least one component out of x or y to fetch
1074
1075                 // save mask as it is zero'd out after each gather
1076                 vMask = mask;
1077
1078                 vGatherResult[0] = GATHERDD(vGatherMaskedVal, pSrcBase, byteOffsets, vMask, C((char)1));
1079                 // e.g. result of first 8x32bit integer gather for 16bit components
1080                 // 256i - 0    1    2    3    4    5    6    7
1081                 //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1082                 //
1083
1084                 // if we have at least one component out of x or y to fetch
1085                 if(info.numComps > 2)
1086                 {
1087                     // offset base to the next components(zw) in the vertex to gather
1088                     pSrcBase = GEP(pSrcBase, C((char)4));
1089                     vMask = mask;
1090
1091                     vGatherResult[1] = GATHERDD(vGatherMaskedVal, pSrcBase, byteOffsets, vMask, C((char)1));
1092                     // e.g. result of second 8x32bit integer gather for 16bit components
1093                     // 256i - 0    1    2    3    4    5    6    7
1094                     //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1095                     //
1096                 }
1097                 else
1098                 {
1099                     vGatherResult[1] = vGatherMaskedVal;
1100                 }
1101
1102                 // Shuffle gathered components into place, each row is a component
1103                 Shuffle16bpcGather4(info, vGatherResult, vGatherComponents, bPackedOutput);
1104
1105             }
1106                 break;
1107             case 32:
1108             {
1109                 // apply defaults
1110                 for (uint32_t i = 0; i < 4; ++i)
1111                 {
1112                     vGatherComponents[i] = VIMMED1((int)info.defaults[i]);
1113                 }
1114
1115                 for(uint32_t i = 0; i < info.numComps; i++)
1116                 {
1117                     uint32_t swizzleIndex = info.swizzle[i];
1118
1119                     // save mask as it is zero'd out after each gather
1120                     Value *vMask = mask;
1121
1122                     // Gather a SIMD of components
1123                     vGatherComponents[swizzleIndex] = GATHERDD(vGatherComponents[swizzleIndex], pSrcBase, byteOffsets, vMask, C((char)1));
1124
1125                     // offset base to the next component to gather
1126                     pSrcBase = GEP(pSrcBase, C((char)4));
1127                 }
1128             }
1129                 break;
1130             default:
1131                 SWR_ASSERT(0, "unsupported format");
1132             break;
1133         }
1134     }
1135
1136     void Builder::Shuffle16bpcGather4(const SWR_FORMAT_INFO &info, Value* vGatherInput[2], Value* vGatherOutput[4], bool bPackedOutput)
1137     {
1138         // cast types
1139         Type* vGatherTy = VectorType::get(IntegerType::getInt32Ty(JM()->mContext), mVWidth);
1140         Type* v32x8Ty = VectorType::get(mInt8Ty, mVWidth * 4); // vwidth is units of 32 bits
1141
1142         // input could either be float or int vector; do shuffle work in int
1143         vGatherInput[0] = BITCAST(vGatherInput[0], mSimdInt32Ty);
1144         vGatherInput[1] = BITCAST(vGatherInput[1], mSimdInt32Ty);
1145
1146         if(bPackedOutput)
1147         {
1148             Type* v128bitTy = VectorType::get(IntegerType::getIntNTy(JM()->mContext, 128), mVWidth / 4); // vwidth is units of 32 bits
1149
1150             // shuffle mask
1151             Value* vConstMask = C<char>({0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15,
1152                                          0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15});
1153             Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherInput[0], v32x8Ty), vConstMask), vGatherTy);
1154             // after pshufb: group components together in each 128bit lane
1155             // 256i - 0    1    2    3    4    5    6    7
1156             //        xxxx xxxx yyyy yyyy xxxx xxxx yyyy yyyy
1157
1158             Value* vi128XY = BITCAST(PERMD(vShufResult, C<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy);
1159             // after PERMD: move and pack xy components into each 128bit lane
1160             // 256i - 0    1    2    3    4    5    6    7
1161             //        xxxx xxxx xxxx xxxx yyyy yyyy yyyy yyyy
1162
1163             // do the same for zw components
1164             Value* vi128ZW = nullptr;
1165             if(info.numComps > 2)
1166             {
1167                 Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherInput[1], v32x8Ty), vConstMask), vGatherTy);
1168                 vi128ZW = BITCAST(PERMD(vShufResult, C<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy);
1169             }
1170
1171             for(uint32_t i = 0; i < 4; i++)
1172             {
1173                 uint32_t swizzleIndex = info.swizzle[i];
1174                 // todo: fixed for packed
1175                 Value* vGatherMaskedVal = VIMMED1((int32_t)(info.defaults[i]));
1176                 if(i >= info.numComps)
1177                 {
1178                     // set the default component val
1179                     vGatherOutput[swizzleIndex] = vGatherMaskedVal;
1180                     continue;
1181                 }
1182
1183                 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1184                 uint32_t lane = ((i == 0) || (i == 2)) ? 0 : 1;
1185                 // if x or y, use vi128XY permute result, else use vi128ZW
1186                 Value* selectedPermute = (i < 2) ? vi128XY : vi128ZW;
1187
1188                 // extract packed component 128 bit lanes
1189                 vGatherOutput[swizzleIndex] = VEXTRACT(selectedPermute, C(lane));
1190             }
1191
1192         }
1193         else
1194         {
1195             // pshufb masks for each component
1196             Value* vConstMask[2];
1197             // x/z shuffle mask
1198             vConstMask[0] = C<char>({0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1,
1199                                      0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1, });
1200
1201             // y/w shuffle mask
1202             vConstMask[1] = C<char>({2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1,
1203                                      2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1});
1204
1205
1206             // shuffle enabled components into lower word of each 32bit lane, 0 extending to 32 bits
1207             // apply defaults
1208             for (uint32_t i = 0; i < 4; ++i)
1209             {
1210                 vGatherOutput[i] = VIMMED1((int32_t)info.defaults[i]);
1211             }
1212
1213             for(uint32_t i = 0; i < info.numComps; i++)
1214             {
1215                 uint32_t swizzleIndex = info.swizzle[i];
1216
1217                 // select correct constMask for x/z or y/w pshufb
1218                 uint32_t selectedMask = ((i == 0) || (i == 2)) ? 0 : 1;
1219                 // if x or y, use vi128XY permute result, else use vi128ZW
1220                 uint32_t selectedGather = (i < 2) ? 0 : 1;
1221
1222                 vGatherOutput[swizzleIndex] = BITCAST(PSHUFB(BITCAST(vGatherInput[selectedGather], v32x8Ty), vConstMask[selectedMask]), vGatherTy);
1223                 // after pshufb mask for x channel; z uses the same shuffle from the second gather
1224                 // 256i - 0    1    2    3    4    5    6    7
1225                 //        xx00 xx00 xx00 xx00 xx00 xx00 xx00 xx00
1226             }
1227         }
1228     }
1229
1230     void Builder::Shuffle8bpcGather4(const SWR_FORMAT_INFO &info, Value* vGatherInput, Value* vGatherOutput[], bool bPackedOutput)
1231     {
1232         // cast types
1233         Type* vGatherTy = VectorType::get(IntegerType::getInt32Ty(JM()->mContext), mVWidth);
1234         Type* v32x8Ty =  VectorType::get(mInt8Ty, mVWidth * 4 ); // vwidth is units of 32 bits
1235
1236         if(bPackedOutput)
1237         {
1238             Type* v128Ty = VectorType::get(IntegerType::getIntNTy(JM()->mContext, 128), mVWidth / 4); // vwidth is units of 32 bits
1239             // shuffle mask
1240             Value* vConstMask = C<char>({0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
1241                                          0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15});
1242             Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherInput, v32x8Ty), vConstMask), vGatherTy);
1243             // after pshufb: group components together in each 128bit lane
1244             // 256i - 0    1    2    3    4    5    6    7
1245             //        xxxx yyyy zzzz wwww xxxx yyyy zzzz wwww
1246
1247             Value* vi128XY = BITCAST(PERMD(vShufResult, C<int32_t>({0, 4, 0, 0, 1, 5, 0, 0})), v128Ty);
1248             // after PERMD: move and pack xy and zw components in low 64 bits of each 128bit lane
1249             // 256i - 0    1    2    3    4    5    6    7
1250             //        xxxx xxxx dcdc dcdc yyyy yyyy dcdc dcdc (dc - don't care)
1251
1252             // do the same for zw components
1253             Value* vi128ZW = nullptr;
1254             if(info.numComps > 2)
1255             {
1256                 vi128ZW = BITCAST(PERMD(vShufResult, C<int32_t>({2, 6, 0, 0, 3, 7, 0, 0})), v128Ty);
1257             }
1258
1259             // sign extend all enabled components. If we have a fill vVertexElements, output to current simdvertex
1260             for(uint32_t i = 0; i < 4; i++)
1261             {
1262                 uint32_t swizzleIndex = info.swizzle[i];
1263                 // todo: fix for packed
1264                 Value* vGatherMaskedVal = VIMMED1((int32_t)(info.defaults[i]));
1265                 if(i >= info.numComps)
1266                 {
1267                     // set the default component val
1268                     vGatherOutput[swizzleIndex] = vGatherMaskedVal;
1269                     continue;
1270                 }
1271
1272                 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1273                 uint32_t lane = ((i == 0) || (i == 2)) ? 0 : 1;
1274                 // if x or y, use vi128XY permute result, else use vi128ZW
1275                 Value* selectedPermute = (i < 2) ? vi128XY : vi128ZW;
1276
1277                 // sign extend
1278                 vGatherOutput[swizzleIndex] = VEXTRACT(selectedPermute, C(lane));
1279             }
1280         }
1281         // else zero extend
1282         else{
1283             // shuffle enabled components into lower byte of each 32bit lane, 0 extending to 32 bits
1284             // apply defaults
1285             for (uint32_t i = 0; i < 4; ++i)
1286             {
1287                 vGatherOutput[i] = VIMMED1((int32_t)info.defaults[i]);
1288             }
1289
1290             for(uint32_t i = 0; i < info.numComps; i++){
1291                 uint32_t swizzleIndex = info.swizzle[i];
1292
1293                 // pshufb masks for each component
1294                 Value* vConstMask;
1295                 switch(i)
1296                 {
1297                     case 0:
1298                         // x shuffle mask
1299                         vConstMask = C<char>({0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1,
1300                                               0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1});
1301                         break;
1302                     case 1:
1303                         // y shuffle mask
1304                         vConstMask = C<char>({1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1,
1305                                               1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1});
1306                         break;
1307                     case 2:
1308                         // z shuffle mask
1309                         vConstMask = C<char>({2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1,
1310                                               2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1});
1311                         break;
1312                     case 3:
1313                         // w shuffle mask
1314                         vConstMask = C<char>({3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1,
1315                                               3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1});
1316                         break;
1317                     default:
1318                         vConstMask = nullptr;
1319                         break;
1320                 }
1321
1322                     vGatherOutput[swizzleIndex] = BITCAST(PSHUFB(BITCAST(vGatherInput, v32x8Ty), vConstMask), vGatherTy);
1323                     // after pshufb for x channel
1324                     // 256i - 0    1    2    3    4    5    6    7
1325                     //        x000 x000 x000 x000 x000 x000 x000 x000
1326             }
1327         }
1328     }
1329
1330     // Helper function to create alloca in entry block of function
1331     Value* Builder::CreateEntryAlloca(Function* pFunc, Type* pType)
1332     {
1333         auto saveIP = IRB()->saveIP();
1334         IRB()->SetInsertPoint(&pFunc->getEntryBlock(),
1335                               pFunc->getEntryBlock().begin());
1336         Value* pAlloca = ALLOCA(pType);
1337         IRB()->restoreIP(saveIP);
1338         return pAlloca;
1339     }
1340
1341     //////////////////////////////////////////////////////////////////////////
1342     /// @brief emulates a scatter operation.
1343     /// @param pDst - pointer to destination
1344     /// @param vSrc - vector of src data to scatter
1345     /// @param vOffsets - vector of byte offsets from pDst
1346     /// @param vMask - mask of valid lanes
1347     void Builder::SCATTERPS(Value* pDst, Value* vSrc, Value* vOffsets, Value* vMask)
1348     {
1349         /* Scatter algorithm
1350
1351            while(Index = BitScanForward(mask))
1352                 srcElem = srcVector[Index]
1353                 offsetElem = offsetVector[Index]
1354                 *(pDst + offsetElem) = srcElem
1355                 Update mask (&= ~(1<<Index)
1356
1357         */
1358
1359         BasicBlock* pCurBB = IRB()->GetInsertBlock();
1360         Function* pFunc = pCurBB->getParent();
1361         Type* pSrcTy = vSrc->getType()->getVectorElementType();
1362
1363         // Store vectors on stack
1364         if (pScatterStackSrc == nullptr)
1365         {
1366             // Save off stack allocations and reuse per scatter. Significantly reduces stack
1367             // requirements for shaders with a lot of scatters.
1368             pScatterStackSrc = CreateEntryAlloca(pFunc, mSimdInt64Ty);
1369             pScatterStackOffsets = CreateEntryAlloca(pFunc, mSimdInt32Ty);
1370         }
1371
1372         Value* pSrcArrayPtr = BITCAST(pScatterStackSrc, PointerType::get(vSrc->getType(), 0));
1373         Value* pOffsetsArrayPtr = pScatterStackOffsets;
1374         STORE(vSrc, pSrcArrayPtr);
1375         STORE(vOffsets, pOffsetsArrayPtr);
1376
1377         // Cast to pointers for random access
1378         pSrcArrayPtr = POINTER_CAST(pSrcArrayPtr, PointerType::get(pSrcTy, 0));
1379         pOffsetsArrayPtr = POINTER_CAST(pOffsetsArrayPtr, PointerType::get(mInt32Ty, 0));
1380
1381         Value* pMask = VMOVMSKPS(BITCAST(vMask, mSimdFP32Ty));
1382
1383         // Get cttz function
1384         Function* pfnCttz = Intrinsic::getDeclaration(mpJitMgr->mpCurrentModule, Intrinsic::cttz, { mInt32Ty });
1385
1386         // Setup loop basic block
1387         BasicBlock* pLoop = BasicBlock::Create(mpJitMgr->mContext, "Scatter Loop", pFunc);
1388
1389         // compute first set bit
1390         Value* pIndex = CALL(pfnCttz, { pMask, C(false) });
1391
1392         Value* pIsUndef = ICMP_EQ(pIndex, C(32));
1393
1394         // Split current block
1395         BasicBlock* pPostLoop = pCurBB->splitBasicBlock(cast<Instruction>(pIsUndef)->getNextNode());
1396
1397         // Remove unconditional jump created by splitBasicBlock
1398         pCurBB->getTerminator()->eraseFromParent();
1399
1400         // Add terminator to end of original block
1401         IRB()->SetInsertPoint(pCurBB);
1402
1403         // Add conditional branch
1404         COND_BR(pIsUndef, pPostLoop, pLoop);
1405
1406         // Add loop basic block contents
1407         IRB()->SetInsertPoint(pLoop);
1408         PHINode* pIndexPhi = PHI(mInt32Ty, 2);
1409         PHINode* pMaskPhi = PHI(mInt32Ty, 2);
1410
1411         pIndexPhi->addIncoming(pIndex, pCurBB);
1412         pMaskPhi->addIncoming(pMask, pCurBB);
1413
1414         // Extract elements for this index
1415         Value* pSrcElem = LOADV(pSrcArrayPtr, { pIndexPhi });
1416         Value* pOffsetElem = LOADV(pOffsetsArrayPtr, { pIndexPhi });
1417
1418         // GEP to this offset in dst
1419         Value* pCurDst = GEP(pDst, pOffsetElem);
1420         pCurDst = POINTER_CAST(pCurDst, PointerType::get(pSrcTy, 0));
1421         STORE(pSrcElem, pCurDst);
1422
1423         // Update the mask
1424         Value* pNewMask = AND(pMaskPhi, NOT(SHL(C(1), pIndexPhi)));
1425
1426         // Terminator
1427         Value* pNewIndex = CALL(pfnCttz, { pNewMask, C(false) });
1428
1429         pIsUndef = ICMP_EQ(pNewIndex, C(32));
1430         COND_BR(pIsUndef, pPostLoop, pLoop);
1431
1432         // Update phi edges
1433         pIndexPhi->addIncoming(pNewIndex, pLoop);
1434         pMaskPhi->addIncoming(pNewMask, pLoop);
1435
1436         // Move builder to beginning of post loop
1437         IRB()->SetInsertPoint(pPostLoop, pPostLoop->begin());
1438     }
1439
1440     Value* Builder::VABSPS(Value* a)
1441     {
1442         Value* asInt = BITCAST(a, mSimdInt32Ty);
1443         Value* result = BITCAST(AND(asInt, VIMMED1(0x7fffffff)), mSimdFP32Ty);
1444         return result;
1445     }
1446
1447     Value *Builder::ICLAMP(Value* src, Value* low, Value* high)
1448     {
1449         Value *lowCmp = ICMP_SLT(src, low);
1450         Value *ret = SELECT(lowCmp, low, src);
1451
1452         Value *highCmp = ICMP_SGT(ret, high);
1453         ret = SELECT(highCmp, high, ret);
1454
1455         return ret;
1456     }
1457
1458     Value *Builder::FCLAMP(Value* src, Value* low, Value* high)
1459     {
1460         Value *lowCmp = FCMP_OLT(src, low);
1461         Value *ret = SELECT(lowCmp, low, src);
1462
1463         Value *highCmp = FCMP_OGT(ret, high);
1464         ret = SELECT(highCmp, high, ret);
1465
1466         return ret;
1467     }
1468
1469     Value *Builder::FCLAMP(Value* src, float low, float high)
1470     {
1471         Value* result = VMAXPS(src, VIMMED1(low));
1472         result = VMINPS(result, VIMMED1(high));
1473
1474         return result;
1475     }
1476
1477     //////////////////////////////////////////////////////////////////////////
1478     /// @brief save/restore stack, providing ability to push/pop the stack and
1479     ///        reduce overall stack requirements for temporary stack use
1480     Value* Builder::STACKSAVE()
1481     {
1482         Function* pfnStackSave = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::stacksave);
1483     #if HAVE_LLVM == 0x306
1484         return CALL(pfnStackSave);
1485     #else
1486         return CALLA(pfnStackSave);
1487     #endif
1488     }
1489
1490     void Builder::STACKRESTORE(Value* pSaved)
1491     {
1492         Function* pfnStackRestore = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::stackrestore);
1493         CALL(pfnStackRestore, std::initializer_list<Value*>{pSaved});
1494     }
1495
1496     Value *Builder::FMADDPS(Value* a, Value* b, Value* c)
1497     {
1498         Value* vOut;
1499         // use FMADs if available
1500         if(JM()->mArch.AVX2())
1501         {
1502             vOut = VFMADDPS(a, b, c);
1503         }
1504         else
1505         {
1506             vOut = FADD(FMUL(a, b), c);
1507         }
1508         return vOut;
1509     }
1510
1511     Value* Builder::POPCNT(Value* a)
1512     {
1513         Function* pCtPop = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::ctpop, { a->getType() });
1514         return CALL(pCtPop, std::initializer_list<Value*>{a});
1515     }
1516
1517     //////////////////////////////////////////////////////////////////////////
1518     /// @brief C functions called by LLVM IR
1519     //////////////////////////////////////////////////////////////////////////
1520
1521     //////////////////////////////////////////////////////////////////////////
1522     /// @brief called in JIT code, inserted by PRINT
1523     /// output to both stdout and visual studio debug console
1524     void __cdecl CallPrint(const char* fmt, ...)
1525     {
1526         va_list args;
1527         va_start(args, fmt);
1528         vprintf(fmt, args);
1529
1530     #if defined( _WIN32 )
1531         char strBuf[1024];
1532         vsnprintf_s(strBuf, _TRUNCATE, fmt, args);
1533         OutputDebugString(strBuf);
1534     #endif
1535
1536         va_end(args);
1537     }
1538
1539     Value *Builder::VEXTRACTI128(Value* a, Constant* imm8)
1540     {
1541     #if HAVE_LLVM == 0x306
1542         Function *func =
1543             Intrinsic::getDeclaration(JM()->mpCurrentModule,
1544                                       Intrinsic::x86_avx_vextractf128_si_256);
1545         return CALL(func, {a, imm8});
1546     #else
1547         bool flag = !imm8->isZeroValue();
1548         SmallVector<Constant*,8> idx;
1549         for (unsigned i = 0; i < mVWidth / 2; i++) {
1550             idx.push_back(C(flag ? i + mVWidth / 2 : i));
1551         }
1552         return VSHUFFLE(a, VUNDEF_I(), ConstantVector::get(idx));
1553     #endif
1554     }
1555
1556     Value *Builder::VINSERTI128(Value* a, Value* b, Constant* imm8)
1557     {
1558     #if HAVE_LLVM == 0x306
1559         Function *func =
1560             Intrinsic::getDeclaration(JM()->mpCurrentModule,
1561                                       Intrinsic::x86_avx_vinsertf128_si_256);
1562         return CALL(func, {a, b, imm8});
1563     #else
1564         bool flag = !imm8->isZeroValue();
1565         SmallVector<Constant*,8> idx;
1566         for (unsigned i = 0; i < mVWidth; i++) {
1567             idx.push_back(C(i));
1568         }
1569         Value *inter = VSHUFFLE(b, VUNDEF_I(), ConstantVector::get(idx));
1570
1571         SmallVector<Constant*,8> idx2;
1572         for (unsigned i = 0; i < mVWidth / 2; i++) {
1573             idx2.push_back(C(flag ? i : i + mVWidth));
1574         }
1575         for (unsigned i = mVWidth / 2; i < mVWidth; i++) {
1576             idx2.push_back(C(flag ? i + mVWidth / 2 : i));
1577         }
1578         return VSHUFFLE(a, inter, ConstantVector::get(idx2));
1579     #endif
1580     }
1581
1582     // rdtsc buckets macros
1583     void Builder::RDTSC_START(Value* pBucketMgr, Value* pId)
1584     {
1585         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1586         // buckets framework when single threaded
1587         if (KNOB_SINGLE_THREADED)
1588         {
1589             std::vector<Type*> args{
1590                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
1591                 mInt32Ty                        // id
1592             };
1593
1594             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
1595             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StartBucket", pFuncTy));
1596             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StartBucket") == nullptr)
1597             {
1598                 sys::DynamicLibrary::AddSymbol("BucketManager_StartBucket", (void*)&BucketManager_StartBucket);
1599             }
1600
1601             CALL(pFunc, { pBucketMgr, pId });
1602         }
1603     }
1604
1605     void Builder::RDTSC_STOP(Value* pBucketMgr, Value* pId)
1606     {
1607         // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1608         // buckets framework when single threaded
1609         if (KNOB_SINGLE_THREADED)
1610         {
1611             std::vector<Type*> args{
1612                 PointerType::get(mInt32Ty, 0),   // pBucketMgr
1613                 mInt32Ty                        // id
1614             };
1615
1616             FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
1617             Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StopBucket", pFuncTy));
1618             if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StopBucket") == nullptr)
1619             {
1620                 sys::DynamicLibrary::AddSymbol("BucketManager_StopBucket", (void*)&BucketManager_StopBucket);
1621             }
1622
1623             CALL(pFunc, { pBucketMgr, pId });
1624         }
1625     }
1626
1627 }