brintos / llvm-project-archived public Read only

1//===-- ConstantFolding.cpp - Fold instructions into constants ------------===//2//3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.4// See https://llvm.org/LICENSE.txt for license information.5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception6//7//===----------------------------------------------------------------------===//8//9// This file defines routines for folding instructions into constants.10//11// Also, to supplement the basic IR ConstantExpr simplifications,12// this file defines some additional folding routines that can make use of13// DataLayout information. These functions cannot go in IR due to library14// dependency issues.15//16//===----------------------------------------------------------------------===//17 18#include "llvm/Analysis/ConstantFolding.h"19#include "llvm/ADT/APFloat.h"20#include "llvm/ADT/APInt.h"21#include "llvm/ADT/APSInt.h"22#include "llvm/ADT/ArrayRef.h"23#include "llvm/ADT/DenseMap.h"24#include "llvm/ADT/STLExtras.h"25#include "llvm/ADT/SmallVector.h"26#include "llvm/ADT/StringRef.h"27#include "llvm/Analysis/TargetFolder.h"28#include "llvm/Analysis/TargetLibraryInfo.h"29#include "llvm/Analysis/ValueTracking.h"30#include "llvm/Analysis/VectorUtils.h"31#include "llvm/Config/config.h"32#include "llvm/IR/Constant.h"33#include "llvm/IR/ConstantFold.h"34#include "llvm/IR/Constants.h"35#include "llvm/IR/DataLayout.h"36#include "llvm/IR/DerivedTypes.h"37#include "llvm/IR/Function.h"38#include "llvm/IR/GlobalValue.h"39#include "llvm/IR/GlobalVariable.h"40#include "llvm/IR/InstrTypes.h"41#include "llvm/IR/Instruction.h"42#include "llvm/IR/Instructions.h"43#include "llvm/IR/IntrinsicInst.h"44#include "llvm/IR/Intrinsics.h"45#include "llvm/IR/IntrinsicsAArch64.h"46#include "llvm/IR/IntrinsicsAMDGPU.h"47#include "llvm/IR/IntrinsicsARM.h"48#include "llvm/IR/IntrinsicsNVPTX.h"49#include "llvm/IR/IntrinsicsWebAssembly.h"50#include "llvm/IR/IntrinsicsX86.h"51#include "llvm/IR/NVVMIntrinsicUtils.h"52#include "llvm/IR/Operator.h"53#include "llvm/IR/Type.h"54#include "llvm/IR/Value.h"55#include "llvm/Support/Casting.h"56#include "llvm/Support/ErrorHandling.h"57#include "llvm/Support/KnownBits.h"58#include "llvm/Support/MathExtras.h"59#include <cassert>60#include <cerrno>61#include <cfenv>62#include <cmath>63#include <cstdint>64 65using namespace llvm;66 67static cl::opt<bool> DisableFPCallFolding(68    "disable-fp-call-folding",69    cl::desc("Disable constant-folding of FP intrinsics and libcalls."),70    cl::init(false), cl::Hidden);71 72namespace {73 74//===----------------------------------------------------------------------===//75// Constant Folding internal helper functions76//===----------------------------------------------------------------------===//77 78static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,79                                        Constant *C, Type *SrcEltTy,80                                        unsigned NumSrcElts,81                                        const DataLayout &DL) {82  // Now that we know that the input value is a vector of integers, just shift83  // and insert them into our result.84  unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy);85  for (unsigned i = 0; i != NumSrcElts; ++i) {86    Constant *Element;87    if (DL.isLittleEndian())88      Element = C->getAggregateElement(NumSrcElts - i - 1);89    else90      Element = C->getAggregateElement(i);91 92    if (isa_and_nonnull<UndefValue>(Element)) {93      Result <<= BitShift;94      continue;95    }96 97    auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);98    if (!ElementCI)99      return ConstantExpr::getBitCast(C, DestTy);100 101    Result <<= BitShift;102    Result |= ElementCI->getValue().zext(Result.getBitWidth());103  }104 105  return nullptr;106}107 108/// Constant fold bitcast, symbolically evaluating it with DataLayout.109/// This always returns a non-null constant, but it may be a110/// ConstantExpr if unfoldable.111Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {112  assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) &&113         "Invalid constantexpr bitcast!");114 115  // Catch the obvious splat cases.116  if (Constant *Res = ConstantFoldLoadFromUniformValue(C, DestTy, DL))117    return Res;118 119  if (auto *VTy = dyn_cast<VectorType>(C->getType())) {120    // Handle a vector->scalar integer/fp cast.121    if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) {122      unsigned NumSrcElts = cast<FixedVectorType>(VTy)->getNumElements();123      Type *SrcEltTy = VTy->getElementType();124 125      // If the vector is a vector of floating point, convert it to vector of int126      // to simplify things.127      if (SrcEltTy->isFloatingPointTy()) {128        unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();129        auto *SrcIVTy = FixedVectorType::get(130            IntegerType::get(C->getContext(), FPWidth), NumSrcElts);131        // Ask IR to do the conversion now that #elts line up.132        C = ConstantExpr::getBitCast(C, SrcIVTy);133      }134 135      APInt Result(DL.getTypeSizeInBits(DestTy), 0);136      if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,137                                                SrcEltTy, NumSrcElts, DL))138        return CE;139 140      if (isa<IntegerType>(DestTy))141        return ConstantInt::get(DestTy, Result);142 143      APFloat FP(DestTy->getFltSemantics(), Result);144      return ConstantFP::get(DestTy->getContext(), FP);145    }146  }147 148  // The code below only handles casts to vectors currently.149  auto *DestVTy = dyn_cast<VectorType>(DestTy);150  if (!DestVTy)151    return ConstantExpr::getBitCast(C, DestTy);152 153  // If this is a scalar -> vector cast, convert the input into a <1 x scalar>154  // vector so the code below can handle it uniformly.155  if (!isa<VectorType>(C->getType()) &&156      (isa<ConstantFP>(C) || isa<ConstantInt>(C))) {157    Constant *Ops = C; // don't take the address of C!158    return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);159  }160 161  // Some of what follows may extend to cover scalable vectors but the current162  // implementation is fixed length specific.163  if (!isa<FixedVectorType>(C->getType()))164    return ConstantExpr::getBitCast(C, DestTy);165 166  // If this is a bitcast from constant vector -> vector, fold it.167  if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C) &&168      !isa<ConstantInt>(C) && !isa<ConstantFP>(C))169    return ConstantExpr::getBitCast(C, DestTy);170 171  // If the element types match, IR can fold it.172  unsigned NumDstElt = cast<FixedVectorType>(DestVTy)->getNumElements();173  unsigned NumSrcElt = cast<FixedVectorType>(C->getType())->getNumElements();174  if (NumDstElt == NumSrcElt)175    return ConstantExpr::getBitCast(C, DestTy);176 177  Type *SrcEltTy = cast<VectorType>(C->getType())->getElementType();178  Type *DstEltTy = DestVTy->getElementType();179 180  // Otherwise, we're changing the number of elements in a vector, which181  // requires endianness information to do the right thing.  For example,182  //    bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)183  // folds to (little endian):184  //    <4 x i32> <i32 0, i32 0, i32 1, i32 0>185  // and to (big endian):186  //    <4 x i32> <i32 0, i32 0, i32 0, i32 1>187 188  // First thing is first.  We only want to think about integer here, so if189  // we have something in FP form, recast it as integer.190  if (DstEltTy->isFloatingPointTy()) {191    // Fold to an vector of integers with same size as our FP type.192    unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();193    auto *DestIVTy = FixedVectorType::get(194        IntegerType::get(C->getContext(), FPWidth), NumDstElt);195    // Recursively handle this integer conversion, if possible.196    C = FoldBitCast(C, DestIVTy, DL);197 198    // Finally, IR can handle this now that #elts line up.199    return ConstantExpr::getBitCast(C, DestTy);200  }201 202  // Okay, we know the destination is integer, if the input is FP, convert203  // it to integer first.204  if (SrcEltTy->isFloatingPointTy()) {205    unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();206    auto *SrcIVTy = FixedVectorType::get(207        IntegerType::get(C->getContext(), FPWidth), NumSrcElt);208    // Ask IR to do the conversion now that #elts line up.209    C = ConstantExpr::getBitCast(C, SrcIVTy);210    assert((isa<ConstantVector>(C) || // FIXME: Remove ConstantVector.211            isa<ConstantDataVector>(C) || isa<ConstantInt>(C)) &&212           "Constant folding cannot fail for plain fp->int bitcast!");213  }214 215  // Now we know that the input and output vectors are both integer vectors216  // of the same size, and that their #elements is not the same.  Do the217  // conversion here, which depends on whether the input or output has218  // more elements.219  bool isLittleEndian = DL.isLittleEndian();220 221  SmallVector<Constant*, 32> Result;222  if (NumDstElt < NumSrcElt) {223    // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)224    Constant *Zero = Constant::getNullValue(DstEltTy);225    unsigned Ratio = NumSrcElt/NumDstElt;226    unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();227    unsigned SrcElt = 0;228    for (unsigned i = 0; i != NumDstElt; ++i) {229      // Build each element of the result.230      Constant *Elt = Zero;231      unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);232      for (unsigned j = 0; j != Ratio; ++j) {233        Constant *Src = C->getAggregateElement(SrcElt++);234        if (isa_and_nonnull<UndefValue>(Src))235          Src = Constant::getNullValue(236              cast<VectorType>(C->getType())->getElementType());237        else238          Src = dyn_cast_or_null<ConstantInt>(Src);239        if (!Src)  // Reject constantexpr elements.240          return ConstantExpr::getBitCast(C, DestTy);241 242        // Zero extend the element to the right size.243        Src = ConstantFoldCastOperand(Instruction::ZExt, Src, Elt->getType(),244                                      DL);245        assert(Src && "Constant folding cannot fail on plain integers");246 247        // Shift it to the right place, depending on endianness.248        Src = ConstantFoldBinaryOpOperands(249            Instruction::Shl, Src, ConstantInt::get(Src->getType(), ShiftAmt),250            DL);251        assert(Src && "Constant folding cannot fail on plain integers");252 253        ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;254 255        // Mix it in.256        Elt = ConstantFoldBinaryOpOperands(Instruction::Or, Elt, Src, DL);257        assert(Elt && "Constant folding cannot fail on plain integers");258      }259      Result.push_back(Elt);260    }261    return ConstantVector::get(Result);262  }263 264  // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)265  unsigned Ratio = NumDstElt/NumSrcElt;266  unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);267 268  // Loop over each source value, expanding into multiple results.269  for (unsigned i = 0; i != NumSrcElt; ++i) {270    auto *Element = C->getAggregateElement(i);271 272    if (!Element) // Reject constantexpr elements.273      return ConstantExpr::getBitCast(C, DestTy);274 275    if (isa<UndefValue>(Element)) {276      // Correctly Propagate undef values.277      Result.append(Ratio, UndefValue::get(DstEltTy));278      continue;279    }280 281    auto *Src = dyn_cast<ConstantInt>(Element);282    if (!Src)283      return ConstantExpr::getBitCast(C, DestTy);284 285    unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);286    for (unsigned j = 0; j != Ratio; ++j) {287      // Shift the piece of the value into the right place, depending on288      // endianness.289      APInt Elt = Src->getValue().lshr(ShiftAmt);290      ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;291 292      // Truncate and remember this piece.293      Result.push_back(ConstantInt::get(DstEltTy, Elt.trunc(DstBitSize)));294    }295  }296 297  return ConstantVector::get(Result);298}299 300} // end anonymous namespace301 302/// If this constant is a constant offset from a global, return the global and303/// the constant. Because of constantexprs, this function is recursive.304bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,305                                      APInt &Offset, const DataLayout &DL,306                                      DSOLocalEquivalent **DSOEquiv) {307  if (DSOEquiv)308    *DSOEquiv = nullptr;309 310  // Trivial case, constant is the global.311  if ((GV = dyn_cast<GlobalValue>(C))) {312    unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType());313    Offset = APInt(BitWidth, 0);314    return true;315  }316 317  if (auto *FoundDSOEquiv = dyn_cast<DSOLocalEquivalent>(C)) {318    if (DSOEquiv)319      *DSOEquiv = FoundDSOEquiv;320    GV = FoundDSOEquiv->getGlobalValue();321    unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType());322    Offset = APInt(BitWidth, 0);323    return true;324  }325 326  // Otherwise, if this isn't a constant expr, bail out.327  auto *CE = dyn_cast<ConstantExpr>(C);328  if (!CE) return false;329 330  // Look through ptr->int and ptr->ptr casts.331  if (CE->getOpcode() == Instruction::PtrToInt ||332      CE->getOpcode() == Instruction::PtrToAddr ||333      CE->getOpcode() == Instruction::BitCast)334    return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL,335                                      DSOEquiv);336 337  // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)338  auto *GEP = dyn_cast<GEPOperator>(CE);339  if (!GEP)340    return false;341 342  unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());343  APInt TmpOffset(BitWidth, 0);344 345  // If the base isn't a global+constant, we aren't either.346  if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL,347                                  DSOEquiv))348    return false;349 350  // Otherwise, add any offset that our operands provide.351  if (!GEP->accumulateConstantOffset(DL, TmpOffset))352    return false;353 354  Offset = TmpOffset;355  return true;356}357 358Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,359                                               const DataLayout &DL) {360  do {361    Type *SrcTy = C->getType();362    if (SrcTy == DestTy)363      return C;364 365    TypeSize DestSize = DL.getTypeSizeInBits(DestTy);366    TypeSize SrcSize = DL.getTypeSizeInBits(SrcTy);367    if (!TypeSize::isKnownGE(SrcSize, DestSize))368      return nullptr;369 370    // Catch the obvious splat cases (since all-zeros can coerce non-integral371    // pointers legally).372    if (Constant *Res = ConstantFoldLoadFromUniformValue(C, DestTy, DL))373      return Res;374 375    // If the type sizes are the same and a cast is legal, just directly376    // cast the constant.377    // But be careful not to coerce non-integral pointers illegally.378    if (SrcSize == DestSize &&379        DL.isNonIntegralPointerType(SrcTy->getScalarType()) ==380            DL.isNonIntegralPointerType(DestTy->getScalarType())) {381      Instruction::CastOps Cast = Instruction::BitCast;382      // If we are going from a pointer to int or vice versa, we spell the cast383      // differently.384      if (SrcTy->isIntegerTy() && DestTy->isPointerTy())385        Cast = Instruction::IntToPtr;386      else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())387        Cast = Instruction::PtrToInt;388 389      if (CastInst::castIsValid(Cast, C, DestTy))390        return ConstantFoldCastOperand(Cast, C, DestTy, DL);391    }392 393    // If this isn't an aggregate type, there is nothing we can do to drill down394    // and find a bitcastable constant.395    if (!SrcTy->isAggregateType() && !SrcTy->isVectorTy())396      return nullptr;397 398    // We're simulating a load through a pointer that was bitcast to point to399    // a different type, so we can try to walk down through the initial400    // elements of an aggregate to see if some part of the aggregate is401    // castable to implement the "load" semantic model.402    if (SrcTy->isStructTy()) {403      // Struct types might have leading zero-length elements like [0 x i32],404      // which are certainly not what we are looking for, so skip them.405      unsigned Elem = 0;406      Constant *ElemC;407      do {408        ElemC = C->getAggregateElement(Elem++);409      } while (ElemC && DL.getTypeSizeInBits(ElemC->getType()).isZero());410      C = ElemC;411    } else {412      // For non-byte-sized vector elements, the first element is not413      // necessarily located at the vector base address.414      if (auto *VT = dyn_cast<VectorType>(SrcTy))415        if (!DL.typeSizeEqualsStoreSize(VT->getElementType()))416          return nullptr;417 418      C = C->getAggregateElement(0u);419    }420  } while (C);421 422  return nullptr;423}424 425namespace {426 427/// Recursive helper to read bits out of global. C is the constant being copied428/// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy429/// results into and BytesLeft is the number of bytes left in430/// the CurPtr buffer. DL is the DataLayout.431bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,432                        unsigned BytesLeft, const DataLayout &DL) {433  assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&434         "Out of range access");435 436  // Reading type padding, return zero.437  if (ByteOffset >= DL.getTypeStoreSize(C->getType()))438    return true;439 440  // If this element is zero or undefined, we can just return since *CurPtr is441  // zero initialized.442  if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))443    return true;444 445  if (auto *CI = dyn_cast<ConstantInt>(C)) {446    if ((CI->getBitWidth() & 7) != 0)447      return false;448    const APInt &Val = CI->getValue();449    unsigned IntBytes = unsigned(CI->getBitWidth()/8);450 451    for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {452      unsigned n = ByteOffset;453      if (!DL.isLittleEndian())454        n = IntBytes - n - 1;455      CurPtr[i] = Val.extractBits(8, n * 8).getZExtValue();456      ++ByteOffset;457    }458    return true;459  }460 461  if (auto *CFP = dyn_cast<ConstantFP>(C)) {462    if (CFP->getType()->isDoubleTy()) {463      C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);464      return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);465    }466    if (CFP->getType()->isFloatTy()){467      C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);468      return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);469    }470    if (CFP->getType()->isHalfTy()){471      C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);472      return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);473    }474    return false;475  }476 477  if (auto *CS = dyn_cast<ConstantStruct>(C)) {478    const StructLayout *SL = DL.getStructLayout(CS->getType());479    unsigned Index = SL->getElementContainingOffset(ByteOffset);480    uint64_t CurEltOffset = SL->getElementOffset(Index);481    ByteOffset -= CurEltOffset;482 483    while (true) {484      // If the element access is to the element itself and not to tail padding,485      // read the bytes from the element.486      uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());487 488      if (ByteOffset < EltSize &&489          !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,490                              BytesLeft, DL))491        return false;492 493      ++Index;494 495      // Check to see if we read from the last struct element, if so we're done.496      if (Index == CS->getType()->getNumElements())497        return true;498 499      // If we read all of the bytes we needed from this element we're done.500      uint64_t NextEltOffset = SL->getElementOffset(Index);501 502      if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)503        return true;504 505      // Move to the next element of the struct.506      CurPtr += NextEltOffset - CurEltOffset - ByteOffset;507      BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;508      ByteOffset = 0;509      CurEltOffset = NextEltOffset;510    }511    // not reached.512  }513 514  if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||515      isa<ConstantDataSequential>(C)) {516    uint64_t NumElts, EltSize;517    Type *EltTy;518    if (auto *AT = dyn_cast<ArrayType>(C->getType())) {519      NumElts = AT->getNumElements();520      EltTy = AT->getElementType();521      EltSize = DL.getTypeAllocSize(EltTy);522    } else {523      NumElts = cast<FixedVectorType>(C->getType())->getNumElements();524      EltTy = cast<FixedVectorType>(C->getType())->getElementType();525      // TODO: For non-byte-sized vectors, current implementation assumes there is526      // padding to the next byte boundary between elements.527      if (!DL.typeSizeEqualsStoreSize(EltTy))528        return false;529 530      EltSize = DL.getTypeStoreSize(EltTy);531    }532    uint64_t Index = ByteOffset / EltSize;533    uint64_t Offset = ByteOffset - Index * EltSize;534 535    for (; Index != NumElts; ++Index) {536      if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,537                              BytesLeft, DL))538        return false;539 540      uint64_t BytesWritten = EltSize - Offset;541      assert(BytesWritten <= EltSize && "Not indexing into this element?");542      if (BytesWritten >= BytesLeft)543        return true;544 545      Offset = 0;546      BytesLeft -= BytesWritten;547      CurPtr += BytesWritten;548    }549    return true;550  }551 552  if (auto *CE = dyn_cast<ConstantExpr>(C)) {553    if (CE->getOpcode() == Instruction::IntToPtr &&554        CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {555      return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,556                                BytesLeft, DL);557    }558  }559 560  // Otherwise, unknown initializer type.561  return false;562}563 564Constant *FoldReinterpretLoadFromConst(Constant *C, Type *LoadTy,565                                       int64_t Offset, const DataLayout &DL) {566  // Bail out early. Not expect to load from scalable global variable.567  if (isa<ScalableVectorType>(LoadTy))568    return nullptr;569 570  auto *IntType = dyn_cast<IntegerType>(LoadTy);571 572  // If this isn't an integer load we can't fold it directly.573  if (!IntType) {574    // If this is a non-integer load, we can try folding it as an int load and575    // then bitcast the result.  This can be useful for union cases.  Note576    // that address spaces don't matter here since we're not going to result in577    // an actual new load.578    if (!LoadTy->isFloatingPointTy() && !LoadTy->isPointerTy() &&579        !LoadTy->isVectorTy())580      return nullptr;581 582    Type *MapTy = Type::getIntNTy(C->getContext(),583                                  DL.getTypeSizeInBits(LoadTy).getFixedValue());584    if (Constant *Res = FoldReinterpretLoadFromConst(C, MapTy, Offset, DL)) {585      if (Res->isNullValue() && !LoadTy->isX86_AMXTy())586        // Materializing a zero can be done trivially without a bitcast587        return Constant::getNullValue(LoadTy);588      Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy;589      Res = FoldBitCast(Res, CastTy, DL);590      if (LoadTy->isPtrOrPtrVectorTy()) {591        // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr592        if (Res->isNullValue() && !LoadTy->isX86_AMXTy())593          return Constant::getNullValue(LoadTy);594        if (DL.isNonIntegralPointerType(LoadTy->getScalarType()))595          // Be careful not to replace a load of an addrspace value with an inttoptr here596          return nullptr;597        Res = ConstantExpr::getIntToPtr(Res, LoadTy);598      }599      return Res;600    }601    return nullptr;602  }603 604  unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;605  if (BytesLoaded > 32 || BytesLoaded == 0)606    return nullptr;607 608  // If we're not accessing anything in this constant, the result is undefined.609  if (Offset <= -1 * static_cast<int64_t>(BytesLoaded))610    return PoisonValue::get(IntType);611 612  // TODO: We should be able to support scalable types.613  TypeSize InitializerSize = DL.getTypeAllocSize(C->getType());614  if (InitializerSize.isScalable())615    return nullptr;616 617  // If we're not accessing anything in this constant, the result is undefined.618  if (Offset >= (int64_t)InitializerSize.getFixedValue())619    return PoisonValue::get(IntType);620 621  unsigned char RawBytes[32] = {0};622  unsigned char *CurPtr = RawBytes;623  unsigned BytesLeft = BytesLoaded;624 625  // If we're loading off the beginning of the global, some bytes may be valid.626  if (Offset < 0) {627    CurPtr += -Offset;628    BytesLeft += Offset;629    Offset = 0;630  }631 632  if (!ReadDataFromGlobal(C, Offset, CurPtr, BytesLeft, DL))633    return nullptr;634 635  APInt ResultVal = APInt(IntType->getBitWidth(), 0);636  if (DL.isLittleEndian()) {637    ResultVal = RawBytes[BytesLoaded - 1];638    for (unsigned i = 1; i != BytesLoaded; ++i) {639      ResultVal <<= 8;640      ResultVal |= RawBytes[BytesLoaded - 1 - i];641    }642  } else {643    ResultVal = RawBytes[0];644    for (unsigned i = 1; i != BytesLoaded; ++i) {645      ResultVal <<= 8;646      ResultVal |= RawBytes[i];647    }648  }649 650  return ConstantInt::get(IntType->getContext(), ResultVal);651}652 653} // anonymous namespace654 655// If GV is a constant with an initializer read its representation starting656// at Offset and return it as a constant array of unsigned char.  Otherwise657// return null.658Constant *llvm::ReadByteArrayFromGlobal(const GlobalVariable *GV,659                                        uint64_t Offset) {660  if (!GV->isConstant() || !GV->hasDefinitiveInitializer())661    return nullptr;662 663  const DataLayout &DL = GV->getDataLayout();664  Constant *Init = const_cast<Constant *>(GV->getInitializer());665  TypeSize InitSize = DL.getTypeAllocSize(Init->getType());666  if (InitSize < Offset)667    return nullptr;668 669  uint64_t NBytes = InitSize - Offset;670  if (NBytes > UINT16_MAX)671    // Bail for large initializers in excess of 64K to avoid allocating672    // too much memory.673    // Offset is assumed to be less than or equal than InitSize (this674    // is enforced in ReadDataFromGlobal).675    return nullptr;676 677  SmallVector<unsigned char, 256> RawBytes(static_cast<size_t>(NBytes));678  unsigned char *CurPtr = RawBytes.data();679 680  if (!ReadDataFromGlobal(Init, Offset, CurPtr, NBytes, DL))681    return nullptr;682 683  return ConstantDataArray::get(GV->getContext(), RawBytes);684}685 686/// If this Offset points exactly to the start of an aggregate element, return687/// that element, otherwise return nullptr.688Constant *getConstantAtOffset(Constant *Base, APInt Offset,689                              const DataLayout &DL) {690  if (Offset.isZero())691    return Base;692 693  if (!isa<ConstantAggregate>(Base) && !isa<ConstantDataSequential>(Base))694    return nullptr;695 696  Type *ElemTy = Base->getType();697  SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(ElemTy, Offset);698  if (!Offset.isZero() || !Indices[0].isZero())699    return nullptr;700 701  Constant *C = Base;702  for (const APInt &Index : drop_begin(Indices)) {703    if (Index.isNegative() || Index.getActiveBits() >= 32)704      return nullptr;705 706    C = C->getAggregateElement(Index.getZExtValue());707    if (!C)708      return nullptr;709  }710 711  return C;712}713 714Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,715                                          const APInt &Offset,716                                          const DataLayout &DL) {717  if (Constant *AtOffset = getConstantAtOffset(C, Offset, DL))718    if (Constant *Result = ConstantFoldLoadThroughBitcast(AtOffset, Ty, DL))719      return Result;720 721  // Explicitly check for out-of-bounds access, so we return poison even if the722  // constant is a uniform value.723  TypeSize Size = DL.getTypeAllocSize(C->getType());724  if (!Size.isScalable() && Offset.sge(Size.getFixedValue()))725    return PoisonValue::get(Ty);726 727  // Try an offset-independent fold of a uniform value.728  if (Constant *Result = ConstantFoldLoadFromUniformValue(C, Ty, DL))729    return Result;730 731  // Try hard to fold loads from bitcasted strange and non-type-safe things.732  if (Offset.getSignificantBits() <= 64)733    if (Constant *Result =734            FoldReinterpretLoadFromConst(C, Ty, Offset.getSExtValue(), DL))735      return Result;736 737  return nullptr;738}739 740Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,741                                          const DataLayout &DL) {742  return ConstantFoldLoadFromConst(C, Ty, APInt(64, 0), DL);743}744 745Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,746                                             APInt Offset,747                                             const DataLayout &DL) {748  // We can only fold loads from constant globals with a definitive initializer.749  // Check this upfront, to skip expensive offset calculations.750  auto *GV = dyn_cast<GlobalVariable>(getUnderlyingObject(C));751  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())752    return nullptr;753 754  C = cast<Constant>(C->stripAndAccumulateConstantOffsets(755          DL, Offset, /* AllowNonInbounds */ true));756 757  if (C == GV)758    if (Constant *Result = ConstantFoldLoadFromConst(GV->getInitializer(), Ty,759                                                     Offset, DL))760      return Result;761 762  // If this load comes from anywhere in a uniform constant global, the value763  // is always the same, regardless of the loaded offset.764  return ConstantFoldLoadFromUniformValue(GV->getInitializer(), Ty, DL);765}766 767Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,768                                             const DataLayout &DL) {769  APInt Offset(DL.getIndexTypeSizeInBits(C->getType()), 0);770  return ConstantFoldLoadFromConstPtr(C, Ty, std::move(Offset), DL);771}772 773Constant *llvm::ConstantFoldLoadFromUniformValue(Constant *C, Type *Ty,774                                                 const DataLayout &DL) {775  if (isa<PoisonValue>(C))776    return PoisonValue::get(Ty);777  if (isa<UndefValue>(C))778    return UndefValue::get(Ty);779  // If padding is needed when storing C to memory, then it isn't considered as780  // uniform.781  if (!DL.typeSizeEqualsStoreSize(C->getType()))782    return nullptr;783  if (C->isNullValue() && !Ty->isX86_AMXTy())784    return Constant::getNullValue(Ty);785  if (C->isAllOnesValue() &&786      (Ty->isIntOrIntVectorTy() || Ty->isFPOrFPVectorTy()))787    return Constant::getAllOnesValue(Ty);788  return nullptr;789}790 791namespace {792 793/// One of Op0/Op1 is a constant expression.794/// Attempt to symbolically evaluate the result of a binary operator merging795/// these together.  If target data info is available, it is provided as DL,796/// otherwise DL is null.797Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,798                                    const DataLayout &DL) {799  // SROA800 801  // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.802  // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute803  // bits.804 805  if (Opc == Instruction::And) {806    KnownBits Known0 = computeKnownBits(Op0, DL);807    KnownBits Known1 = computeKnownBits(Op1, DL);808    if ((Known1.One | Known0.Zero).isAllOnes()) {809      // All the bits of Op0 that the 'and' could be masking are already zero.810      return Op0;811    }812    if ((Known0.One | Known1.Zero).isAllOnes()) {813      // All the bits of Op1 that the 'and' could be masking are already zero.814      return Op1;815    }816 817    Known0 &= Known1;818    if (Known0.isConstant())819      return ConstantInt::get(Op0->getType(), Known0.getConstant());820  }821 822  // If the constant expr is something like &A[123] - &A[4].f, fold this into a823  // constant.  This happens frequently when iterating over a global array.824  if (Opc == Instruction::Sub) {825    GlobalValue *GV1, *GV2;826    APInt Offs1, Offs2;827 828    if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))829      if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {830        unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());831 832        // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.833        // PtrToInt may change the bitwidth so we have convert to the right size834        // first.835        return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -836                                                Offs2.zextOrTrunc(OpSize));837      }838  }839 840  return nullptr;841}842 843/// If array indices are not pointer-sized integers, explicitly cast them so844/// that they aren't implicitly casted by the getelementptr.845Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,846                         Type *ResultTy, GEPNoWrapFlags NW,847                         std::optional<ConstantRange> InRange,848                         const DataLayout &DL, const TargetLibraryInfo *TLI) {849  Type *IntIdxTy = DL.getIndexType(ResultTy);850  Type *IntIdxScalarTy = IntIdxTy->getScalarType();851 852  bool Any = false;853  SmallVector<Constant*, 32> NewIdxs;854  for (unsigned i = 1, e = Ops.size(); i != e; ++i) {855    if ((i == 1 ||856         !isa<StructType>(GetElementPtrInst::getIndexedType(857             SrcElemTy, Ops.slice(1, i - 1)))) &&858        Ops[i]->getType()->getScalarType() != IntIdxScalarTy) {859      Any = true;860      Type *NewType =861          Ops[i]->getType()->isVectorTy() ? IntIdxTy : IntIdxScalarTy;862      Constant *NewIdx = ConstantFoldCastOperand(863          CastInst::getCastOpcode(Ops[i], true, NewType, true), Ops[i], NewType,864          DL);865      if (!NewIdx)866        return nullptr;867      NewIdxs.push_back(NewIdx);868    } else869      NewIdxs.push_back(Ops[i]);870  }871 872  if (!Any)873    return nullptr;874 875  Constant *C =876      ConstantExpr::getGetElementPtr(SrcElemTy, Ops[0], NewIdxs, NW, InRange);877  return ConstantFoldConstant(C, DL, TLI);878}879 880/// If we can symbolically evaluate the GEP constant expression, do so.881Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,882                                  ArrayRef<Constant *> Ops,883                                  const DataLayout &DL,884                                  const TargetLibraryInfo *TLI) {885  Type *SrcElemTy = GEP->getSourceElementType();886  Type *ResTy = GEP->getType();887  if (!SrcElemTy->isSized() || isa<ScalableVectorType>(SrcElemTy))888    return nullptr;889 890  if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy, GEP->getNoWrapFlags(),891                                   GEP->getInRange(), DL, TLI))892    return C;893 894  Constant *Ptr = Ops[0];895  if (!Ptr->getType()->isPointerTy())896    return nullptr;897 898  Type *IntIdxTy = DL.getIndexType(Ptr->getType());899 900  for (unsigned i = 1, e = Ops.size(); i != e; ++i)901    if (!isa<ConstantInt>(Ops[i]) || !Ops[i]->getType()->isIntegerTy())902      return nullptr;903 904  unsigned BitWidth = DL.getTypeSizeInBits(IntIdxTy);905  APInt Offset = APInt(906      BitWidth,907      DL.getIndexedOffsetInType(908          SrcElemTy, ArrayRef((Value *const *)Ops.data() + 1, Ops.size() - 1)),909      /*isSigned=*/true, /*implicitTrunc=*/true);910 911  std::optional<ConstantRange> InRange = GEP->getInRange();912  if (InRange)913    InRange = InRange->sextOrTrunc(BitWidth);914 915  // If this is a GEP of a GEP, fold it all into a single GEP.916  GEPNoWrapFlags NW = GEP->getNoWrapFlags();917  bool Overflow = false;918  while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {919    NW &= GEP->getNoWrapFlags();920 921    SmallVector<Value *, 4> NestedOps(llvm::drop_begin(GEP->operands()));922 923    // Do not try the incorporate the sub-GEP if some index is not a number.924    bool AllConstantInt = true;925    for (Value *NestedOp : NestedOps)926      if (!isa<ConstantInt>(NestedOp)) {927        AllConstantInt = false;928        break;929      }930    if (!AllConstantInt)931      break;932 933    // Adjust inrange offset and intersect inrange attributes934    if (auto GEPRange = GEP->getInRange()) {935      auto AdjustedGEPRange = GEPRange->sextOrTrunc(BitWidth).subtract(Offset);936      InRange =937          InRange ? InRange->intersectWith(AdjustedGEPRange) : AdjustedGEPRange;938    }939 940    Ptr = cast<Constant>(GEP->getOperand(0));941    SrcElemTy = GEP->getSourceElementType();942    Offset = Offset.sadd_ov(943        APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps),944              /*isSigned=*/true, /*implicitTrunc=*/true),945        Overflow);946  }947 948  // Preserving nusw (without inbounds) also requires that the offset949  // additions did not overflow.950  if (NW.hasNoUnsignedSignedWrap() && !NW.isInBounds() && Overflow)951    NW = NW.withoutNoUnsignedSignedWrap();952 953  // If the base value for this address is a literal integer value, fold the954  // getelementptr to the resulting integer value casted to the pointer type.955  APInt BaseIntVal(DL.getPointerTypeSizeInBits(Ptr->getType()), 0);956  if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) {957    if (CE->getOpcode() == Instruction::IntToPtr) {958      if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))959        BaseIntVal = Base->getValue().zextOrTrunc(BaseIntVal.getBitWidth());960    }961  }962 963  if ((Ptr->isNullValue() || BaseIntVal != 0) &&964      !DL.mustNotIntroduceIntToPtr(Ptr->getType())) {965 966    // If the index size is smaller than the pointer size, add to the low967    // bits only.968    BaseIntVal.insertBits(BaseIntVal.trunc(BitWidth) + Offset, 0);969    Constant *C = ConstantInt::get(Ptr->getContext(), BaseIntVal);970    return ConstantExpr::getIntToPtr(C, ResTy);971  }972 973  // Try to infer inbounds for GEPs of globals.974  if (!NW.isInBounds() && Offset.isNonNegative()) {975    bool CanBeNull, CanBeFreed;976    uint64_t DerefBytes =977        Ptr->getPointerDereferenceableBytes(DL, CanBeNull, CanBeFreed);978    if (DerefBytes != 0 && !CanBeNull && Offset.sle(DerefBytes))979      NW |= GEPNoWrapFlags::inBounds();980  }981 982  // nusw + nneg -> nuw983  if (NW.hasNoUnsignedSignedWrap() && Offset.isNonNegative())984    NW |= GEPNoWrapFlags::noUnsignedWrap();985 986  // Otherwise canonicalize this to a single ptradd.987  LLVMContext &Ctx = Ptr->getContext();988  return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ctx), Ptr,989                                        ConstantInt::get(Ctx, Offset), NW,990                                        InRange);991}992 993/// Attempt to constant fold an instruction with the994/// specified opcode and operands.  If successful, the constant result is995/// returned, if not, null is returned.  Note that this function can fail when996/// attempting to fold instructions like loads and stores, which have no997/// constant expression form.998Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,999                                       ArrayRef<Constant *> Ops,1000                                       const DataLayout &DL,1001                                       const TargetLibraryInfo *TLI,1002                                       bool AllowNonDeterministic) {1003  Type *DestTy = InstOrCE->getType();1004 1005  if (Instruction::isUnaryOp(Opcode))1006    return ConstantFoldUnaryOpOperand(Opcode, Ops[0], DL);1007 1008  if (Instruction::isBinaryOp(Opcode)) {1009    switch (Opcode) {1010    default:1011      break;1012    case Instruction::FAdd:1013    case Instruction::FSub:1014    case Instruction::FMul:1015    case Instruction::FDiv:1016    case Instruction::FRem:1017      // Handle floating point instructions separately to account for denormals1018      // TODO: If a constant expression is being folded rather than an1019      // instruction, denormals will not be flushed/treated as zero1020      if (const auto *I = dyn_cast<Instruction>(InstOrCE)) {1021        return ConstantFoldFPInstOperands(Opcode, Ops[0], Ops[1], DL, I,1022                                          AllowNonDeterministic);1023      }1024    }1025    return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL);1026  }1027 1028  if (Instruction::isCast(Opcode))1029    return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL);1030 1031  if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) {1032    Type *SrcElemTy = GEP->getSourceElementType();1033    if (!ConstantExpr::isSupportedGetElementPtr(SrcElemTy))1034      return nullptr;1035 1036    if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))1037      return C;1038 1039    return ConstantExpr::getGetElementPtr(SrcElemTy, Ops[0], Ops.slice(1),1040                                          GEP->getNoWrapFlags(),1041                                          GEP->getInRange());1042  }1043 1044  if (auto *CE = dyn_cast<ConstantExpr>(InstOrCE))1045    return CE->getWithOperands(Ops);1046 1047  switch (Opcode) {1048  default: return nullptr;1049  case Instruction::ICmp:1050  case Instruction::FCmp: {1051    auto *C = cast<CmpInst>(InstOrCE);1052    return ConstantFoldCompareInstOperands(C->getPredicate(), Ops[0], Ops[1],1053                                           DL, TLI, C);1054  }1055  case Instruction::Freeze:1056    return isGuaranteedNotToBeUndefOrPoison(Ops[0]) ? Ops[0] : nullptr;1057  case Instruction::Call:1058    if (auto *F = dyn_cast<Function>(Ops.back())) {1059      const auto *Call = cast<CallBase>(InstOrCE);1060      if (canConstantFoldCallTo(Call, F))1061        return ConstantFoldCall(Call, F, Ops.slice(0, Ops.size() - 1), TLI,1062                                AllowNonDeterministic);1063    }1064    return nullptr;1065  case Instruction::Select:1066    return ConstantFoldSelectInstruction(Ops[0], Ops[1], Ops[2]);1067  case Instruction::ExtractElement:1068    return ConstantExpr::getExtractElement(Ops[0], Ops[1]);1069  case Instruction::ExtractValue:1070    return ConstantFoldExtractValueInstruction(1071        Ops[0], cast<ExtractValueInst>(InstOrCE)->getIndices());1072  case Instruction::InsertElement:1073    return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);1074  case Instruction::InsertValue:1075    return ConstantFoldInsertValueInstruction(1076        Ops[0], Ops[1], cast<InsertValueInst>(InstOrCE)->getIndices());1077  case Instruction::ShuffleVector:1078    return ConstantExpr::getShuffleVector(1079        Ops[0], Ops[1], cast<ShuffleVectorInst>(InstOrCE)->getShuffleMask());1080  case Instruction::Load: {1081    const auto *LI = dyn_cast<LoadInst>(InstOrCE);1082    if (LI->isVolatile())1083      return nullptr;1084    return ConstantFoldLoadFromConstPtr(Ops[0], LI->getType(), DL);1085  }1086  }1087}1088 1089} // end anonymous namespace1090 1091//===----------------------------------------------------------------------===//1092// Constant Folding public APIs1093//===----------------------------------------------------------------------===//1094 1095namespace {1096 1097Constant *1098ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,1099                         const TargetLibraryInfo *TLI,1100                         SmallDenseMap<Constant *, Constant *> &FoldedOps) {1101  if (!isa<ConstantVector>(C) && !isa<ConstantExpr>(C))1102    return const_cast<Constant *>(C);1103 1104  SmallVector<Constant *, 8> Ops;1105  for (const Use &OldU : C->operands()) {1106    Constant *OldC = cast<Constant>(&OldU);1107    Constant *NewC = OldC;1108    // Recursively fold the ConstantExpr's operands. If we have already folded1109    // a ConstantExpr, we don't have to process it again.1110    if (isa<ConstantVector>(OldC) || isa<ConstantExpr>(OldC)) {1111      auto It = FoldedOps.find(OldC);1112      if (It == FoldedOps.end()) {1113        NewC = ConstantFoldConstantImpl(OldC, DL, TLI, FoldedOps);1114        FoldedOps.insert({OldC, NewC});1115      } else {1116        NewC = It->second;1117      }1118    }1119    Ops.push_back(NewC);1120  }1121 1122  if (auto *CE = dyn_cast<ConstantExpr>(C)) {1123    if (Constant *Res = ConstantFoldInstOperandsImpl(1124            CE, CE->getOpcode(), Ops, DL, TLI, /*AllowNonDeterministic=*/true))1125      return Res;1126    return const_cast<Constant *>(C);1127  }1128 1129  assert(isa<ConstantVector>(C));1130  return ConstantVector::get(Ops);1131}1132 1133} // end anonymous namespace1134 1135Constant *llvm::ConstantFoldInstruction(const Instruction *I,1136                                        const DataLayout &DL,1137                                        const TargetLibraryInfo *TLI) {1138  // Handle PHI nodes quickly here...1139  if (auto *PN = dyn_cast<PHINode>(I)) {1140    Constant *CommonValue = nullptr;1141 1142    SmallDenseMap<Constant *, Constant *> FoldedOps;1143    for (Value *Incoming : PN->incoming_values()) {1144      // If the incoming value is undef then skip it.  Note that while we could1145      // skip the value if it is equal to the phi node itself we choose not to1146      // because that would break the rule that constant folding only applies if1147      // all operands are constants.1148      if (isa<UndefValue>(Incoming))1149        continue;1150      // If the incoming value is not a constant, then give up.1151      auto *C = dyn_cast<Constant>(Incoming);1152      if (!C)1153        return nullptr;1154      // Fold the PHI's operands.1155      C = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);1156      // If the incoming value is a different constant to1157      // the one we saw previously, then give up.1158      if (CommonValue && C != CommonValue)1159        return nullptr;1160      CommonValue = C;1161    }1162 1163    // If we reach here, all incoming values are the same constant or undef.1164    return CommonValue ? CommonValue : UndefValue::get(PN->getType());1165  }1166 1167  // Scan the operand list, checking to see if they are all constants, if so,1168  // hand off to ConstantFoldInstOperandsImpl.1169  if (!all_of(I->operands(), [](const Use &U) { return isa<Constant>(U); }))1170    return nullptr;1171 1172  SmallDenseMap<Constant *, Constant *> FoldedOps;1173  SmallVector<Constant *, 8> Ops;1174  for (const Use &OpU : I->operands()) {1175    auto *Op = cast<Constant>(&OpU);1176    // Fold the Instruction's operands.1177    Op = ConstantFoldConstantImpl(Op, DL, TLI, FoldedOps);1178    Ops.push_back(Op);1179  }1180 1181  return ConstantFoldInstOperands(I, Ops, DL, TLI);1182}1183 1184Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,1185                                     const TargetLibraryInfo *TLI) {1186  SmallDenseMap<Constant *, Constant *> FoldedOps;1187  return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);1188}1189 1190Constant *llvm::ConstantFoldInstOperands(const Instruction *I,1191                                         ArrayRef<Constant *> Ops,1192                                         const DataLayout &DL,1193                                         const TargetLibraryInfo *TLI,1194                                         bool AllowNonDeterministic) {1195  return ConstantFoldInstOperandsImpl(I, I->getOpcode(), Ops, DL, TLI,1196                                      AllowNonDeterministic);1197}1198 1199Constant *llvm::ConstantFoldCompareInstOperands(1200    unsigned IntPredicate, Constant *Ops0, Constant *Ops1, const DataLayout &DL,1201    const TargetLibraryInfo *TLI, const Instruction *I) {1202  CmpInst::Predicate Predicate = (CmpInst::Predicate)IntPredicate;1203  // fold: icmp (inttoptr x), null         -> icmp x, 01204  // fold: icmp null, (inttoptr x)         -> icmp 0, x1205  // fold: icmp (ptrtoint x), 0            -> icmp x, null1206  // fold: icmp 0, (ptrtoint x)            -> icmp null, x1207  // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y1208  // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y1209  //1210  // FIXME: The following comment is out of data and the DataLayout is here now.1211  // ConstantExpr::getCompare cannot do this, because it doesn't have DL1212  // around to know if bit truncation is happening.1213  if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) {1214    if (Ops1->isNullValue()) {1215      if (CE0->getOpcode() == Instruction::IntToPtr) {1216        Type *IntPtrTy = DL.getIntPtrType(CE0->getType());1217        // Convert the integer value to the right size to ensure we get the1218        // proper extension or truncation.1219        if (Constant *C = ConstantFoldIntegerCast(CE0->getOperand(0), IntPtrTy,1220                                                  /*IsSigned*/ false, DL)) {1221          Constant *Null = Constant::getNullValue(C->getType());1222          return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);1223        }1224      }1225 1226      // Only do this transformation if the int is intptrty in size, otherwise1227      // there is a truncation or extension that we aren't modeling.1228      if (CE0->getOpcode() == Instruction::PtrToInt) {1229        Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());1230        if (CE0->getType() == IntPtrTy) {1231          Constant *C = CE0->getOperand(0);1232          Constant *Null = Constant::getNullValue(C->getType());1233          return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);1234        }1235      }1236    }1237 1238    if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) {1239      if (CE0->getOpcode() == CE1->getOpcode()) {1240        if (CE0->getOpcode() == Instruction::IntToPtr) {1241          Type *IntPtrTy = DL.getIntPtrType(CE0->getType());1242 1243          // Convert the integer value to the right size to ensure we get the1244          // proper extension or truncation.1245          Constant *C0 = ConstantFoldIntegerCast(CE0->getOperand(0), IntPtrTy,1246                                                 /*IsSigned*/ false, DL);1247          Constant *C1 = ConstantFoldIntegerCast(CE1->getOperand(0), IntPtrTy,1248                                                 /*IsSigned*/ false, DL);1249          if (C0 && C1)1250            return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);1251        }1252 1253        // Only do this transformation if the int is intptrty in size, otherwise1254        // there is a truncation or extension that we aren't modeling.1255        if (CE0->getOpcode() == Instruction::PtrToInt) {1256          Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());1257          if (CE0->getType() == IntPtrTy &&1258              CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {1259            return ConstantFoldCompareInstOperands(1260                Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);1261          }1262        }1263      }1264    }1265 1266    // Convert pointer comparison (base+offset1) pred (base+offset2) into1267    // offset1 pred offset2, for the case where the offset is inbounds. This1268    // only works for equality and unsigned comparison, as inbounds permits1269    // crossing the sign boundary. However, the offset comparison itself is1270    // signed.1271    if (Ops0->getType()->isPointerTy() && !ICmpInst::isSigned(Predicate)) {1272      unsigned IndexWidth = DL.getIndexTypeSizeInBits(Ops0->getType());1273      APInt Offset0(IndexWidth, 0);1274      bool IsEqPred = ICmpInst::isEquality(Predicate);1275      Value *Stripped0 = Ops0->stripAndAccumulateConstantOffsets(1276          DL, Offset0, /*AllowNonInbounds=*/IsEqPred,1277          /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr,1278          /*LookThroughIntToPtr=*/IsEqPred);1279      APInt Offset1(IndexWidth, 0);1280      Value *Stripped1 = Ops1->stripAndAccumulateConstantOffsets(1281          DL, Offset1, /*AllowNonInbounds=*/IsEqPred,1282          /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr,1283          /*LookThroughIntToPtr=*/IsEqPred);1284      if (Stripped0 == Stripped1)1285        return ConstantInt::getBool(1286            Ops0->getContext(),1287            ICmpInst::compare(Offset0, Offset1,1288                              ICmpInst::getSignedPredicate(Predicate)));1289    }1290  } else if (isa<ConstantExpr>(Ops1)) {1291    // If RHS is a constant expression, but the left side isn't, swap the1292    // operands and try again.1293    Predicate = ICmpInst::getSwappedPredicate(Predicate);1294    return ConstantFoldCompareInstOperands(Predicate, Ops1, Ops0, DL, TLI);1295  }1296 1297  if (CmpInst::isFPPredicate(Predicate)) {1298    // Flush any denormal constant float input according to denormal handling1299    // mode.1300    Ops0 = FlushFPConstant(Ops0, I, /*IsOutput=*/false);1301    if (!Ops0)1302      return nullptr;1303    Ops1 = FlushFPConstant(Ops1, I, /*IsOutput=*/false);1304    if (!Ops1)1305      return nullptr;1306  }1307 1308  return ConstantFoldCompareInstruction(Predicate, Ops0, Ops1);1309}1310 1311Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op,1312                                           const DataLayout &DL) {1313  assert(Instruction::isUnaryOp(Opcode));1314 1315  return ConstantFoldUnaryInstruction(Opcode, Op);1316}1317 1318Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,1319                                             Constant *RHS,1320                                             const DataLayout &DL) {1321  assert(Instruction::isBinaryOp(Opcode));1322  if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS))1323    if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL))1324      return C;1325 1326  if (ConstantExpr::isDesirableBinOp(Opcode))1327    return ConstantExpr::get(Opcode, LHS, RHS);1328  return ConstantFoldBinaryInstruction(Opcode, LHS, RHS);1329}1330 1331static ConstantFP *flushDenormalConstant(Type *Ty, const APFloat &APF,1332                                         DenormalMode::DenormalModeKind Mode) {1333  switch (Mode) {1334  case DenormalMode::Dynamic:1335    return nullptr;1336  case DenormalMode::IEEE:1337    return ConstantFP::get(Ty->getContext(), APF);1338  case DenormalMode::PreserveSign:1339    return ConstantFP::get(1340        Ty->getContext(),1341        APFloat::getZero(APF.getSemantics(), APF.isNegative()));1342  case DenormalMode::PositiveZero:1343    return ConstantFP::get(Ty->getContext(),1344                           APFloat::getZero(APF.getSemantics(), false));1345  default:1346    break;1347  }1348 1349  llvm_unreachable("unknown denormal mode");1350}1351 1352/// Return the denormal mode that can be assumed when executing a floating point1353/// operation at \p CtxI.1354static DenormalMode getInstrDenormalMode(const Instruction *CtxI, Type *Ty) {1355  if (!CtxI || !CtxI->getParent() || !CtxI->getFunction())1356    return DenormalMode::getDynamic();1357  return CtxI->getFunction()->getDenormalMode(Ty->getFltSemantics());1358}1359 1360static ConstantFP *flushDenormalConstantFP(ConstantFP *CFP,1361                                           const Instruction *Inst,1362                                           bool IsOutput) {1363  const APFloat &APF = CFP->getValueAPF();1364  if (!APF.isDenormal())1365    return CFP;1366 1367  DenormalMode Mode = getInstrDenormalMode(Inst, CFP->getType());1368  return flushDenormalConstant(CFP->getType(), APF,1369                               IsOutput ? Mode.Output : Mode.Input);1370}1371 1372Constant *llvm::FlushFPConstant(Constant *Operand, const Instruction *Inst,1373                                bool IsOutput) {1374  if (ConstantFP *CFP = dyn_cast<ConstantFP>(Operand))1375    return flushDenormalConstantFP(CFP, Inst, IsOutput);1376 1377  if (isa<ConstantAggregateZero, UndefValue>(Operand))1378    return Operand;1379 1380  Type *Ty = Operand->getType();1381  VectorType *VecTy = dyn_cast<VectorType>(Ty);1382  if (VecTy) {1383    if (auto *Splat = dyn_cast_or_null<ConstantFP>(Operand->getSplatValue())) {1384      ConstantFP *Folded = flushDenormalConstantFP(Splat, Inst, IsOutput);1385      if (!Folded)1386        return nullptr;1387      return ConstantVector::getSplat(VecTy->getElementCount(), Folded);1388    }1389 1390    Ty = VecTy->getElementType();1391  }1392 1393  if (isa<ConstantExpr>(Operand))1394    return Operand;1395 1396  if (const auto *CV = dyn_cast<ConstantVector>(Operand)) {1397    SmallVector<Constant *, 16> NewElts;1398    for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {1399      Constant *Element = CV->getAggregateElement(i);1400      if (isa<UndefValue>(Element)) {1401        NewElts.push_back(Element);1402        continue;1403      }1404 1405      ConstantFP *CFP = dyn_cast<ConstantFP>(Element);1406      if (!CFP)1407        return nullptr;1408 1409      ConstantFP *Folded = flushDenormalConstantFP(CFP, Inst, IsOutput);1410      if (!Folded)1411        return nullptr;1412      NewElts.push_back(Folded);1413    }1414 1415    return ConstantVector::get(NewElts);1416  }1417 1418  if (const auto *CDV = dyn_cast<ConstantDataVector>(Operand)) {1419    SmallVector<Constant *, 16> NewElts;1420    for (unsigned I = 0, E = CDV->getNumElements(); I < E; ++I) {1421      const APFloat &Elt = CDV->getElementAsAPFloat(I);1422      if (!Elt.isDenormal()) {1423        NewElts.push_back(ConstantFP::get(Ty, Elt));1424      } else {1425        DenormalMode Mode = getInstrDenormalMode(Inst, Ty);1426        ConstantFP *Folded =1427            flushDenormalConstant(Ty, Elt, IsOutput ? Mode.Output : Mode.Input);1428        if (!Folded)1429          return nullptr;1430        NewElts.push_back(Folded);1431      }1432    }1433 1434    return ConstantVector::get(NewElts);1435  }1436 1437  return nullptr;1438}1439 1440Constant *llvm::ConstantFoldFPInstOperands(unsigned Opcode, Constant *LHS,1441                                           Constant *RHS, const DataLayout &DL,1442                                           const Instruction *I,1443                                           bool AllowNonDeterministic) {1444  if (Instruction::isBinaryOp(Opcode)) {1445    // Flush denormal inputs if needed.1446    Constant *Op0 = FlushFPConstant(LHS, I, /* IsOutput */ false);1447    if (!Op0)1448      return nullptr;1449    Constant *Op1 = FlushFPConstant(RHS, I, /* IsOutput */ false);1450    if (!Op1)1451      return nullptr;1452 1453    // If nsz or an algebraic FMF flag is set, the result of the FP operation1454    // may change due to future optimization. Don't constant fold them if1455    // non-deterministic results are not allowed.1456    if (!AllowNonDeterministic)1457      if (auto *FP = dyn_cast_or_null<FPMathOperator>(I))1458        if (FP->hasNoSignedZeros() || FP->hasAllowReassoc() ||1459            FP->hasAllowContract() || FP->hasAllowReciprocal())1460          return nullptr;1461 1462    // Calculate constant result.1463    Constant *C = ConstantFoldBinaryOpOperands(Opcode, Op0, Op1, DL);1464    if (!C)1465      return nullptr;1466 1467    // Flush denormal output if needed.1468    C = FlushFPConstant(C, I, /* IsOutput */ true);1469    if (!C)1470      return nullptr;1471 1472    // The precise NaN value is non-deterministic.1473    if (!AllowNonDeterministic && C->isNaN())1474      return nullptr;1475 1476    return C;1477  }1478  // If instruction lacks a parent/function and the denormal mode cannot be1479  // determined, use the default (IEEE).1480  return ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL);1481}1482 1483Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,1484                                        Type *DestTy, const DataLayout &DL) {1485  assert(Instruction::isCast(Opcode));1486 1487  if (auto *CE = dyn_cast<ConstantExpr>(C))1488    if (CE->isCast())1489      if (unsigned NewOp = CastInst::isEliminableCastPair(1490              Instruction::CastOps(CE->getOpcode()),1491              Instruction::CastOps(Opcode), CE->getOperand(0)->getType(),1492              C->getType(), DestTy, &DL))1493        return ConstantFoldCastOperand(NewOp, CE->getOperand(0), DestTy, DL);1494 1495  switch (Opcode) {1496  default:1497    llvm_unreachable("Missing case");1498  case Instruction::PtrToAddr:1499  case Instruction::PtrToInt:1500    if (auto *CE = dyn_cast<ConstantExpr>(C)) {1501      Constant *FoldedValue = nullptr;1502      // If the input is an inttoptr, eliminate the pair.  This requires knowing1503      // the width of a pointer, so it can't be done in ConstantExpr::getCast.1504      if (CE->getOpcode() == Instruction::IntToPtr) {1505        // zext/trunc the inttoptr to pointer/address size.1506        Type *MidTy = Opcode == Instruction::PtrToInt1507                          ? DL.getAddressType(CE->getType())1508                          : DL.getIntPtrType(CE->getType());1509        FoldedValue = ConstantFoldIntegerCast(CE->getOperand(0), MidTy,1510                                              /*IsSigned=*/false, DL);1511      } else if (auto *GEP = dyn_cast<GEPOperator>(CE)) {1512        // If we have GEP, we can perform the following folds:1513        // (ptrtoint/ptrtoaddr (gep null, x)) -> x1514        // (ptrtoint/ptrtoaddr (gep (gep null, x), y) -> x + y, etc.1515        unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());1516        APInt BaseOffset(BitWidth, 0);1517        auto *Base = cast<Constant>(GEP->stripAndAccumulateConstantOffsets(1518            DL, BaseOffset, /*AllowNonInbounds=*/true));1519        if (Base->isNullValue()) {1520          FoldedValue = ConstantInt::get(CE->getContext(), BaseOffset);1521        } else {1522          // ptrtoint/ptrtoaddr (gep i8, Ptr, (sub 0, V))1523          //   -> sub (ptrtoint/ptrtoaddr Ptr), V1524          if (GEP->getNumIndices() == 1 &&1525              GEP->getSourceElementType()->isIntegerTy(8)) {1526            auto *Ptr = cast<Constant>(GEP->getPointerOperand());1527            auto *Sub = dyn_cast<ConstantExpr>(GEP->getOperand(1));1528            Type *IntIdxTy = DL.getIndexType(Ptr->getType());1529            if (Sub && Sub->getType() == IntIdxTy &&1530                Sub->getOpcode() == Instruction::Sub &&1531                Sub->getOperand(0)->isNullValue())1532              FoldedValue = ConstantExpr::getSub(1533                  ConstantExpr::getCast(Opcode, Ptr, IntIdxTy),1534                  Sub->getOperand(1));1535          }1536        }1537      }1538      if (FoldedValue) {1539        // Do a zext or trunc to get to the ptrtoint/ptrtoaddr dest size.1540        return ConstantFoldIntegerCast(FoldedValue, DestTy, /*IsSigned=*/false,1541                                       DL);1542      }1543    }1544    break;1545  case Instruction::IntToPtr:1546    // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if1547    // the int size is >= the ptr size and the address spaces are the same.1548    // This requires knowing the width of a pointer, so it can't be done in1549    // ConstantExpr::getCast.1550    if (auto *CE = dyn_cast<ConstantExpr>(C)) {1551      if (CE->getOpcode() == Instruction::PtrToInt) {1552        Constant *SrcPtr = CE->getOperand(0);1553        unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());1554        unsigned MidIntSize = CE->getType()->getScalarSizeInBits();1555 1556        if (MidIntSize >= SrcPtrSize) {1557          unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();1558          if (SrcAS == DestTy->getPointerAddressSpace())1559            return FoldBitCast(CE->getOperand(0), DestTy, DL);1560        }1561      }1562    }1563    break;1564  case Instruction::Trunc:1565  case Instruction::ZExt:1566  case Instruction::SExt:1567  case Instruction::FPTrunc:1568  case Instruction::FPExt:1569  case Instruction::UIToFP:1570  case Instruction::SIToFP:1571  case Instruction::FPToUI:1572  case Instruction::FPToSI:1573  case Instruction::AddrSpaceCast:1574    break;1575  case Instruction::BitCast:1576    return FoldBitCast(C, DestTy, DL);1577  }1578 1579  if (ConstantExpr::isDesirableCastOp(Opcode))1580    return ConstantExpr::getCast(Opcode, C, DestTy);1581  return ConstantFoldCastInstruction(Opcode, C, DestTy);1582}1583 1584Constant *llvm::ConstantFoldIntegerCast(Constant *C, Type *DestTy,1585                                        bool IsSigned, const DataLayout &DL) {1586  Type *SrcTy = C->getType();1587  if (SrcTy == DestTy)1588    return C;1589  if (SrcTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits())1590    return ConstantFoldCastOperand(Instruction::Trunc, C, DestTy, DL);1591  if (IsSigned)1592    return ConstantFoldCastOperand(Instruction::SExt, C, DestTy, DL);1593  return ConstantFoldCastOperand(Instruction::ZExt, C, DestTy, DL);1594}1595 1596//===----------------------------------------------------------------------===//1597//  Constant Folding for Calls1598//1599 1600bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {1601  if (Call->isNoBuiltin())1602    return false;1603  if (Call->getFunctionType() != F->getFunctionType())1604    return false;1605 1606  // Allow FP calls (both libcalls and intrinsics) to avoid being folded.1607  // This can be useful for GPU targets or in cross-compilation scenarios1608  // when the exact target FP behaviour is required, and the host compiler's1609  // behaviour may be slightly different from the device's run-time behaviour.1610  if (DisableFPCallFolding && (F->getReturnType()->isFloatingPointTy() ||1611                               any_of(F->args(), [](const Argument &Arg) {1612                                 return Arg.getType()->isFloatingPointTy();1613                               })))1614    return false;1615 1616  switch (F->getIntrinsicID()) {1617  // Operations that do not operate floating-point numbers and do not depend on1618  // FP environment can be folded even in strictfp functions.1619  case Intrinsic::bswap:1620  case Intrinsic::ctpop:1621  case Intrinsic::ctlz:1622  case Intrinsic::cttz:1623  case Intrinsic::fshl:1624  case Intrinsic::fshr:1625  case Intrinsic::launder_invariant_group:1626  case Intrinsic::strip_invariant_group:1627  case Intrinsic::masked_load:1628  case Intrinsic::get_active_lane_mask:1629  case Intrinsic::abs:1630  case Intrinsic::smax:1631  case Intrinsic::smin:1632  case Intrinsic::umax:1633  case Intrinsic::umin:1634  case Intrinsic::scmp:1635  case Intrinsic::ucmp:1636  case Intrinsic::sadd_with_overflow:1637  case Intrinsic::uadd_with_overflow:1638  case Intrinsic::ssub_with_overflow:1639  case Intrinsic::usub_with_overflow:1640  case Intrinsic::smul_with_overflow:1641  case Intrinsic::umul_with_overflow:1642  case Intrinsic::sadd_sat:1643  case Intrinsic::uadd_sat:1644  case Intrinsic::ssub_sat:1645  case Intrinsic::usub_sat:1646  case Intrinsic::smul_fix:1647  case Intrinsic::smul_fix_sat:1648  case Intrinsic::bitreverse:1649  case Intrinsic::is_constant:1650  case Intrinsic::vector_reduce_add:1651  case Intrinsic::vector_reduce_mul:1652  case Intrinsic::vector_reduce_and:1653  case Intrinsic::vector_reduce_or:1654  case Intrinsic::vector_reduce_xor:1655  case Intrinsic::vector_reduce_smin:1656  case Intrinsic::vector_reduce_smax:1657  case Intrinsic::vector_reduce_umin:1658  case Intrinsic::vector_reduce_umax:1659  case Intrinsic::vector_extract:1660  case Intrinsic::vector_insert:1661  case Intrinsic::vector_interleave2:1662  case Intrinsic::vector_interleave3:1663  case Intrinsic::vector_interleave4:1664  case Intrinsic::vector_interleave5:1665  case Intrinsic::vector_interleave6:1666  case Intrinsic::vector_interleave7:1667  case Intrinsic::vector_interleave8:1668  case Intrinsic::vector_deinterleave2:1669  case Intrinsic::vector_deinterleave3:1670  case Intrinsic::vector_deinterleave4:1671  case Intrinsic::vector_deinterleave5:1672  case Intrinsic::vector_deinterleave6:1673  case Intrinsic::vector_deinterleave7:1674  case Intrinsic::vector_deinterleave8:1675  // Target intrinsics1676  case Intrinsic::amdgcn_perm:1677  case Intrinsic::amdgcn_wave_reduce_umin:1678  case Intrinsic::amdgcn_wave_reduce_umax:1679  case Intrinsic::amdgcn_wave_reduce_max:1680  case Intrinsic::amdgcn_wave_reduce_min:1681  case Intrinsic::amdgcn_wave_reduce_add:1682  case Intrinsic::amdgcn_wave_reduce_sub:1683  case Intrinsic::amdgcn_wave_reduce_and:1684  case Intrinsic::amdgcn_wave_reduce_or:1685  case Intrinsic::amdgcn_wave_reduce_xor:1686  case Intrinsic::amdgcn_s_wqm:1687  case Intrinsic::amdgcn_s_quadmask:1688  case Intrinsic::amdgcn_s_bitreplicate:1689  case Intrinsic::arm_mve_vctp8:1690  case Intrinsic::arm_mve_vctp16:1691  case Intrinsic::arm_mve_vctp32:1692  case Intrinsic::arm_mve_vctp64:1693  case Intrinsic::aarch64_sve_convert_from_svbool:1694  case Intrinsic::wasm_alltrue:1695  case Intrinsic::wasm_anytrue:1696  case Intrinsic::wasm_dot:1697  // WebAssembly float semantics are always known1698  case Intrinsic::wasm_trunc_signed:1699  case Intrinsic::wasm_trunc_unsigned:1700    return true;1701 1702  // Floating point operations cannot be folded in strictfp functions in1703  // general case. They can be folded if FP environment is known to compiler.1704  case Intrinsic::minnum:1705  case Intrinsic::maxnum:1706  case Intrinsic::minimum:1707  case Intrinsic::maximum:1708  case Intrinsic::minimumnum:1709  case Intrinsic::maximumnum:1710  case Intrinsic::log:1711  case Intrinsic::log2:1712  case Intrinsic::log10:1713  case Intrinsic::exp:1714  case Intrinsic::exp2:1715  case Intrinsic::exp10:1716  case Intrinsic::sqrt:1717  case Intrinsic::sin:1718  case Intrinsic::cos:1719  case Intrinsic::sincos:1720  case Intrinsic::sinh:1721  case Intrinsic::cosh:1722  case Intrinsic::atan:1723  case Intrinsic::pow:1724  case Intrinsic::powi:1725  case Intrinsic::ldexp:1726  case Intrinsic::fma:1727  case Intrinsic::fmuladd:1728  case Intrinsic::frexp:1729  case Intrinsic::fptoui_sat:1730  case Intrinsic::fptosi_sat:1731  case Intrinsic::convert_from_fp16:1732  case Intrinsic::convert_to_fp16:1733  case Intrinsic::amdgcn_cos:1734  case Intrinsic::amdgcn_cubeid:1735  case Intrinsic::amdgcn_cubema:1736  case Intrinsic::amdgcn_cubesc:1737  case Intrinsic::amdgcn_cubetc:1738  case Intrinsic::amdgcn_fmul_legacy:1739  case Intrinsic::amdgcn_fma_legacy:1740  case Intrinsic::amdgcn_fract:1741  case Intrinsic::amdgcn_sin:1742  // The intrinsics below depend on rounding mode in MXCSR.1743  case Intrinsic::x86_sse_cvtss2si:1744  case Intrinsic::x86_sse_cvtss2si64:1745  case Intrinsic::x86_sse_cvttss2si:1746  case Intrinsic::x86_sse_cvttss2si64:1747  case Intrinsic::x86_sse2_cvtsd2si:1748  case Intrinsic::x86_sse2_cvtsd2si64:1749  case Intrinsic::x86_sse2_cvttsd2si:1750  case Intrinsic::x86_sse2_cvttsd2si64:1751  case Intrinsic::x86_avx512_vcvtss2si32:1752  case Intrinsic::x86_avx512_vcvtss2si64:1753  case Intrinsic::x86_avx512_cvttss2si:1754  case Intrinsic::x86_avx512_cvttss2si64:1755  case Intrinsic::x86_avx512_vcvtsd2si32:1756  case Intrinsic::x86_avx512_vcvtsd2si64:1757  case Intrinsic::x86_avx512_cvttsd2si:1758  case Intrinsic::x86_avx512_cvttsd2si64:1759  case Intrinsic::x86_avx512_vcvtss2usi32:1760  case Intrinsic::x86_avx512_vcvtss2usi64:1761  case Intrinsic::x86_avx512_cvttss2usi:1762  case Intrinsic::x86_avx512_cvttss2usi64:1763  case Intrinsic::x86_avx512_vcvtsd2usi32:1764  case Intrinsic::x86_avx512_vcvtsd2usi64:1765  case Intrinsic::x86_avx512_cvttsd2usi:1766  case Intrinsic::x86_avx512_cvttsd2usi64:1767 1768  // NVVM FMax intrinsics1769  case Intrinsic::nvvm_fmax_d:1770  case Intrinsic::nvvm_fmax_f:1771  case Intrinsic::nvvm_fmax_ftz_f:1772  case Intrinsic::nvvm_fmax_ftz_nan_f:1773  case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:1774  case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:1775  case Intrinsic::nvvm_fmax_nan_f:1776  case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:1777  case Intrinsic::nvvm_fmax_xorsign_abs_f:1778 1779  // NVVM FMin intrinsics1780  case Intrinsic::nvvm_fmin_d:1781  case Intrinsic::nvvm_fmin_f:1782  case Intrinsic::nvvm_fmin_ftz_f:1783  case Intrinsic::nvvm_fmin_ftz_nan_f:1784  case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:1785  case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:1786  case Intrinsic::nvvm_fmin_nan_f:1787  case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:1788  case Intrinsic::nvvm_fmin_xorsign_abs_f:1789 1790  // NVVM float/double to int32/uint32 conversion intrinsics1791  case Intrinsic::nvvm_f2i_rm:1792  case Intrinsic::nvvm_f2i_rn:1793  case Intrinsic::nvvm_f2i_rp:1794  case Intrinsic::nvvm_f2i_rz:1795  case Intrinsic::nvvm_f2i_rm_ftz:1796  case Intrinsic::nvvm_f2i_rn_ftz:1797  case Intrinsic::nvvm_f2i_rp_ftz:1798  case Intrinsic::nvvm_f2i_rz_ftz:1799  case Intrinsic::nvvm_f2ui_rm:1800  case Intrinsic::nvvm_f2ui_rn:1801  case Intrinsic::nvvm_f2ui_rp:1802  case Intrinsic::nvvm_f2ui_rz:1803  case Intrinsic::nvvm_f2ui_rm_ftz:1804  case Intrinsic::nvvm_f2ui_rn_ftz:1805  case Intrinsic::nvvm_f2ui_rp_ftz:1806  case Intrinsic::nvvm_f2ui_rz_ftz:1807  case Intrinsic::nvvm_d2i_rm:1808  case Intrinsic::nvvm_d2i_rn:1809  case Intrinsic::nvvm_d2i_rp:1810  case Intrinsic::nvvm_d2i_rz:1811  case Intrinsic::nvvm_d2ui_rm:1812  case Intrinsic::nvvm_d2ui_rn:1813  case Intrinsic::nvvm_d2ui_rp:1814  case Intrinsic::nvvm_d2ui_rz:1815 1816  // NVVM float/double to int64/uint64 conversion intrinsics1817  case Intrinsic::nvvm_f2ll_rm:1818  case Intrinsic::nvvm_f2ll_rn:1819  case Intrinsic::nvvm_f2ll_rp:1820  case Intrinsic::nvvm_f2ll_rz:1821  case Intrinsic::nvvm_f2ll_rm_ftz:1822  case Intrinsic::nvvm_f2ll_rn_ftz:1823  case Intrinsic::nvvm_f2ll_rp_ftz:1824  case Intrinsic::nvvm_f2ll_rz_ftz:1825  case Intrinsic::nvvm_f2ull_rm:1826  case Intrinsic::nvvm_f2ull_rn:1827  case Intrinsic::nvvm_f2ull_rp:1828  case Intrinsic::nvvm_f2ull_rz:1829  case Intrinsic::nvvm_f2ull_rm_ftz:1830  case Intrinsic::nvvm_f2ull_rn_ftz:1831  case Intrinsic::nvvm_f2ull_rp_ftz:1832  case Intrinsic::nvvm_f2ull_rz_ftz:1833  case Intrinsic::nvvm_d2ll_rm:1834  case Intrinsic::nvvm_d2ll_rn:1835  case Intrinsic::nvvm_d2ll_rp:1836  case Intrinsic::nvvm_d2ll_rz:1837  case Intrinsic::nvvm_d2ull_rm:1838  case Intrinsic::nvvm_d2ull_rn:1839  case Intrinsic::nvvm_d2ull_rp:1840  case Intrinsic::nvvm_d2ull_rz:1841 1842  // NVVM math intrinsics:1843  case Intrinsic::nvvm_ceil_d:1844  case Intrinsic::nvvm_ceil_f:1845  case Intrinsic::nvvm_ceil_ftz_f:1846 1847  case Intrinsic::nvvm_fabs:1848  case Intrinsic::nvvm_fabs_ftz:1849 1850  case Intrinsic::nvvm_floor_d:1851  case Intrinsic::nvvm_floor_f:1852  case Intrinsic::nvvm_floor_ftz_f:1853 1854  case Intrinsic::nvvm_rcp_rm_d:1855  case Intrinsic::nvvm_rcp_rm_f:1856  case Intrinsic::nvvm_rcp_rm_ftz_f:1857  case Intrinsic::nvvm_rcp_rn_d:1858  case Intrinsic::nvvm_rcp_rn_f:1859  case Intrinsic::nvvm_rcp_rn_ftz_f:1860  case Intrinsic::nvvm_rcp_rp_d:1861  case Intrinsic::nvvm_rcp_rp_f:1862  case Intrinsic::nvvm_rcp_rp_ftz_f:1863  case Intrinsic::nvvm_rcp_rz_d:1864  case Intrinsic::nvvm_rcp_rz_f:1865  case Intrinsic::nvvm_rcp_rz_ftz_f:1866 1867  case Intrinsic::nvvm_round_d:1868  case Intrinsic::nvvm_round_f:1869  case Intrinsic::nvvm_round_ftz_f:1870 1871  case Intrinsic::nvvm_saturate_d:1872  case Intrinsic::nvvm_saturate_f:1873  case Intrinsic::nvvm_saturate_ftz_f:1874 1875  case Intrinsic::nvvm_sqrt_f:1876  case Intrinsic::nvvm_sqrt_rn_d:1877  case Intrinsic::nvvm_sqrt_rn_f:1878  case Intrinsic::nvvm_sqrt_rn_ftz_f:1879    return !Call->isStrictFP();1880 1881  // NVVM add intrinsics with explicit rounding modes1882  case Intrinsic::nvvm_add_rm_d:1883  case Intrinsic::nvvm_add_rn_d:1884  case Intrinsic::nvvm_add_rp_d:1885  case Intrinsic::nvvm_add_rz_d:1886  case Intrinsic::nvvm_add_rm_f:1887  case Intrinsic::nvvm_add_rn_f:1888  case Intrinsic::nvvm_add_rp_f:1889  case Intrinsic::nvvm_add_rz_f:1890  case Intrinsic::nvvm_add_rm_ftz_f:1891  case Intrinsic::nvvm_add_rn_ftz_f:1892  case Intrinsic::nvvm_add_rp_ftz_f:1893  case Intrinsic::nvvm_add_rz_ftz_f:1894 1895  // NVVM div intrinsics with explicit rounding modes1896  case Intrinsic::nvvm_div_rm_d:1897  case Intrinsic::nvvm_div_rn_d:1898  case Intrinsic::nvvm_div_rp_d:1899  case Intrinsic::nvvm_div_rz_d:1900  case Intrinsic::nvvm_div_rm_f:1901  case Intrinsic::nvvm_div_rn_f:1902  case Intrinsic::nvvm_div_rp_f:1903  case Intrinsic::nvvm_div_rz_f:1904  case Intrinsic::nvvm_div_rm_ftz_f:1905  case Intrinsic::nvvm_div_rn_ftz_f:1906  case Intrinsic::nvvm_div_rp_ftz_f:1907  case Intrinsic::nvvm_div_rz_ftz_f:1908 1909  // NVVM mul intrinsics with explicit rounding modes1910  case Intrinsic::nvvm_mul_rm_d:1911  case Intrinsic::nvvm_mul_rn_d:1912  case Intrinsic::nvvm_mul_rp_d:1913  case Intrinsic::nvvm_mul_rz_d:1914  case Intrinsic::nvvm_mul_rm_f:1915  case Intrinsic::nvvm_mul_rn_f:1916  case Intrinsic::nvvm_mul_rp_f:1917  case Intrinsic::nvvm_mul_rz_f:1918  case Intrinsic::nvvm_mul_rm_ftz_f:1919  case Intrinsic::nvvm_mul_rn_ftz_f:1920  case Intrinsic::nvvm_mul_rp_ftz_f:1921  case Intrinsic::nvvm_mul_rz_ftz_f:1922 1923  // NVVM fma intrinsics with explicit rounding modes1924  case Intrinsic::nvvm_fma_rm_d:1925  case Intrinsic::nvvm_fma_rn_d:1926  case Intrinsic::nvvm_fma_rp_d:1927  case Intrinsic::nvvm_fma_rz_d:1928  case Intrinsic::nvvm_fma_rm_f:1929  case Intrinsic::nvvm_fma_rn_f:1930  case Intrinsic::nvvm_fma_rp_f:1931  case Intrinsic::nvvm_fma_rz_f:1932  case Intrinsic::nvvm_fma_rm_ftz_f:1933  case Intrinsic::nvvm_fma_rn_ftz_f:1934  case Intrinsic::nvvm_fma_rp_ftz_f:1935  case Intrinsic::nvvm_fma_rz_ftz_f:1936 1937  // Sign operations are actually bitwise operations, they do not raise1938  // exceptions even for SNANs.1939  case Intrinsic::fabs:1940  case Intrinsic::copysign:1941  case Intrinsic::is_fpclass:1942  // Non-constrained variants of rounding operations means default FP1943  // environment, they can be folded in any case.1944  case Intrinsic::ceil:1945  case Intrinsic::floor:1946  case Intrinsic::round:1947  case Intrinsic::roundeven:1948  case Intrinsic::trunc:1949  case Intrinsic::nearbyint:1950  case Intrinsic::rint:1951  case Intrinsic::canonicalize:1952 1953  // Constrained intrinsics can be folded if FP environment is known1954  // to compiler.1955  case Intrinsic::experimental_constrained_fma:1956  case Intrinsic::experimental_constrained_fmuladd:1957  case Intrinsic::experimental_constrained_fadd:1958  case Intrinsic::experimental_constrained_fsub:1959  case Intrinsic::experimental_constrained_fmul:1960  case Intrinsic::experimental_constrained_fdiv:1961  case Intrinsic::experimental_constrained_frem:1962  case Intrinsic::experimental_constrained_ceil:1963  case Intrinsic::experimental_constrained_floor:1964  case Intrinsic::experimental_constrained_round:1965  case Intrinsic::experimental_constrained_roundeven:1966  case Intrinsic::experimental_constrained_trunc:1967  case Intrinsic::experimental_constrained_nearbyint:1968  case Intrinsic::experimental_constrained_rint:1969  case Intrinsic::experimental_constrained_fcmp:1970  case Intrinsic::experimental_constrained_fcmps:1971    return true;1972  default:1973    return false;1974  case Intrinsic::not_intrinsic: break;1975  }1976 1977  if (!F->hasName() || Call->isStrictFP())1978    return false;1979 1980  // In these cases, the check of the length is required.  We don't want to1981  // return true for a name like "cos\0blah" which strcmp would return equal to1982  // "cos", but has length 8.1983  StringRef Name = F->getName();1984  switch (Name[0]) {1985  default:1986    return false;1987  case 'a':1988    return Name == "acos" || Name == "acosf" ||1989           Name == "asin" || Name == "asinf" ||1990           Name == "atan" || Name == "atanf" ||1991           Name == "atan2" || Name == "atan2f";1992  case 'c':1993    return Name == "ceil" || Name == "ceilf" ||1994           Name == "cos" || Name == "cosf" ||1995           Name == "cosh" || Name == "coshf";1996  case 'e':1997    return Name == "exp" || Name == "expf" || Name == "exp2" ||1998           Name == "exp2f" || Name == "erf" || Name == "erff";1999  case 'f':2000    return Name == "fabs" || Name == "fabsf" ||2001           Name == "floor" || Name == "floorf" ||2002           Name == "fmod" || Name == "fmodf";2003  case 'i':2004    return Name == "ilogb" || Name == "ilogbf";2005  case 'l':2006    return Name == "log" || Name == "logf" || Name == "logl" ||2007           Name == "log2" || Name == "log2f" || Name == "log10" ||2008           Name == "log10f" || Name == "logb" || Name == "logbf" ||2009           Name == "log1p" || Name == "log1pf";2010  case 'n':2011    return Name == "nearbyint" || Name == "nearbyintf";2012  case 'p':2013    return Name == "pow" || Name == "powf";2014  case 'r':2015    return Name == "remainder" || Name == "remainderf" ||2016           Name == "rint" || Name == "rintf" ||2017           Name == "round" || Name == "roundf";2018  case 's':2019    return Name == "sin" || Name == "sinf" ||2020           Name == "sinh" || Name == "sinhf" ||2021           Name == "sqrt" || Name == "sqrtf";2022  case 't':2023    return Name == "tan" || Name == "tanf" ||2024           Name == "tanh" || Name == "tanhf" ||2025           Name == "trunc" || Name == "truncf";2026  case '_':2027    // Check for various function names that get used for the math functions2028    // when the header files are preprocessed with the macro2029    // __FINITE_MATH_ONLY__ enabled.2030    // The '12' here is the length of the shortest name that can match.2031    // We need to check the size before looking at Name[1] and Name[2]2032    // so we may as well check a limit that will eliminate mismatches.2033    if (Name.size() < 12 || Name[1] != '_')2034      return false;2035    switch (Name[2]) {2036    default:2037      return false;2038    case 'a':2039      return Name == "__acos_finite" || Name == "__acosf_finite" ||2040             Name == "__asin_finite" || Name == "__asinf_finite" ||2041             Name == "__atan2_finite" || Name == "__atan2f_finite";2042    case 'c':2043      return Name == "__cosh_finite" || Name == "__coshf_finite";2044    case 'e':2045      return Name == "__exp_finite" || Name == "__expf_finite" ||2046             Name == "__exp2_finite" || Name == "__exp2f_finite";2047    case 'l':2048      return Name == "__log_finite" || Name == "__logf_finite" ||2049             Name == "__log10_finite" || Name == "__log10f_finite";2050    case 'p':2051      return Name == "__pow_finite" || Name == "__powf_finite";2052    case 's':2053      return Name == "__sinh_finite" || Name == "__sinhf_finite";2054    }2055  }2056}2057 2058namespace {2059 2060Constant *GetConstantFoldFPValue(double V, Type *Ty) {2061  if (Ty->isHalfTy() || Ty->isFloatTy()) {2062    APFloat APF(V);2063    bool unused;2064    APF.convert(Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &unused);2065    return ConstantFP::get(Ty->getContext(), APF);2066  }2067  if (Ty->isDoubleTy())2068    return ConstantFP::get(Ty->getContext(), APFloat(V));2069  llvm_unreachable("Can only constant fold half/float/double");2070}2071 2072#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)2073Constant *GetConstantFoldFPValue128(float128 V, Type *Ty) {2074  if (Ty->isFP128Ty())2075    return ConstantFP::get(Ty, V);2076  llvm_unreachable("Can only constant fold fp128");2077}2078#endif2079 2080/// Clear the floating-point exception state.2081inline void llvm_fenv_clearexcept() {2082#if HAVE_DECL_FE_ALL_EXCEPT2083  feclearexcept(FE_ALL_EXCEPT);2084#endif2085  errno = 0;2086}2087 2088/// Test if a floating-point exception was raised.2089inline bool llvm_fenv_testexcept() {2090  int errno_val = errno;2091  if (errno_val == ERANGE || errno_val == EDOM)2092    return true;2093#if HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT2094  if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))2095    return true;2096#endif2097  return false;2098}2099 2100static APFloat FTZPreserveSign(const APFloat &V) {2101  if (V.isDenormal())2102    return APFloat::getZero(V.getSemantics(), V.isNegative());2103  return V;2104}2105 2106static APFloat FlushToPositiveZero(const APFloat &V) {2107  if (V.isDenormal())2108    return APFloat::getZero(V.getSemantics(), false);2109  return V;2110}2111 2112static APFloat FlushWithDenormKind(const APFloat &V,2113                                   DenormalMode::DenormalModeKind DenormKind) {2114  assert(DenormKind != DenormalMode::DenormalModeKind::Invalid &&2115         DenormKind != DenormalMode::DenormalModeKind::Dynamic);2116  switch (DenormKind) {2117  case DenormalMode::DenormalModeKind::IEEE:2118    return V;2119  case DenormalMode::DenormalModeKind::PreserveSign:2120    return FTZPreserveSign(V);2121  case DenormalMode::DenormalModeKind::PositiveZero:2122    return FlushToPositiveZero(V);2123  default:2124    llvm_unreachable("Invalid denormal mode!");2125  }2126}2127 2128Constant *ConstantFoldFP(double (*NativeFP)(double), const APFloat &V, Type *Ty,2129                         DenormalMode DenormMode = DenormalMode::getIEEE()) {2130  if (!DenormMode.isValid() ||2131      DenormMode.Input == DenormalMode::DenormalModeKind::Dynamic ||2132      DenormMode.Output == DenormalMode::DenormalModeKind::Dynamic)2133    return nullptr;2134 2135  llvm_fenv_clearexcept();2136  auto Input = FlushWithDenormKind(V, DenormMode.Input);2137  double Result = NativeFP(Input.convertToDouble());2138  if (llvm_fenv_testexcept()) {2139    llvm_fenv_clearexcept();2140    return nullptr;2141  }2142 2143  Constant *Output = GetConstantFoldFPValue(Result, Ty);2144  if (DenormMode.Output == DenormalMode::DenormalModeKind::IEEE)2145    return Output;2146  const auto *CFP = static_cast<ConstantFP *>(Output);2147  const auto Res = FlushWithDenormKind(CFP->getValueAPF(), DenormMode.Output);2148  return ConstantFP::get(Ty->getContext(), Res);2149}2150 2151#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)2152Constant *ConstantFoldFP128(float128 (*NativeFP)(float128), const APFloat &V,2153                            Type *Ty) {2154  llvm_fenv_clearexcept();2155  float128 Result = NativeFP(V.convertToQuad());2156  if (llvm_fenv_testexcept()) {2157    llvm_fenv_clearexcept();2158    return nullptr;2159  }2160 2161  return GetConstantFoldFPValue128(Result, Ty);2162}2163#endif2164 2165Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),2166                               const APFloat &V, const APFloat &W, Type *Ty) {2167  llvm_fenv_clearexcept();2168  double Result = NativeFP(V.convertToDouble(), W.convertToDouble());2169  if (llvm_fenv_testexcept()) {2170    llvm_fenv_clearexcept();2171    return nullptr;2172  }2173 2174  return GetConstantFoldFPValue(Result, Ty);2175}2176 2177Constant *constantFoldVectorReduce(Intrinsic::ID IID, Constant *Op) {2178  auto *OpVT = cast<VectorType>(Op->getType());2179 2180  // This is the same as the underlying binops - poison propagates.2181  if (Op->containsPoisonElement())2182    return PoisonValue::get(OpVT->getElementType());2183 2184  // Shortcut non-accumulating reductions.2185  if (Constant *SplatVal = Op->getSplatValue()) {2186    switch (IID) {2187    case Intrinsic::vector_reduce_and:2188    case Intrinsic::vector_reduce_or:2189    case Intrinsic::vector_reduce_smin:2190    case Intrinsic::vector_reduce_smax:2191    case Intrinsic::vector_reduce_umin:2192    case Intrinsic::vector_reduce_umax:2193      return SplatVal;2194    case Intrinsic::vector_reduce_add:2195      if (SplatVal->isNullValue())2196        return SplatVal;2197      break;2198    case Intrinsic::vector_reduce_mul:2199      if (SplatVal->isNullValue() || SplatVal->isOneValue())2200        return SplatVal;2201      break;2202    case Intrinsic::vector_reduce_xor:2203      if (SplatVal->isNullValue())2204        return SplatVal;2205      if (OpVT->getElementCount().isKnownMultipleOf(2))2206        return Constant::getNullValue(OpVT->getElementType());2207      break;2208    }2209  }2210 2211  FixedVectorType *VT = dyn_cast<FixedVectorType>(OpVT);2212  if (!VT)2213    return nullptr;2214 2215  // TODO: Handle undef.2216  auto *EltC = dyn_cast_or_null<ConstantInt>(Op->getAggregateElement(0U));2217  if (!EltC)2218    return nullptr;2219 2220  APInt Acc = EltC->getValue();2221  for (unsigned I = 1, E = VT->getNumElements(); I != E; I++) {2222    if (!(EltC = dyn_cast_or_null<ConstantInt>(Op->getAggregateElement(I))))2223      return nullptr;2224    const APInt &X = EltC->getValue();2225    switch (IID) {2226    case Intrinsic::vector_reduce_add:2227      Acc = Acc + X;2228      break;2229    case Intrinsic::vector_reduce_mul:2230      Acc = Acc * X;2231      break;2232    case Intrinsic::vector_reduce_and:2233      Acc = Acc & X;2234      break;2235    case Intrinsic::vector_reduce_or:2236      Acc = Acc | X;2237      break;2238    case Intrinsic::vector_reduce_xor:2239      Acc = Acc ^ X;2240      break;2241    case Intrinsic::vector_reduce_smin:2242      Acc = APIntOps::smin(Acc, X);2243      break;2244    case Intrinsic::vector_reduce_smax:2245      Acc = APIntOps::smax(Acc, X);2246      break;2247    case Intrinsic::vector_reduce_umin:2248      Acc = APIntOps::umin(Acc, X);2249      break;2250    case Intrinsic::vector_reduce_umax:2251      Acc = APIntOps::umax(Acc, X);2252      break;2253    }2254  }2255 2256  return ConstantInt::get(Op->getContext(), Acc);2257}2258 2259/// Attempt to fold an SSE floating point to integer conversion of a constant2260/// floating point. If roundTowardZero is false, the default IEEE rounding is2261/// used (toward nearest, ties to even). This matches the behavior of the2262/// non-truncating SSE instructions in the default rounding mode. The desired2263/// integer type Ty is used to select how many bits are available for the2264/// result. Returns null if the conversion cannot be performed, otherwise2265/// returns the Constant value resulting from the conversion.2266Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,2267                                      Type *Ty, bool IsSigned) {2268  // All of these conversion intrinsics form an integer of at most 64bits.2269  unsigned ResultWidth = Ty->getIntegerBitWidth();2270  assert(ResultWidth <= 64 &&2271         "Can only constant fold conversions to 64 and 32 bit ints");2272 2273  uint64_t UIntVal;2274  bool isExact = false;2275  APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero2276                                              : APFloat::rmNearestTiesToEven;2277  APFloat::opStatus status =2278      Val.convertToInteger(MutableArrayRef(UIntVal), ResultWidth,2279                           IsSigned, mode, &isExact);2280  if (status != APFloat::opOK &&2281      (!roundTowardZero || status != APFloat::opInexact))2282    return nullptr;2283  return ConstantInt::get(Ty, UIntVal, IsSigned);2284}2285 2286double getValueAsDouble(ConstantFP *Op) {2287  Type *Ty = Op->getType();2288 2289  if (Ty->isBFloatTy() || Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())2290    return Op->getValueAPF().convertToDouble();2291 2292  bool unused;2293  APFloat APF = Op->getValueAPF();2294  APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused);2295  return APF.convertToDouble();2296}2297 2298static bool getConstIntOrUndef(Value *Op, const APInt *&C) {2299  if (auto *CI = dyn_cast<ConstantInt>(Op)) {2300    C = &CI->getValue();2301    return true;2302  }2303  if (isa<UndefValue>(Op)) {2304    C = nullptr;2305    return true;2306  }2307  return false;2308}2309 2310/// Checks if the given intrinsic call, which evaluates to constant, is allowed2311/// to be folded.2312///2313/// \param CI Constrained intrinsic call.2314/// \param St Exception flags raised during constant evaluation.2315static bool mayFoldConstrained(ConstrainedFPIntrinsic *CI,2316                               APFloat::opStatus St) {2317  std::optional<RoundingMode> ORM = CI->getRoundingMode();2318  std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();2319 2320  // If the operation does not change exception status flags, it is safe2321  // to fold.2322  if (St == APFloat::opStatus::opOK)2323    return true;2324 2325  // If evaluation raised FP exception, the result can depend on rounding2326  // mode. If the latter is unknown, folding is not possible.2327  if (ORM == RoundingMode::Dynamic)2328    return false;2329 2330  // If FP exceptions are ignored, fold the call, even if such exception is2331  // raised.2332  if (EB && *EB != fp::ExceptionBehavior::ebStrict)2333    return true;2334 2335  // Leave the calculation for runtime so that exception flags be correctly set2336  // in hardware.2337  return false;2338}2339 2340/// Returns the rounding mode that should be used for constant evaluation.2341static RoundingMode2342getEvaluationRoundingMode(const ConstrainedFPIntrinsic *CI) {2343  std::optional<RoundingMode> ORM = CI->getRoundingMode();2344  if (!ORM || *ORM == RoundingMode::Dynamic)2345    // Even if the rounding mode is unknown, try evaluating the operation.2346    // If it does not raise inexact exception, rounding was not applied,2347    // so the result is exact and does not depend on rounding mode. Whether2348    // other FP exceptions are raised, it does not depend on rounding mode.2349    return RoundingMode::NearestTiesToEven;2350  return *ORM;2351}2352 2353/// Try to constant fold llvm.canonicalize for the given caller and value.2354static Constant *constantFoldCanonicalize(const Type *Ty, const CallBase *CI,2355                                          const APFloat &Src) {2356  // Zero, positive and negative, is always OK to fold.2357  if (Src.isZero()) {2358    // Get a fresh 0, since ppc_fp128 does have non-canonical zeros.2359    return ConstantFP::get(2360        CI->getContext(),2361        APFloat::getZero(Src.getSemantics(), Src.isNegative()));2362  }2363 2364  if (!Ty->isIEEELikeFPTy())2365    return nullptr;2366 2367  // Zero is always canonical and the sign must be preserved.2368  //2369  // Denorms and nans may have special encodings, but it should be OK to fold a2370  // totally average number.2371  if (Src.isNormal() || Src.isInfinity())2372    return ConstantFP::get(CI->getContext(), Src);2373 2374  if (Src.isDenormal() && CI->getParent() && CI->getFunction()) {2375    DenormalMode DenormMode =2376        CI->getFunction()->getDenormalMode(Src.getSemantics());2377 2378    if (DenormMode == DenormalMode::getIEEE())2379      return ConstantFP::get(CI->getContext(), Src);2380 2381    if (DenormMode.Input == DenormalMode::Dynamic)2382      return nullptr;2383 2384    // If we know if either input or output is flushed, we can fold.2385    if ((DenormMode.Input == DenormalMode::Dynamic &&2386         DenormMode.Output == DenormalMode::IEEE) ||2387        (DenormMode.Input == DenormalMode::IEEE &&2388         DenormMode.Output == DenormalMode::Dynamic))2389      return nullptr;2390 2391    bool IsPositive =2392        (!Src.isNegative() || DenormMode.Input == DenormalMode::PositiveZero ||2393         (DenormMode.Output == DenormalMode::PositiveZero &&2394          DenormMode.Input == DenormalMode::IEEE));2395 2396    return ConstantFP::get(CI->getContext(),2397                           APFloat::getZero(Src.getSemantics(), !IsPositive));2398  }2399 2400  return nullptr;2401}2402 2403static Constant *ConstantFoldScalarCall1(StringRef Name,2404                                         Intrinsic::ID IntrinsicID,2405                                         Type *Ty,2406                                         ArrayRef<Constant *> Operands,2407                                         const TargetLibraryInfo *TLI,2408                                         const CallBase *Call) {2409  assert(Operands.size() == 1 && "Wrong number of operands.");2410 2411  if (IntrinsicID == Intrinsic::is_constant) {2412    // We know we have a "Constant" argument. But we want to only2413    // return true for manifest constants, not those that depend on2414    // constants with unknowable values, e.g. GlobalValue or BlockAddress.2415    if (Operands[0]->isManifestConstant())2416      return ConstantInt::getTrue(Ty->getContext());2417    return nullptr;2418  }2419 2420  if (isa<UndefValue>(Operands[0])) {2421    // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN.2422    // ctpop() is between 0 and bitwidth, pick 0 for undef.2423    // fptoui.sat and fptosi.sat can always fold to zero (for a zero input).2424    if (IntrinsicID == Intrinsic::cos ||2425        IntrinsicID == Intrinsic::ctpop ||2426        IntrinsicID == Intrinsic::fptoui_sat ||2427        IntrinsicID == Intrinsic::fptosi_sat ||2428        IntrinsicID == Intrinsic::canonicalize)2429      return Constant::getNullValue(Ty);2430    if (IntrinsicID == Intrinsic::bswap ||2431        IntrinsicID == Intrinsic::bitreverse ||2432        IntrinsicID == Intrinsic::launder_invariant_group ||2433        IntrinsicID == Intrinsic::strip_invariant_group)2434      return Operands[0];2435  }2436 2437  if (isa<ConstantPointerNull>(Operands[0])) {2438    // launder(null) == null == strip(null) iff in addrspace 02439    if (IntrinsicID == Intrinsic::launder_invariant_group ||2440        IntrinsicID == Intrinsic::strip_invariant_group) {2441      // If instruction is not yet put in a basic block (e.g. when cloning2442      // a function during inlining), Call's caller may not be available.2443      // So check Call's BB first before querying Call->getCaller.2444      const Function *Caller =2445          Call->getParent() ? Call->getCaller() : nullptr;2446      if (Caller &&2447          !NullPointerIsDefined(2448              Caller, Operands[0]->getType()->getPointerAddressSpace())) {2449        return Operands[0];2450      }2451      return nullptr;2452    }2453  }2454 2455  if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) {2456    if (IntrinsicID == Intrinsic::convert_to_fp16) {2457      APFloat Val(Op->getValueAPF());2458 2459      bool lost = false;2460      Val.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &lost);2461 2462      return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());2463    }2464 2465    APFloat U = Op->getValueAPF();2466 2467    if (IntrinsicID == Intrinsic::wasm_trunc_signed ||2468        IntrinsicID == Intrinsic::wasm_trunc_unsigned) {2469      bool Signed = IntrinsicID == Intrinsic::wasm_trunc_signed;2470 2471      if (U.isNaN())2472        return nullptr;2473 2474      unsigned Width = Ty->getIntegerBitWidth();2475      APSInt Int(Width, !Signed);2476      bool IsExact = false;2477      APFloat::opStatus Status =2478          U.convertToInteger(Int, APFloat::rmTowardZero, &IsExact);2479 2480      if (Status == APFloat::opOK || Status == APFloat::opInexact)2481        return ConstantInt::get(Ty, Int);2482 2483      return nullptr;2484    }2485 2486    if (IntrinsicID == Intrinsic::fptoui_sat ||2487        IntrinsicID == Intrinsic::fptosi_sat) {2488      // convertToInteger() already has the desired saturation semantics.2489      APSInt Int(Ty->getIntegerBitWidth(),2490                 IntrinsicID == Intrinsic::fptoui_sat);2491      bool IsExact;2492      U.convertToInteger(Int, APFloat::rmTowardZero, &IsExact);2493      return ConstantInt::get(Ty, Int);2494    }2495 2496    if (IntrinsicID == Intrinsic::canonicalize)2497      return constantFoldCanonicalize(Ty, Call, U);2498 2499#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)2500    if (Ty->isFP128Ty()) {2501      if (IntrinsicID == Intrinsic::log) {2502        float128 Result = logf128(Op->getValueAPF().convertToQuad());2503        return GetConstantFoldFPValue128(Result, Ty);2504      }2505 2506      LibFunc Fp128Func = NotLibFunc;2507      if (TLI && TLI->getLibFunc(Name, Fp128Func) && TLI->has(Fp128Func) &&2508          Fp128Func == LibFunc_logl)2509        return ConstantFoldFP128(logf128, Op->getValueAPF(), Ty);2510    }2511#endif2512 2513    if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy() &&2514        !Ty->isIntegerTy())2515      return nullptr;2516 2517    // Use internal versions of these intrinsics.2518 2519    if (IntrinsicID == Intrinsic::nearbyint || IntrinsicID == Intrinsic::rint) {2520      U.roundToIntegral(APFloat::rmNearestTiesToEven);2521      return ConstantFP::get(Ty->getContext(), U);2522    }2523 2524    if (IntrinsicID == Intrinsic::round) {2525      U.roundToIntegral(APFloat::rmNearestTiesToAway);2526      return ConstantFP::get(Ty->getContext(), U);2527    }2528 2529    if (IntrinsicID == Intrinsic::roundeven) {2530      U.roundToIntegral(APFloat::rmNearestTiesToEven);2531      return ConstantFP::get(Ty->getContext(), U);2532    }2533 2534    if (IntrinsicID == Intrinsic::ceil) {2535      U.roundToIntegral(APFloat::rmTowardPositive);2536      return ConstantFP::get(Ty->getContext(), U);2537    }2538 2539    if (IntrinsicID == Intrinsic::floor) {2540      U.roundToIntegral(APFloat::rmTowardNegative);2541      return ConstantFP::get(Ty->getContext(), U);2542    }2543 2544    if (IntrinsicID == Intrinsic::trunc) {2545      U.roundToIntegral(APFloat::rmTowardZero);2546      return ConstantFP::get(Ty->getContext(), U);2547    }2548 2549    if (IntrinsicID == Intrinsic::fabs) {2550      U.clearSign();2551      return ConstantFP::get(Ty->getContext(), U);2552    }2553 2554    if (IntrinsicID == Intrinsic::amdgcn_fract) {2555      // The v_fract instruction behaves like the OpenCL spec, which defines2556      // fract(x) as fmin(x - floor(x), 0x1.fffffep-1f): "The min() operator is2557      //   there to prevent fract(-small) from returning 1.0. It returns the2558      //   largest positive floating-point number less than 1.0."2559      APFloat FloorU(U);2560      FloorU.roundToIntegral(APFloat::rmTowardNegative);2561      APFloat FractU(U - FloorU);2562      APFloat AlmostOne(U.getSemantics(), 1);2563      AlmostOne.next(/*nextDown*/ true);2564      return ConstantFP::get(Ty->getContext(), minimum(FractU, AlmostOne));2565    }2566 2567    // Rounding operations (floor, trunc, ceil, round and nearbyint) do not2568    // raise FP exceptions, unless the argument is signaling NaN.2569 2570    std::optional<APFloat::roundingMode> RM;2571    switch (IntrinsicID) {2572    default:2573      break;2574    case Intrinsic::experimental_constrained_nearbyint:2575    case Intrinsic::experimental_constrained_rint: {2576      auto CI = cast<ConstrainedFPIntrinsic>(Call);2577      RM = CI->getRoundingMode();2578      if (!RM || *RM == RoundingMode::Dynamic)2579        return nullptr;2580      break;2581    }2582    case Intrinsic::experimental_constrained_round:2583      RM = APFloat::rmNearestTiesToAway;2584      break;2585    case Intrinsic::experimental_constrained_ceil:2586      RM = APFloat::rmTowardPositive;2587      break;2588    case Intrinsic::experimental_constrained_floor:2589      RM = APFloat::rmTowardNegative;2590      break;2591    case Intrinsic::experimental_constrained_trunc:2592      RM = APFloat::rmTowardZero;2593      break;2594    }2595    if (RM) {2596      auto CI = cast<ConstrainedFPIntrinsic>(Call);2597      if (U.isFinite()) {2598        APFloat::opStatus St = U.roundToIntegral(*RM);2599        if (IntrinsicID == Intrinsic::experimental_constrained_rint &&2600            St == APFloat::opInexact) {2601          std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();2602          if (EB == fp::ebStrict)2603            return nullptr;2604        }2605      } else if (U.isSignaling()) {2606        std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();2607        if (EB && *EB != fp::ebIgnore)2608          return nullptr;2609        U = APFloat::getQNaN(U.getSemantics());2610      }2611      return ConstantFP::get(Ty->getContext(), U);2612    }2613 2614    // NVVM float/double to signed/unsigned int32/int64 conversions:2615    switch (IntrinsicID) {2616    // f2i2617    case Intrinsic::nvvm_f2i_rm:2618    case Intrinsic::nvvm_f2i_rn:2619    case Intrinsic::nvvm_f2i_rp:2620    case Intrinsic::nvvm_f2i_rz:2621    case Intrinsic::nvvm_f2i_rm_ftz:2622    case Intrinsic::nvvm_f2i_rn_ftz:2623    case Intrinsic::nvvm_f2i_rp_ftz:2624    case Intrinsic::nvvm_f2i_rz_ftz:2625    // f2ui2626    case Intrinsic::nvvm_f2ui_rm:2627    case Intrinsic::nvvm_f2ui_rn:2628    case Intrinsic::nvvm_f2ui_rp:2629    case Intrinsic::nvvm_f2ui_rz:2630    case Intrinsic::nvvm_f2ui_rm_ftz:2631    case Intrinsic::nvvm_f2ui_rn_ftz:2632    case Intrinsic::nvvm_f2ui_rp_ftz:2633    case Intrinsic::nvvm_f2ui_rz_ftz:2634    // d2i2635    case Intrinsic::nvvm_d2i_rm:2636    case Intrinsic::nvvm_d2i_rn:2637    case Intrinsic::nvvm_d2i_rp:2638    case Intrinsic::nvvm_d2i_rz:2639    // d2ui2640    case Intrinsic::nvvm_d2ui_rm:2641    case Intrinsic::nvvm_d2ui_rn:2642    case Intrinsic::nvvm_d2ui_rp:2643    case Intrinsic::nvvm_d2ui_rz:2644    // f2ll2645    case Intrinsic::nvvm_f2ll_rm:2646    case Intrinsic::nvvm_f2ll_rn:2647    case Intrinsic::nvvm_f2ll_rp:2648    case Intrinsic::nvvm_f2ll_rz:2649    case Intrinsic::nvvm_f2ll_rm_ftz:2650    case Intrinsic::nvvm_f2ll_rn_ftz:2651    case Intrinsic::nvvm_f2ll_rp_ftz:2652    case Intrinsic::nvvm_f2ll_rz_ftz:2653    // f2ull2654    case Intrinsic::nvvm_f2ull_rm:2655    case Intrinsic::nvvm_f2ull_rn:2656    case Intrinsic::nvvm_f2ull_rp:2657    case Intrinsic::nvvm_f2ull_rz:2658    case Intrinsic::nvvm_f2ull_rm_ftz:2659    case Intrinsic::nvvm_f2ull_rn_ftz:2660    case Intrinsic::nvvm_f2ull_rp_ftz:2661    case Intrinsic::nvvm_f2ull_rz_ftz:2662    // d2ll2663    case Intrinsic::nvvm_d2ll_rm:2664    case Intrinsic::nvvm_d2ll_rn:2665    case Intrinsic::nvvm_d2ll_rp:2666    case Intrinsic::nvvm_d2ll_rz:2667    // d2ull2668    case Intrinsic::nvvm_d2ull_rm:2669    case Intrinsic::nvvm_d2ull_rn:2670    case Intrinsic::nvvm_d2ull_rp:2671    case Intrinsic::nvvm_d2ull_rz: {2672      // In float-to-integer conversion, NaN inputs are converted to 0.2673      if (U.isNaN()) {2674        // In float-to-integer conversion, NaN inputs are converted to 02675        // when the source and destination bitwidths are both less than 64.2676        if (nvvm::FPToIntegerIntrinsicNaNZero(IntrinsicID))2677          return ConstantInt::get(Ty, 0);2678 2679        // Otherwise, the most significant bit is set.2680        unsigned BitWidth = Ty->getIntegerBitWidth();2681        uint64_t Val = 1ULL << (BitWidth - 1);2682        return ConstantInt::get(Ty, APInt(BitWidth, Val, /*IsSigned=*/false));2683      }2684 2685      APFloat::roundingMode RMode =2686          nvvm::GetFPToIntegerRoundingMode(IntrinsicID);2687      bool IsFTZ = nvvm::FPToIntegerIntrinsicShouldFTZ(IntrinsicID);2688      bool IsSigned = nvvm::FPToIntegerIntrinsicResultIsSigned(IntrinsicID);2689 2690      APSInt ResInt(Ty->getIntegerBitWidth(), !IsSigned);2691      auto FloatToRound = IsFTZ ? FTZPreserveSign(U) : U;2692 2693      // Return max/min value for integers if the result is +/-inf or2694      // is too large to fit in the result's integer bitwidth.2695      bool IsExact = false;2696      FloatToRound.convertToInteger(ResInt, RMode, &IsExact);2697      return ConstantInt::get(Ty, ResInt);2698    }2699    }2700 2701    /// We only fold functions with finite arguments. Folding NaN and inf is2702    /// likely to be aborted with an exception anyway, and some host libms2703    /// have known errors raising exceptions.2704    if (!U.isFinite())2705      return nullptr;2706 2707    /// Currently APFloat versions of these functions do not exist, so we use2708    /// the host native double versions.  Float versions are not called2709    /// directly but for all these it is true (float)(f((double)arg)) ==2710    /// f(arg).  Long double not supported yet.2711    const APFloat &APF = Op->getValueAPF();2712 2713    switch (IntrinsicID) {2714      default: break;2715      case Intrinsic::log:2716        return ConstantFoldFP(log, APF, Ty);2717      case Intrinsic::log2:2718        // TODO: What about hosts that lack a C99 library?2719        return ConstantFoldFP(log2, APF, Ty);2720      case Intrinsic::log10:2721        // TODO: What about hosts that lack a C99 library?2722        return ConstantFoldFP(log10, APF, Ty);2723      case Intrinsic::exp:2724        return ConstantFoldFP(exp, APF, Ty);2725      case Intrinsic::exp2:2726        // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.2727        return ConstantFoldBinaryFP(pow, APFloat(2.0), APF, Ty);2728      case Intrinsic::exp10:2729        // Fold exp10(x) as pow(10, x), in case the host lacks a C99 library.2730        return ConstantFoldBinaryFP(pow, APFloat(10.0), APF, Ty);2731      case Intrinsic::sin:2732        return ConstantFoldFP(sin, APF, Ty);2733      case Intrinsic::cos:2734        return ConstantFoldFP(cos, APF, Ty);2735      case Intrinsic::sinh:2736        return ConstantFoldFP(sinh, APF, Ty);2737      case Intrinsic::cosh:2738        return ConstantFoldFP(cosh, APF, Ty);2739      case Intrinsic::atan:2740        // Implement optional behavior from C's Annex F for +/-0.0.2741        if (U.isZero())2742          return ConstantFP::get(Ty->getContext(), U);2743        return ConstantFoldFP(atan, APF, Ty);2744      case Intrinsic::sqrt:2745        return ConstantFoldFP(sqrt, APF, Ty);2746 2747      // NVVM Intrinsics:2748      case Intrinsic::nvvm_ceil_ftz_f:2749      case Intrinsic::nvvm_ceil_f:2750      case Intrinsic::nvvm_ceil_d:2751        return ConstantFoldFP(2752            ceil, APF, Ty,2753            nvvm::GetNVVMDenormMode(2754                nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));2755 2756      case Intrinsic::nvvm_fabs_ftz:2757      case Intrinsic::nvvm_fabs:2758        return ConstantFoldFP(2759            fabs, APF, Ty,2760            nvvm::GetNVVMDenormMode(2761                nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));2762 2763      case Intrinsic::nvvm_floor_ftz_f:2764      case Intrinsic::nvvm_floor_f:2765      case Intrinsic::nvvm_floor_d:2766        return ConstantFoldFP(2767            floor, APF, Ty,2768            nvvm::GetNVVMDenormMode(2769                nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));2770 2771      case Intrinsic::nvvm_rcp_rm_ftz_f:2772      case Intrinsic::nvvm_rcp_rn_ftz_f:2773      case Intrinsic::nvvm_rcp_rp_ftz_f:2774      case Intrinsic::nvvm_rcp_rz_ftz_f:2775      case Intrinsic::nvvm_rcp_rm_d:2776      case Intrinsic::nvvm_rcp_rm_f:2777      case Intrinsic::nvvm_rcp_rn_d:2778      case Intrinsic::nvvm_rcp_rn_f:2779      case Intrinsic::nvvm_rcp_rp_d:2780      case Intrinsic::nvvm_rcp_rp_f:2781      case Intrinsic::nvvm_rcp_rz_d:2782      case Intrinsic::nvvm_rcp_rz_f: {2783        APFloat::roundingMode RoundMode = nvvm::GetRCPRoundingMode(IntrinsicID);2784        bool IsFTZ = nvvm::RCPShouldFTZ(IntrinsicID);2785 2786        auto Denominator = IsFTZ ? FTZPreserveSign(APF) : APF;2787        APFloat Res = APFloat::getOne(APF.getSemantics());2788        APFloat::opStatus Status = Res.divide(Denominator, RoundMode);2789 2790        if (Status == APFloat::opOK || Status == APFloat::opInexact) {2791          if (IsFTZ)2792            Res = FTZPreserveSign(Res);2793          return ConstantFP::get(Ty->getContext(), Res);2794        }2795        return nullptr;2796      }2797 2798      case Intrinsic::nvvm_round_ftz_f:2799      case Intrinsic::nvvm_round_f:2800      case Intrinsic::nvvm_round_d: {2801        // nvvm_round is lowered to PTX cvt.rni, which will round to nearest2802        // integer, choosing even integer if source is equidistant between two2803        // integers, so the semantics are closer to "rint" rather than "round".2804        bool IsFTZ = nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID);2805        auto V = IsFTZ ? FTZPreserveSign(APF) : APF;2806        V.roundToIntegral(APFloat::rmNearestTiesToEven);2807        return ConstantFP::get(Ty->getContext(), V);2808      }2809 2810      case Intrinsic::nvvm_saturate_ftz_f:2811      case Intrinsic::nvvm_saturate_d:2812      case Intrinsic::nvvm_saturate_f: {2813        bool IsFTZ = nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID);2814        auto V = IsFTZ ? FTZPreserveSign(APF) : APF;2815        if (V.isNegative() || V.isZero() || V.isNaN())2816          return ConstantFP::getZero(Ty);2817        APFloat One = APFloat::getOne(APF.getSemantics());2818        if (V > One)2819          return ConstantFP::get(Ty->getContext(), One);2820        return ConstantFP::get(Ty->getContext(), APF);2821      }2822 2823      case Intrinsic::nvvm_sqrt_rn_ftz_f:2824      case Intrinsic::nvvm_sqrt_f:2825      case Intrinsic::nvvm_sqrt_rn_d:2826      case Intrinsic::nvvm_sqrt_rn_f:2827        if (APF.isNegative())2828          return nullptr;2829        return ConstantFoldFP(2830            sqrt, APF, Ty,2831            nvvm::GetNVVMDenormMode(2832                nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));2833 2834      // AMDGCN Intrinsics:2835      case Intrinsic::amdgcn_cos:2836      case Intrinsic::amdgcn_sin: {2837        double V = getValueAsDouble(Op);2838        if (V < -256.0 || V > 256.0)2839          // The gfx8 and gfx9 architectures handle arguments outside the range2840          // [-256, 256] differently. This should be a rare case so bail out2841          // rather than trying to handle the difference.2842          return nullptr;2843        bool IsCos = IntrinsicID == Intrinsic::amdgcn_cos;2844        double V4 = V * 4.0;2845        if (V4 == floor(V4)) {2846          // Force exact results for quarter-integer inputs.2847          const double SinVals[4] = { 0.0, 1.0, 0.0, -1.0 };2848          V = SinVals[((int)V4 + (IsCos ? 1 : 0)) & 3];2849        } else {2850          if (IsCos)2851            V = cos(V * 2.0 * numbers::pi);2852          else2853            V = sin(V * 2.0 * numbers::pi);2854        }2855        return GetConstantFoldFPValue(V, Ty);2856      }2857    }2858 2859    if (!TLI)2860      return nullptr;2861 2862    LibFunc Func = NotLibFunc;2863    if (!TLI->getLibFunc(Name, Func))2864      return nullptr;2865 2866    switch (Func) {2867    default:2868      break;2869    case LibFunc_acos:2870    case LibFunc_acosf:2871    case LibFunc_acos_finite:2872    case LibFunc_acosf_finite:2873      if (TLI->has(Func))2874        return ConstantFoldFP(acos, APF, Ty);2875      break;2876    case LibFunc_asin:2877    case LibFunc_asinf:2878    case LibFunc_asin_finite:2879    case LibFunc_asinf_finite:2880      if (TLI->has(Func))2881        return ConstantFoldFP(asin, APF, Ty);2882      break;2883    case LibFunc_atan:2884    case LibFunc_atanf:2885      // Implement optional behavior from C's Annex F for +/-0.0.2886      if (U.isZero())2887        return ConstantFP::get(Ty->getContext(), U);2888      if (TLI->has(Func))2889        return ConstantFoldFP(atan, APF, Ty);2890      break;2891    case LibFunc_ceil:2892    case LibFunc_ceilf:2893      if (TLI->has(Func)) {2894        U.roundToIntegral(APFloat::rmTowardPositive);2895        return ConstantFP::get(Ty->getContext(), U);2896      }2897      break;2898    case LibFunc_cos:2899    case LibFunc_cosf:2900      if (TLI->has(Func))2901        return ConstantFoldFP(cos, APF, Ty);2902      break;2903    case LibFunc_cosh:2904    case LibFunc_coshf:2905    case LibFunc_cosh_finite:2906    case LibFunc_coshf_finite:2907      if (TLI->has(Func))2908        return ConstantFoldFP(cosh, APF, Ty);2909      break;2910    case LibFunc_exp:2911    case LibFunc_expf:2912    case LibFunc_exp_finite:2913    case LibFunc_expf_finite:2914      if (TLI->has(Func))2915        return ConstantFoldFP(exp, APF, Ty);2916      break;2917    case LibFunc_exp2:2918    case LibFunc_exp2f:2919    case LibFunc_exp2_finite:2920    case LibFunc_exp2f_finite:2921      if (TLI->has(Func))2922        // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.2923        return ConstantFoldBinaryFP(pow, APFloat(2.0), APF, Ty);2924      break;2925    case LibFunc_fabs:2926    case LibFunc_fabsf:2927      if (TLI->has(Func)) {2928        U.clearSign();2929        return ConstantFP::get(Ty->getContext(), U);2930      }2931      break;2932    case LibFunc_floor:2933    case LibFunc_floorf:2934      if (TLI->has(Func)) {2935        U.roundToIntegral(APFloat::rmTowardNegative);2936        return ConstantFP::get(Ty->getContext(), U);2937      }2938      break;2939    case LibFunc_log:2940    case LibFunc_logf:2941    case LibFunc_log_finite:2942    case LibFunc_logf_finite:2943      if (!APF.isNegative() && !APF.isZero() && TLI->has(Func))2944        return ConstantFoldFP(log, APF, Ty);2945      break;2946    case LibFunc_log2:2947    case LibFunc_log2f:2948    case LibFunc_log2_finite:2949    case LibFunc_log2f_finite:2950      if (!APF.isNegative() && !APF.isZero() && TLI->has(Func))2951        // TODO: What about hosts that lack a C99 library?2952        return ConstantFoldFP(log2, APF, Ty);2953      break;2954    case LibFunc_log10:2955    case LibFunc_log10f:2956    case LibFunc_log10_finite:2957    case LibFunc_log10f_finite:2958      if (!APF.isNegative() && !APF.isZero() && TLI->has(Func))2959        // TODO: What about hosts that lack a C99 library?2960        return ConstantFoldFP(log10, APF, Ty);2961      break;2962    case LibFunc_ilogb:2963    case LibFunc_ilogbf:2964      if (!APF.isZero() && TLI->has(Func))2965        return ConstantInt::get(Ty, ilogb(APF), true);2966      break;2967    case LibFunc_logb:2968    case LibFunc_logbf:2969      if (!APF.isZero() && TLI->has(Func))2970        return ConstantFoldFP(logb, APF, Ty);2971      break;2972    case LibFunc_log1p:2973    case LibFunc_log1pf:2974      // Implement optional behavior from C's Annex F for +/-0.0.2975      if (U.isZero())2976        return ConstantFP::get(Ty->getContext(), U);2977      if (APF > APFloat::getOne(APF.getSemantics(), true) && TLI->has(Func))2978        return ConstantFoldFP(log1p, APF, Ty);2979      break;2980    case LibFunc_logl:2981      return nullptr;2982    case LibFunc_erf:2983    case LibFunc_erff:2984      if (TLI->has(Func))2985        return ConstantFoldFP(erf, APF, Ty);2986      break;2987    case LibFunc_nearbyint:2988    case LibFunc_nearbyintf:2989    case LibFunc_rint:2990    case LibFunc_rintf:2991      if (TLI->has(Func)) {2992        U.roundToIntegral(APFloat::rmNearestTiesToEven);2993        return ConstantFP::get(Ty->getContext(), U);2994      }2995      break;2996    case LibFunc_round:2997    case LibFunc_roundf:2998      if (TLI->has(Func)) {2999        U.roundToIntegral(APFloat::rmNearestTiesToAway);3000        return ConstantFP::get(Ty->getContext(), U);3001      }3002      break;3003    case LibFunc_sin:3004    case LibFunc_sinf:3005      if (TLI->has(Func))3006        return ConstantFoldFP(sin, APF, Ty);3007      break;3008    case LibFunc_sinh:3009    case LibFunc_sinhf:3010    case LibFunc_sinh_finite:3011    case LibFunc_sinhf_finite:3012      if (TLI->has(Func))3013        return ConstantFoldFP(sinh, APF, Ty);3014      break;3015    case LibFunc_sqrt:3016    case LibFunc_sqrtf:3017      if (!APF.isNegative() && TLI->has(Func))3018        return ConstantFoldFP(sqrt, APF, Ty);3019      break;3020    case LibFunc_tan:3021    case LibFunc_tanf:3022      if (TLI->has(Func))3023        return ConstantFoldFP(tan, APF, Ty);3024      break;3025    case LibFunc_tanh:3026    case LibFunc_tanhf:3027      if (TLI->has(Func))3028        return ConstantFoldFP(tanh, APF, Ty);3029      break;3030    case LibFunc_trunc:3031    case LibFunc_truncf:3032      if (TLI->has(Func)) {3033        U.roundToIntegral(APFloat::rmTowardZero);3034        return ConstantFP::get(Ty->getContext(), U);3035      }3036      break;3037    }3038    return nullptr;3039  }3040 3041  if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {3042    switch (IntrinsicID) {3043    case Intrinsic::bswap:3044      return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());3045    case Intrinsic::ctpop:3046      return ConstantInt::get(Ty, Op->getValue().popcount());3047    case Intrinsic::bitreverse:3048      return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits());3049    case Intrinsic::convert_from_fp16: {3050      APFloat Val(APFloat::IEEEhalf(), Op->getValue());3051 3052      bool lost = false;3053      APFloat::opStatus status = Val.convert(3054          Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost);3055 3056      // Conversion is always precise.3057      (void)status;3058      assert(status != APFloat::opInexact && !lost &&3059             "Precision lost during fp16 constfolding");3060 3061      return ConstantFP::get(Ty->getContext(), Val);3062    }3063 3064    case Intrinsic::amdgcn_s_wqm: {3065      uint64_t Val = Op->getZExtValue();3066      Val |= (Val & 0x5555555555555555ULL) << 1 |3067             ((Val >> 1) & 0x5555555555555555ULL);3068      Val |= (Val & 0x3333333333333333ULL) << 2 |3069             ((Val >> 2) & 0x3333333333333333ULL);3070      return ConstantInt::get(Ty, Val);3071    }3072 3073    case Intrinsic::amdgcn_s_quadmask: {3074      uint64_t Val = Op->getZExtValue();3075      uint64_t QuadMask = 0;3076      for (unsigned I = 0; I < Op->getBitWidth() / 4; ++I, Val >>= 4) {3077        if (!(Val & 0xF))3078          continue;3079 3080        QuadMask |= (1ULL << I);3081      }3082      return ConstantInt::get(Ty, QuadMask);3083    }3084 3085    case Intrinsic::amdgcn_s_bitreplicate: {3086      uint64_t Val = Op->getZExtValue();3087      Val = (Val & 0x000000000000FFFFULL) | (Val & 0x00000000FFFF0000ULL) << 16;3088      Val = (Val & 0x000000FF000000FFULL) | (Val & 0x0000FF000000FF00ULL) << 8;3089      Val = (Val & 0x000F000F000F000FULL) | (Val & 0x00F000F000F000F0ULL) << 4;3090      Val = (Val & 0x0303030303030303ULL) | (Val & 0x0C0C0C0C0C0C0C0CULL) << 2;3091      Val = (Val & 0x1111111111111111ULL) | (Val & 0x2222222222222222ULL) << 1;3092      Val = Val | Val << 1;3093      return ConstantInt::get(Ty, Val);3094    }3095    }3096  }3097 3098  if (Operands[0]->getType()->isVectorTy()) {3099    auto *Op = cast<Constant>(Operands[0]);3100    switch (IntrinsicID) {3101    default: break;3102    case Intrinsic::vector_reduce_add:3103    case Intrinsic::vector_reduce_mul:3104    case Intrinsic::vector_reduce_and:3105    case Intrinsic::vector_reduce_or:3106    case Intrinsic::vector_reduce_xor:3107    case Intrinsic::vector_reduce_smin:3108    case Intrinsic::vector_reduce_smax:3109    case Intrinsic::vector_reduce_umin:3110    case Intrinsic::vector_reduce_umax:3111      if (Constant *C = constantFoldVectorReduce(IntrinsicID, Operands[0]))3112        return C;3113      break;3114    case Intrinsic::x86_sse_cvtss2si:3115    case Intrinsic::x86_sse_cvtss2si64:3116    case Intrinsic::x86_sse2_cvtsd2si:3117    case Intrinsic::x86_sse2_cvtsd2si64:3118      if (ConstantFP *FPOp =3119              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))3120        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),3121                                           /*roundTowardZero=*/false, Ty,3122                                           /*IsSigned*/true);3123      break;3124    case Intrinsic::x86_sse_cvttss2si:3125    case Intrinsic::x86_sse_cvttss2si64:3126    case Intrinsic::x86_sse2_cvttsd2si:3127    case Intrinsic::x86_sse2_cvttsd2si64:3128      if (ConstantFP *FPOp =3129              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))3130        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),3131                                           /*roundTowardZero=*/true, Ty,3132                                           /*IsSigned*/true);3133      break;3134 3135    case Intrinsic::wasm_anytrue:3136      return Op->isZeroValue() ? ConstantInt::get(Ty, 0)3137                               : ConstantInt::get(Ty, 1);3138 3139    case Intrinsic::wasm_alltrue:3140      // Check each element individually3141      unsigned E = cast<FixedVectorType>(Op->getType())->getNumElements();3142      for (unsigned I = 0; I != E; ++I) {3143        Constant *Elt = Op->getAggregateElement(I);3144        // Return false as soon as we find a non-true element.3145        if (Elt && Elt->isZeroValue())3146          return ConstantInt::get(Ty, 0);3147        // Bail as soon as we find an element we cannot prove to be true.3148        if (!Elt || !isa<ConstantInt>(Elt))3149          return nullptr;3150      }3151 3152      return ConstantInt::get(Ty, 1);3153    }3154  }3155 3156  return nullptr;3157}3158 3159static Constant *evaluateCompare(const APFloat &Op1, const APFloat &Op2,3160                                 const ConstrainedFPIntrinsic *Call) {3161  APFloat::opStatus St = APFloat::opOK;3162  auto *FCmp = cast<ConstrainedFPCmpIntrinsic>(Call);3163  FCmpInst::Predicate Cond = FCmp->getPredicate();3164  if (FCmp->isSignaling()) {3165    if (Op1.isNaN() || Op2.isNaN())3166      St = APFloat::opInvalidOp;3167  } else {3168    if (Op1.isSignaling() || Op2.isSignaling())3169      St = APFloat::opInvalidOp;3170  }3171  bool Result = FCmpInst::compare(Op1, Op2, Cond);3172  if (mayFoldConstrained(const_cast<ConstrainedFPCmpIntrinsic *>(FCmp), St))3173    return ConstantInt::get(Call->getType()->getScalarType(), Result);3174  return nullptr;3175}3176 3177static Constant *ConstantFoldLibCall2(StringRef Name, Type *Ty,3178                                      ArrayRef<Constant *> Operands,3179                                      const TargetLibraryInfo *TLI) {3180  if (!TLI)3181    return nullptr;3182 3183  LibFunc Func = NotLibFunc;3184  if (!TLI->getLibFunc(Name, Func))3185    return nullptr;3186 3187  const auto *Op1 = dyn_cast<ConstantFP>(Operands[0]);3188  if (!Op1)3189    return nullptr;3190 3191  const auto *Op2 = dyn_cast<ConstantFP>(Operands[1]);3192  if (!Op2)3193    return nullptr;3194 3195  const APFloat &Op1V = Op1->getValueAPF();3196  const APFloat &Op2V = Op2->getValueAPF();3197 3198  switch (Func) {3199  default:3200    break;3201  case LibFunc_pow:3202  case LibFunc_powf:3203  case LibFunc_pow_finite:3204  case LibFunc_powf_finite:3205    if (TLI->has(Func))3206      return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);3207    break;3208  case LibFunc_fmod:3209  case LibFunc_fmodf:3210    if (TLI->has(Func)) {3211      APFloat V = Op1->getValueAPF();3212      if (APFloat::opStatus::opOK == V.mod(Op2->getValueAPF()))3213        return ConstantFP::get(Ty->getContext(), V);3214    }3215    break;3216  case LibFunc_remainder:3217  case LibFunc_remainderf:3218    if (TLI->has(Func)) {3219      APFloat V = Op1->getValueAPF();3220      if (APFloat::opStatus::opOK == V.remainder(Op2->getValueAPF()))3221        return ConstantFP::get(Ty->getContext(), V);3222    }3223    break;3224  case LibFunc_atan2:3225  case LibFunc_atan2f:3226    // atan2(+/-0.0, +/-0.0) is known to raise an exception on some libm3227    // (Solaris), so we do not assume a known result for that.3228    if (Op1V.isZero() && Op2V.isZero())3229      return nullptr;3230    [[fallthrough]];3231  case LibFunc_atan2_finite:3232  case LibFunc_atan2f_finite:3233    if (TLI->has(Func))3234      return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);3235    break;3236  }3237 3238  return nullptr;3239}3240 3241static Constant *ConstantFoldIntrinsicCall2(Intrinsic::ID IntrinsicID, Type *Ty,3242                                            ArrayRef<Constant *> Operands,3243                                            const CallBase *Call) {3244  assert(Operands.size() == 2 && "Wrong number of operands.");3245 3246  if (Ty->isFloatingPointTy()) {3247    // TODO: We should have undef handling for all of the FP intrinsics that3248    //       are attempted to be folded in this function.3249    bool IsOp0Undef = isa<UndefValue>(Operands[0]);3250    bool IsOp1Undef = isa<UndefValue>(Operands[1]);3251    switch (IntrinsicID) {3252    case Intrinsic::maxnum:3253    case Intrinsic::minnum:3254    case Intrinsic::maximum:3255    case Intrinsic::minimum:3256    case Intrinsic::maximumnum:3257    case Intrinsic::minimumnum:3258    case Intrinsic::nvvm_fmax_d:3259    case Intrinsic::nvvm_fmin_d:3260      // If one argument is undef, return the other argument.3261      if (IsOp0Undef)3262        return Operands[1];3263      if (IsOp1Undef)3264        return Operands[0];3265      break;3266 3267    case Intrinsic::nvvm_fmax_f:3268    case Intrinsic::nvvm_fmax_ftz_f:3269    case Intrinsic::nvvm_fmax_ftz_nan_f:3270    case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:3271    case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:3272    case Intrinsic::nvvm_fmax_nan_f:3273    case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:3274    case Intrinsic::nvvm_fmax_xorsign_abs_f:3275 3276    case Intrinsic::nvvm_fmin_f:3277    case Intrinsic::nvvm_fmin_ftz_f:3278    case Intrinsic::nvvm_fmin_ftz_nan_f:3279    case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:3280    case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:3281    case Intrinsic::nvvm_fmin_nan_f:3282    case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:3283    case Intrinsic::nvvm_fmin_xorsign_abs_f:3284      // If one arg is undef, the other arg can be returned only if it is3285      // constant, as we may need to flush it to sign-preserving zero or3286      // canonicalize the NaN.3287      if (!IsOp0Undef && !IsOp1Undef)3288        break;3289      if (auto *Op = dyn_cast<ConstantFP>(Operands[IsOp0Undef ? 1 : 0])) {3290        if (Op->isNaN()) {3291          APInt NVCanonicalNaN(32, 0x7fffffff);3292          return ConstantFP::get(3293              Ty, APFloat(Ty->getFltSemantics(), NVCanonicalNaN));3294        }3295        if (nvvm::FMinFMaxShouldFTZ(IntrinsicID))3296          return ConstantFP::get(Ty, FTZPreserveSign(Op->getValueAPF()));3297        else3298          return Op;3299      }3300      break;3301    }3302  }3303 3304  if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {3305    const APFloat &Op1V = Op1->getValueAPF();3306 3307    if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {3308      if (Op2->getType() != Op1->getType())3309        return nullptr;3310      const APFloat &Op2V = Op2->getValueAPF();3311 3312      if (const auto *ConstrIntr =3313              dyn_cast_if_present<ConstrainedFPIntrinsic>(Call)) {3314        RoundingMode RM = getEvaluationRoundingMode(ConstrIntr);3315        APFloat Res = Op1V;3316        APFloat::opStatus St;3317        switch (IntrinsicID) {3318        default:3319          return nullptr;3320        case Intrinsic::experimental_constrained_fadd:3321          St = Res.add(Op2V, RM);3322          break;3323        case Intrinsic::experimental_constrained_fsub:3324          St = Res.subtract(Op2V, RM);3325          break;3326        case Intrinsic::experimental_constrained_fmul:3327          St = Res.multiply(Op2V, RM);3328          break;3329        case Intrinsic::experimental_constrained_fdiv:3330          St = Res.divide(Op2V, RM);3331          break;3332        case Intrinsic::experimental_constrained_frem:3333          St = Res.mod(Op2V);3334          break;3335        case Intrinsic::experimental_constrained_fcmp:3336        case Intrinsic::experimental_constrained_fcmps:3337          return evaluateCompare(Op1V, Op2V, ConstrIntr);3338        }3339        if (mayFoldConstrained(const_cast<ConstrainedFPIntrinsic *>(ConstrIntr),3340                               St))3341          return ConstantFP::get(Ty->getContext(), Res);3342        return nullptr;3343      }3344 3345      switch (IntrinsicID) {3346      default:3347        break;3348      case Intrinsic::copysign:3349        return ConstantFP::get(Ty->getContext(), APFloat::copySign(Op1V, Op2V));3350      case Intrinsic::minnum:3351        return ConstantFP::get(Ty->getContext(), minnum(Op1V, Op2V));3352      case Intrinsic::maxnum:3353        return ConstantFP::get(Ty->getContext(), maxnum(Op1V, Op2V));3354      case Intrinsic::minimum:3355        return ConstantFP::get(Ty->getContext(), minimum(Op1V, Op2V));3356      case Intrinsic::maximum:3357        return ConstantFP::get(Ty->getContext(), maximum(Op1V, Op2V));3358      case Intrinsic::minimumnum:3359        return ConstantFP::get(Ty->getContext(), minimumnum(Op1V, Op2V));3360      case Intrinsic::maximumnum:3361        return ConstantFP::get(Ty->getContext(), maximumnum(Op1V, Op2V));3362 3363      case Intrinsic::nvvm_fmax_d:3364      case Intrinsic::nvvm_fmax_f:3365      case Intrinsic::nvvm_fmax_ftz_f:3366      case Intrinsic::nvvm_fmax_ftz_nan_f:3367      case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:3368      case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:3369      case Intrinsic::nvvm_fmax_nan_f:3370      case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:3371      case Intrinsic::nvvm_fmax_xorsign_abs_f:3372 3373      case Intrinsic::nvvm_fmin_d:3374      case Intrinsic::nvvm_fmin_f:3375      case Intrinsic::nvvm_fmin_ftz_f:3376      case Intrinsic::nvvm_fmin_ftz_nan_f:3377      case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:3378      case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:3379      case Intrinsic::nvvm_fmin_nan_f:3380      case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:3381      case Intrinsic::nvvm_fmin_xorsign_abs_f: {3382 3383        bool ShouldCanonicalizeNaNs = !(IntrinsicID == Intrinsic::nvvm_fmax_d ||3384                                        IntrinsicID == Intrinsic::nvvm_fmin_d);3385        bool IsFTZ = nvvm::FMinFMaxShouldFTZ(IntrinsicID);3386        bool IsNaNPropagating = nvvm::FMinFMaxPropagatesNaNs(IntrinsicID);3387        bool IsXorSignAbs = nvvm::FMinFMaxIsXorSignAbs(IntrinsicID);3388 3389        APFloat A = IsFTZ ? FTZPreserveSign(Op1V) : Op1V;3390        APFloat B = IsFTZ ? FTZPreserveSign(Op2V) : Op2V;3391 3392        bool XorSign = false;3393        if (IsXorSignAbs) {3394          XorSign = A.isNegative() ^ B.isNegative();3395          A = abs(A);3396          B = abs(B);3397        }3398 3399        bool IsFMax = false;3400        switch (IntrinsicID) {3401        case Intrinsic::nvvm_fmax_d:3402        case Intrinsic::nvvm_fmax_f:3403        case Intrinsic::nvvm_fmax_ftz_f:3404        case Intrinsic::nvvm_fmax_ftz_nan_f:3405        case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:3406        case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:3407        case Intrinsic::nvvm_fmax_nan_f:3408        case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:3409        case Intrinsic::nvvm_fmax_xorsign_abs_f:3410          IsFMax = true;3411          break;3412        }3413        APFloat Res = IsFMax ? maximum(A, B) : minimum(A, B);3414 3415        if (ShouldCanonicalizeNaNs) {3416          APFloat NVCanonicalNaN(Res.getSemantics(), APInt(32, 0x7fffffff));3417          if (A.isNaN() && B.isNaN())3418            return ConstantFP::get(Ty, NVCanonicalNaN);3419          else if (IsNaNPropagating && (A.isNaN() || B.isNaN()))3420            return ConstantFP::get(Ty, NVCanonicalNaN);3421        }3422 3423        if (A.isNaN() && B.isNaN())3424          return Operands[1];3425        else if (A.isNaN())3426          Res = B;3427        else if (B.isNaN())3428          Res = A;3429 3430        if (IsXorSignAbs && XorSign != Res.isNegative())3431          Res.changeSign();3432 3433        return ConstantFP::get(Ty->getContext(), Res);3434      }3435 3436      case Intrinsic::nvvm_add_rm_f:3437      case Intrinsic::nvvm_add_rn_f:3438      case Intrinsic::nvvm_add_rp_f:3439      case Intrinsic::nvvm_add_rz_f:3440      case Intrinsic::nvvm_add_rm_d:3441      case Intrinsic::nvvm_add_rn_d:3442      case Intrinsic::nvvm_add_rp_d:3443      case Intrinsic::nvvm_add_rz_d:3444      case Intrinsic::nvvm_add_rm_ftz_f:3445      case Intrinsic::nvvm_add_rn_ftz_f:3446      case Intrinsic::nvvm_add_rp_ftz_f:3447      case Intrinsic::nvvm_add_rz_ftz_f: {3448 3449        bool IsFTZ = nvvm::FAddShouldFTZ(IntrinsicID);3450        APFloat A = IsFTZ ? FTZPreserveSign(Op1V) : Op1V;3451        APFloat B = IsFTZ ? FTZPreserveSign(Op2V) : Op2V;3452 3453        APFloat::roundingMode RoundMode =3454            nvvm::GetFAddRoundingMode(IntrinsicID);3455 3456        APFloat Res = A;3457        APFloat::opStatus Status = Res.add(B, RoundMode);3458 3459        if (!Res.isNaN() &&3460            (Status == APFloat::opOK || Status == APFloat::opInexact)) {3461          Res = IsFTZ ? FTZPreserveSign(Res) : Res;3462          return ConstantFP::get(Ty->getContext(), Res);3463        }3464        return nullptr;3465      }3466 3467      case Intrinsic::nvvm_mul_rm_f:3468      case Intrinsic::nvvm_mul_rn_f:3469      case Intrinsic::nvvm_mul_rp_f:3470      case Intrinsic::nvvm_mul_rz_f:3471      case Intrinsic::nvvm_mul_rm_d:3472      case Intrinsic::nvvm_mul_rn_d:3473      case Intrinsic::nvvm_mul_rp_d:3474      case Intrinsic::nvvm_mul_rz_d:3475      case Intrinsic::nvvm_mul_rm_ftz_f:3476      case Intrinsic::nvvm_mul_rn_ftz_f:3477      case Intrinsic::nvvm_mul_rp_ftz_f:3478      case Intrinsic::nvvm_mul_rz_ftz_f: {3479 3480        bool IsFTZ = nvvm::FMulShouldFTZ(IntrinsicID);3481        APFloat A = IsFTZ ? FTZPreserveSign(Op1V) : Op1V;3482        APFloat B = IsFTZ ? FTZPreserveSign(Op2V) : Op2V;3483 3484        APFloat::roundingMode RoundMode =3485            nvvm::GetFMulRoundingMode(IntrinsicID);3486 3487        APFloat Res = A;3488        APFloat::opStatus Status = Res.multiply(B, RoundMode);3489 3490        if (!Res.isNaN() &&3491            (Status == APFloat::opOK || Status == APFloat::opInexact)) {3492          Res = IsFTZ ? FTZPreserveSign(Res) : Res;3493          return ConstantFP::get(Ty->getContext(), Res);3494        }3495        return nullptr;3496      }3497 3498      case Intrinsic::nvvm_div_rm_f:3499      case Intrinsic::nvvm_div_rn_f:3500      case Intrinsic::nvvm_div_rp_f:3501      case Intrinsic::nvvm_div_rz_f:3502      case Intrinsic::nvvm_div_rm_d:3503      case Intrinsic::nvvm_div_rn_d:3504      case Intrinsic::nvvm_div_rp_d:3505      case Intrinsic::nvvm_div_rz_d:3506      case Intrinsic::nvvm_div_rm_ftz_f:3507      case Intrinsic::nvvm_div_rn_ftz_f:3508      case Intrinsic::nvvm_div_rp_ftz_f:3509      case Intrinsic::nvvm_div_rz_ftz_f: {3510        bool IsFTZ = nvvm::FDivShouldFTZ(IntrinsicID);3511        APFloat A = IsFTZ ? FTZPreserveSign(Op1V) : Op1V;3512        APFloat B = IsFTZ ? FTZPreserveSign(Op2V) : Op2V;3513        APFloat::roundingMode RoundMode =3514            nvvm::GetFDivRoundingMode(IntrinsicID);3515 3516        APFloat Res = A;3517        APFloat::opStatus Status = Res.divide(B, RoundMode);3518        if (!Res.isNaN() &&3519            (Status == APFloat::opOK || Status == APFloat::opInexact)) {3520          Res = IsFTZ ? FTZPreserveSign(Res) : Res;3521          return ConstantFP::get(Ty->getContext(), Res);3522        }3523        return nullptr;3524      }3525      }3526 3527      if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())3528        return nullptr;3529 3530      switch (IntrinsicID) {3531      default:3532        break;3533      case Intrinsic::pow:3534        return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);3535      case Intrinsic::amdgcn_fmul_legacy:3536        // The legacy behaviour is that multiplying +/- 0.0 by anything, even3537        // NaN or infinity, gives +0.0.3538        if (Op1V.isZero() || Op2V.isZero())3539          return ConstantFP::getZero(Ty);3540        return ConstantFP::get(Ty->getContext(), Op1V * Op2V);3541      }3542 3543    } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) {3544      switch (IntrinsicID) {3545      case Intrinsic::ldexp: {3546        return ConstantFP::get(3547            Ty->getContext(),3548            scalbn(Op1V, Op2C->getSExtValue(), APFloat::rmNearestTiesToEven));3549      }3550      case Intrinsic::is_fpclass: {3551        FPClassTest Mask = static_cast<FPClassTest>(Op2C->getZExtValue());3552        bool Result =3553          ((Mask & fcSNan) && Op1V.isNaN() && Op1V.isSignaling()) ||3554          ((Mask & fcQNan) && Op1V.isNaN() && !Op1V.isSignaling()) ||3555          ((Mask & fcNegInf) && Op1V.isNegInfinity()) ||3556          ((Mask & fcNegNormal) && Op1V.isNormal() && Op1V.isNegative()) ||3557          ((Mask & fcNegSubnormal) && Op1V.isDenormal() && Op1V.isNegative()) ||3558          ((Mask & fcNegZero) && Op1V.isZero() && Op1V.isNegative()) ||3559          ((Mask & fcPosZero) && Op1V.isZero() && !Op1V.isNegative()) ||3560          ((Mask & fcPosSubnormal) && Op1V.isDenormal() && !Op1V.isNegative()) ||3561          ((Mask & fcPosNormal) && Op1V.isNormal() && !Op1V.isNegative()) ||3562          ((Mask & fcPosInf) && Op1V.isPosInfinity());3563        return ConstantInt::get(Ty, Result);3564      }3565      case Intrinsic::powi: {3566        int Exp = static_cast<int>(Op2C->getSExtValue());3567        switch (Ty->getTypeID()) {3568        case Type::HalfTyID:3569        case Type::FloatTyID: {3570          APFloat Res(static_cast<float>(std::pow(Op1V.convertToFloat(), Exp)));3571          if (Ty->isHalfTy()) {3572            bool Unused;3573            Res.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven,3574                        &Unused);3575          }3576          return ConstantFP::get(Ty->getContext(), Res);3577        }3578        case Type::DoubleTyID:3579          return ConstantFP::get(Ty, std::pow(Op1V.convertToDouble(), Exp));3580        default:3581          return nullptr;3582        }3583      }3584      default:3585        break;3586      }3587    }3588    return nullptr;3589  }3590 3591  if (Operands[0]->getType()->isIntegerTy() &&3592      Operands[1]->getType()->isIntegerTy()) {3593    const APInt *C0, *C1;3594    if (!getConstIntOrUndef(Operands[0], C0) ||3595        !getConstIntOrUndef(Operands[1], C1))3596      return nullptr;3597 3598    switch (IntrinsicID) {3599    default: break;3600    case Intrinsic::smax:3601    case Intrinsic::smin:3602    case Intrinsic::umax:3603    case Intrinsic::umin:3604      if (!C0 && !C1)3605        return UndefValue::get(Ty);3606      if (!C0 || !C1)3607        return MinMaxIntrinsic::getSaturationPoint(IntrinsicID, Ty);3608      return ConstantInt::get(3609          Ty, ICmpInst::compare(*C0, *C1,3610                                MinMaxIntrinsic::getPredicate(IntrinsicID))3611                  ? *C03612                  : *C1);3613 3614    case Intrinsic::scmp:3615    case Intrinsic::ucmp:3616      if (!C0 || !C1)3617        return ConstantInt::get(Ty, 0);3618 3619      int Res;3620      if (IntrinsicID == Intrinsic::scmp)3621        Res = C0->sgt(*C1) ? 1 : C0->slt(*C1) ? -1 : 0;3622      else3623        Res = C0->ugt(*C1) ? 1 : C0->ult(*C1) ? -1 : 0;3624      return ConstantInt::get(Ty, Res, /*IsSigned=*/true);3625 3626    case Intrinsic::usub_with_overflow:3627    case Intrinsic::ssub_with_overflow:3628      // X - undef -> { 0, false }3629      // undef - X -> { 0, false }3630      if (!C0 || !C1)3631        return Constant::getNullValue(Ty);3632      [[fallthrough]];3633    case Intrinsic::uadd_with_overflow:3634    case Intrinsic::sadd_with_overflow:3635      // X + undef -> { -1, false }3636      // undef + x -> { -1, false }3637      if (!C0 || !C1) {3638        return ConstantStruct::get(3639            cast<StructType>(Ty),3640            {Constant::getAllOnesValue(Ty->getStructElementType(0)),3641             Constant::getNullValue(Ty->getStructElementType(1))});3642      }3643      [[fallthrough]];3644    case Intrinsic::smul_with_overflow:3645    case Intrinsic::umul_with_overflow: {3646      // undef * X -> { 0, false }3647      // X * undef -> { 0, false }3648      if (!C0 || !C1)3649        return Constant::getNullValue(Ty);3650 3651      APInt Res;3652      bool Overflow;3653      switch (IntrinsicID) {3654      default: llvm_unreachable("Invalid case");3655      case Intrinsic::sadd_with_overflow:3656        Res = C0->sadd_ov(*C1, Overflow);3657        break;3658      case Intrinsic::uadd_with_overflow:3659        Res = C0->uadd_ov(*C1, Overflow);3660        break;3661      case Intrinsic::ssub_with_overflow:3662        Res = C0->ssub_ov(*C1, Overflow);3663        break;3664      case Intrinsic::usub_with_overflow:3665        Res = C0->usub_ov(*C1, Overflow);3666        break;3667      case Intrinsic::smul_with_overflow:3668        Res = C0->smul_ov(*C1, Overflow);3669        break;3670      case Intrinsic::umul_with_overflow:3671        Res = C0->umul_ov(*C1, Overflow);3672        break;3673      }3674      Constant *Ops[] = {3675        ConstantInt::get(Ty->getContext(), Res),3676        ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)3677      };3678      return ConstantStruct::get(cast<StructType>(Ty), Ops);3679    }3680    case Intrinsic::uadd_sat:3681    case Intrinsic::sadd_sat:3682      if (!C0 && !C1)3683        return UndefValue::get(Ty);3684      if (!C0 || !C1)3685        return Constant::getAllOnesValue(Ty);3686      if (IntrinsicID == Intrinsic::uadd_sat)3687        return ConstantInt::get(Ty, C0->uadd_sat(*C1));3688      else3689        return ConstantInt::get(Ty, C0->sadd_sat(*C1));3690    case Intrinsic::usub_sat:3691    case Intrinsic::ssub_sat:3692      if (!C0 && !C1)3693        return UndefValue::get(Ty);3694      if (!C0 || !C1)3695        return Constant::getNullValue(Ty);3696      if (IntrinsicID == Intrinsic::usub_sat)3697        return ConstantInt::get(Ty, C0->usub_sat(*C1));3698      else3699        return ConstantInt::get(Ty, C0->ssub_sat(*C1));3700    case Intrinsic::cttz:3701    case Intrinsic::ctlz:3702      assert(C1 && "Must be constant int");3703 3704      // cttz(0, 1) and ctlz(0, 1) are poison.3705      if (C1->isOne() && (!C0 || C0->isZero()))3706        return PoisonValue::get(Ty);3707      if (!C0)3708        return Constant::getNullValue(Ty);3709      if (IntrinsicID == Intrinsic::cttz)3710        return ConstantInt::get(Ty, C0->countr_zero());3711      else3712        return ConstantInt::get(Ty, C0->countl_zero());3713 3714    case Intrinsic::abs:3715      assert(C1 && "Must be constant int");3716      assert((C1->isOne() || C1->isZero()) && "Must be 0 or 1");3717 3718      // Undef or minimum val operand with poison min --> poison3719      if (C1->isOne() && (!C0 || C0->isMinSignedValue()))3720        return PoisonValue::get(Ty);3721 3722      // Undef operand with no poison min --> 0 (sign bit must be clear)3723      if (!C0)3724        return Constant::getNullValue(Ty);3725 3726      return ConstantInt::get(Ty, C0->abs());3727    case Intrinsic::amdgcn_wave_reduce_umin:3728    case Intrinsic::amdgcn_wave_reduce_umax:3729    case Intrinsic::amdgcn_wave_reduce_max:3730    case Intrinsic::amdgcn_wave_reduce_min:3731    case Intrinsic::amdgcn_wave_reduce_add:3732    case Intrinsic::amdgcn_wave_reduce_sub:3733    case Intrinsic::amdgcn_wave_reduce_and:3734    case Intrinsic::amdgcn_wave_reduce_or:3735    case Intrinsic::amdgcn_wave_reduce_xor:3736      return dyn_cast<Constant>(Operands[0]);3737    }3738 3739    return nullptr;3740  }3741 3742  // Support ConstantVector in case we have an Undef in the top.3743  if ((isa<ConstantVector>(Operands[0]) ||3744       isa<ConstantDataVector>(Operands[0])) &&3745      // Check for default rounding mode.3746      // FIXME: Support other rounding modes?3747      isa<ConstantInt>(Operands[1]) &&3748      cast<ConstantInt>(Operands[1])->getValue() == 4) {3749    auto *Op = cast<Constant>(Operands[0]);3750    switch (IntrinsicID) {3751    default: break;3752    case Intrinsic::x86_avx512_vcvtss2si32:3753    case Intrinsic::x86_avx512_vcvtss2si64:3754    case Intrinsic::x86_avx512_vcvtsd2si32:3755    case Intrinsic::x86_avx512_vcvtsd2si64:3756      if (ConstantFP *FPOp =3757              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))3758        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),3759                                           /*roundTowardZero=*/false, Ty,3760                                           /*IsSigned*/true);3761      break;3762    case Intrinsic::x86_avx512_vcvtss2usi32:3763    case Intrinsic::x86_avx512_vcvtss2usi64:3764    case Intrinsic::x86_avx512_vcvtsd2usi32:3765    case Intrinsic::x86_avx512_vcvtsd2usi64:3766      if (ConstantFP *FPOp =3767              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))3768        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),3769                                           /*roundTowardZero=*/false, Ty,3770                                           /*IsSigned*/false);3771      break;3772    case Intrinsic::x86_avx512_cvttss2si:3773    case Intrinsic::x86_avx512_cvttss2si64:3774    case Intrinsic::x86_avx512_cvttsd2si:3775    case Intrinsic::x86_avx512_cvttsd2si64:3776      if (ConstantFP *FPOp =3777              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))3778        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),3779                                           /*roundTowardZero=*/true, Ty,3780                                           /*IsSigned*/true);3781      break;3782    case Intrinsic::x86_avx512_cvttss2usi:3783    case Intrinsic::x86_avx512_cvttss2usi64:3784    case Intrinsic::x86_avx512_cvttsd2usi:3785    case Intrinsic::x86_avx512_cvttsd2usi64:3786      if (ConstantFP *FPOp =3787              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))3788        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),3789                                           /*roundTowardZero=*/true, Ty,3790                                           /*IsSigned*/false);3791      break;3792    }3793  }3794  return nullptr;3795}3796 3797static APFloat ConstantFoldAMDGCNCubeIntrinsic(Intrinsic::ID IntrinsicID,3798                                               const APFloat &S0,3799                                               const APFloat &S1,3800                                               const APFloat &S2) {3801  unsigned ID;3802  const fltSemantics &Sem = S0.getSemantics();3803  APFloat MA(Sem), SC(Sem), TC(Sem);3804  if (abs(S2) >= abs(S0) && abs(S2) >= abs(S1)) {3805    if (S2.isNegative() && S2.isNonZero() && !S2.isNaN()) {3806      // S2 < 03807      ID = 5;3808      SC = -S0;3809    } else {3810      ID = 4;3811      SC = S0;3812    }3813    MA = S2;3814    TC = -S1;3815  } else if (abs(S1) >= abs(S0)) {3816    if (S1.isNegative() && S1.isNonZero() && !S1.isNaN()) {3817      // S1 < 03818      ID = 3;3819      TC = -S2;3820    } else {3821      ID = 2;3822      TC = S2;3823    }3824    MA = S1;3825    SC = S0;3826  } else {3827    if (S0.isNegative() && S0.isNonZero() && !S0.isNaN()) {3828      // S0 < 03829      ID = 1;3830      SC = S2;3831    } else {3832      ID = 0;3833      SC = -S2;3834    }3835    MA = S0;3836    TC = -S1;3837  }3838  switch (IntrinsicID) {3839  default:3840    llvm_unreachable("unhandled amdgcn cube intrinsic");3841  case Intrinsic::amdgcn_cubeid:3842    return APFloat(Sem, ID);3843  case Intrinsic::amdgcn_cubema:3844    return MA + MA;3845  case Intrinsic::amdgcn_cubesc:3846    return SC;3847  case Intrinsic::amdgcn_cubetc:3848    return TC;3849  }3850}3851 3852static Constant *ConstantFoldAMDGCNPermIntrinsic(ArrayRef<Constant *> Operands,3853                                                 Type *Ty) {3854  const APInt *C0, *C1, *C2;3855  if (!getConstIntOrUndef(Operands[0], C0) ||3856      !getConstIntOrUndef(Operands[1], C1) ||3857      !getConstIntOrUndef(Operands[2], C2))3858    return nullptr;3859 3860  if (!C2)3861    return UndefValue::get(Ty);3862 3863  APInt Val(32, 0);3864  unsigned NumUndefBytes = 0;3865  for (unsigned I = 0; I < 32; I += 8) {3866    unsigned Sel = C2->extractBitsAsZExtValue(8, I);3867    unsigned B = 0;3868 3869    if (Sel >= 13)3870      B = 0xff;3871    else if (Sel == 12)3872      B = 0x00;3873    else {3874      const APInt *Src = ((Sel & 10) == 10 || (Sel & 12) == 4) ? C0 : C1;3875      if (!Src)3876        ++NumUndefBytes;3877      else if (Sel < 8)3878        B = Src->extractBitsAsZExtValue(8, (Sel & 3) * 8);3879      else3880        B = Src->extractBitsAsZExtValue(1, (Sel & 1) ? 31 : 15) * 0xff;3881    }3882 3883    Val.insertBits(B, I, 8);3884  }3885 3886  if (NumUndefBytes == 4)3887    return UndefValue::get(Ty);3888 3889  return ConstantInt::get(Ty, Val);3890}3891 3892static Constant *ConstantFoldScalarCall3(StringRef Name,3893                                         Intrinsic::ID IntrinsicID,3894                                         Type *Ty,3895                                         ArrayRef<Constant *> Operands,3896                                         const TargetLibraryInfo *TLI,3897                                         const CallBase *Call) {3898  assert(Operands.size() == 3 && "Wrong number of operands.");3899 3900  if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {3901    if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {3902      if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) {3903        const APFloat &C1 = Op1->getValueAPF();3904        const APFloat &C2 = Op2->getValueAPF();3905        const APFloat &C3 = Op3->getValueAPF();3906 3907        if (const auto *ConstrIntr = dyn_cast<ConstrainedFPIntrinsic>(Call)) {3908          RoundingMode RM = getEvaluationRoundingMode(ConstrIntr);3909          APFloat Res = C1;3910          APFloat::opStatus St;3911          switch (IntrinsicID) {3912          default:3913            return nullptr;3914          case Intrinsic::experimental_constrained_fma:3915          case Intrinsic::experimental_constrained_fmuladd:3916            St = Res.fusedMultiplyAdd(C2, C3, RM);3917            break;3918          }3919          if (mayFoldConstrained(3920                  const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), St))3921            return ConstantFP::get(Ty->getContext(), Res);3922          return nullptr;3923        }3924 3925        switch (IntrinsicID) {3926        default: break;3927        case Intrinsic::amdgcn_fma_legacy: {3928          // The legacy behaviour is that multiplying +/- 0.0 by anything, even3929          // NaN or infinity, gives +0.0.3930          if (C1.isZero() || C2.isZero()) {3931            // It's tempting to just return C3 here, but that would give the3932            // wrong result if C3 was -0.0.3933            return ConstantFP::get(Ty->getContext(), APFloat(0.0f) + C3);3934          }3935          [[fallthrough]];3936        }3937        case Intrinsic::fma:3938        case Intrinsic::fmuladd: {3939          APFloat V = C1;3940          V.fusedMultiplyAdd(C2, C3, APFloat::rmNearestTiesToEven);3941          return ConstantFP::get(Ty->getContext(), V);3942        }3943 3944        case Intrinsic::nvvm_fma_rm_f:3945        case Intrinsic::nvvm_fma_rn_f:3946        case Intrinsic::nvvm_fma_rp_f:3947        case Intrinsic::nvvm_fma_rz_f:3948        case Intrinsic::nvvm_fma_rm_d:3949        case Intrinsic::nvvm_fma_rn_d:3950        case Intrinsic::nvvm_fma_rp_d:3951        case Intrinsic::nvvm_fma_rz_d:3952        case Intrinsic::nvvm_fma_rm_ftz_f:3953        case Intrinsic::nvvm_fma_rn_ftz_f:3954        case Intrinsic::nvvm_fma_rp_ftz_f:3955        case Intrinsic::nvvm_fma_rz_ftz_f: {3956          bool IsFTZ = nvvm::FMAShouldFTZ(IntrinsicID);3957          APFloat A = IsFTZ ? FTZPreserveSign(C1) : C1;3958          APFloat B = IsFTZ ? FTZPreserveSign(C2) : C2;3959          APFloat C = IsFTZ ? FTZPreserveSign(C3) : C3;3960 3961          APFloat::roundingMode RoundMode =3962              nvvm::GetFMARoundingMode(IntrinsicID);3963 3964          APFloat Res = A;3965          APFloat::opStatus Status = Res.fusedMultiplyAdd(B, C, RoundMode);3966 3967          if (!Res.isNaN() &&3968              (Status == APFloat::opOK || Status == APFloat::opInexact)) {3969            Res = IsFTZ ? FTZPreserveSign(Res) : Res;3970            return ConstantFP::get(Ty->getContext(), Res);3971          }3972          return nullptr;3973        }3974 3975        case Intrinsic::amdgcn_cubeid:3976        case Intrinsic::amdgcn_cubema:3977        case Intrinsic::amdgcn_cubesc:3978        case Intrinsic::amdgcn_cubetc: {3979          APFloat V = ConstantFoldAMDGCNCubeIntrinsic(IntrinsicID, C1, C2, C3);3980          return ConstantFP::get(Ty->getContext(), V);3981        }3982        }3983      }3984    }3985  }3986 3987  if (IntrinsicID == Intrinsic::smul_fix ||3988      IntrinsicID == Intrinsic::smul_fix_sat) {3989    const APInt *C0, *C1;3990    if (!getConstIntOrUndef(Operands[0], C0) ||3991        !getConstIntOrUndef(Operands[1], C1))3992      return nullptr;3993 3994    // undef * C -> 03995    // C * undef -> 03996    if (!C0 || !C1)3997      return Constant::getNullValue(Ty);3998 3999    // This code performs rounding towards negative infinity in case the result4000    // cannot be represented exactly for the given scale. Targets that do care4001    // about rounding should use a target hook for specifying how rounding4002    // should be done, and provide their own folding to be consistent with4003    // rounding. This is the same approach as used by4004    // DAGTypeLegalizer::ExpandIntRes_MULFIX.4005    unsigned Scale = cast<ConstantInt>(Operands[2])->getZExtValue();4006    unsigned Width = C0->getBitWidth();4007    assert(Scale < Width && "Illegal scale.");4008    unsigned ExtendedWidth = Width * 2;4009    APInt Product =4010        (C0->sext(ExtendedWidth) * C1->sext(ExtendedWidth)).ashr(Scale);4011    if (IntrinsicID == Intrinsic::smul_fix_sat) {4012      APInt Max = APInt::getSignedMaxValue(Width).sext(ExtendedWidth);4013      APInt Min = APInt::getSignedMinValue(Width).sext(ExtendedWidth);4014      Product = APIntOps::smin(Product, Max);4015      Product = APIntOps::smax(Product, Min);4016    }4017    return ConstantInt::get(Ty->getContext(), Product.sextOrTrunc(Width));4018  }4019 4020  if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) {4021    const APInt *C0, *C1, *C2;4022    if (!getConstIntOrUndef(Operands[0], C0) ||4023        !getConstIntOrUndef(Operands[1], C1) ||4024        !getConstIntOrUndef(Operands[2], C2))4025      return nullptr;4026 4027    bool IsRight = IntrinsicID == Intrinsic::fshr;4028    if (!C2)4029      return Operands[IsRight ? 1 : 0];4030    if (!C0 && !C1)4031      return UndefValue::get(Ty);4032 4033    // The shift amount is interpreted as modulo the bitwidth. If the shift4034    // amount is effectively 0, avoid UB due to oversized inverse shift below.4035    unsigned BitWidth = C2->getBitWidth();4036    unsigned ShAmt = C2->urem(BitWidth);4037    if (!ShAmt)4038      return Operands[IsRight ? 1 : 0];4039 4040    // (C0 << ShlAmt) | (C1 >> LshrAmt)4041    unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt;4042    unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt;4043    if (!C0)4044      return ConstantInt::get(Ty, C1->lshr(LshrAmt));4045    if (!C1)4046      return ConstantInt::get(Ty, C0->shl(ShlAmt));4047    return ConstantInt::get(Ty, C0->shl(ShlAmt) | C1->lshr(LshrAmt));4048  }4049 4050  if (IntrinsicID == Intrinsic::amdgcn_perm)4051    return ConstantFoldAMDGCNPermIntrinsic(Operands, Ty);4052 4053  return nullptr;4054}4055 4056static Constant *ConstantFoldScalarCall(StringRef Name,4057                                        Intrinsic::ID IntrinsicID,4058                                        Type *Ty,4059                                        ArrayRef<Constant *> Operands,4060                                        const TargetLibraryInfo *TLI,4061                                        const CallBase *Call) {4062  if (IntrinsicID != Intrinsic::not_intrinsic &&4063      any_of(Operands, IsaPred<PoisonValue>) &&4064      intrinsicPropagatesPoison(IntrinsicID))4065    return PoisonValue::get(Ty);4066 4067  if (Operands.size() == 1)4068    return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call);4069 4070  if (Operands.size() == 2) {4071    if (Constant *FoldedLibCall =4072            ConstantFoldLibCall2(Name, Ty, Operands, TLI)) {4073      return FoldedLibCall;4074    }4075    return ConstantFoldIntrinsicCall2(IntrinsicID, Ty, Operands, Call);4076  }4077 4078  if (Operands.size() == 3)4079    return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call);4080 4081  return nullptr;4082}4083 4084static Constant *ConstantFoldFixedVectorCall(4085    StringRef Name, Intrinsic::ID IntrinsicID, FixedVectorType *FVTy,4086    ArrayRef<Constant *> Operands, const DataLayout &DL,4087    const TargetLibraryInfo *TLI, const CallBase *Call) {4088  SmallVector<Constant *, 4> Result(FVTy->getNumElements());4089  SmallVector<Constant *, 4> Lane(Operands.size());4090  Type *Ty = FVTy->getElementType();4091 4092  switch (IntrinsicID) {4093  case Intrinsic::masked_load: {4094    auto *SrcPtr = Operands[0];4095    auto *Mask = Operands[1];4096    auto *Passthru = Operands[2];4097 4098    Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, FVTy, DL);4099 4100    SmallVector<Constant *, 32> NewElements;4101    for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {4102      auto *MaskElt = Mask->getAggregateElement(I);4103      if (!MaskElt)4104        break;4105      auto *PassthruElt = Passthru->getAggregateElement(I);4106      auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr;4107      if (isa<UndefValue>(MaskElt)) {4108        if (PassthruElt)4109          NewElements.push_back(PassthruElt);4110        else if (VecElt)4111          NewElements.push_back(VecElt);4112        else4113          return nullptr;4114      }4115      if (MaskElt->isNullValue()) {4116        if (!PassthruElt)4117          return nullptr;4118        NewElements.push_back(PassthruElt);4119      } else if (MaskElt->isOneValue()) {4120        if (!VecElt)4121          return nullptr;4122        NewElements.push_back(VecElt);4123      } else {4124        return nullptr;4125      }4126    }4127    if (NewElements.size() != FVTy->getNumElements())4128      return nullptr;4129    return ConstantVector::get(NewElements);4130  }4131  case Intrinsic::arm_mve_vctp8:4132  case Intrinsic::arm_mve_vctp16:4133  case Intrinsic::arm_mve_vctp32:4134  case Intrinsic::arm_mve_vctp64: {4135    if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {4136      unsigned Lanes = FVTy->getNumElements();4137      uint64_t Limit = Op->getZExtValue();4138 4139      SmallVector<Constant *, 16> NCs;4140      for (unsigned i = 0; i < Lanes; i++) {4141        if (i < Limit)4142          NCs.push_back(ConstantInt::getTrue(Ty));4143        else4144          NCs.push_back(ConstantInt::getFalse(Ty));4145      }4146      return ConstantVector::get(NCs);4147    }4148    return nullptr;4149  }4150  case Intrinsic::get_active_lane_mask: {4151    auto *Op0 = dyn_cast<ConstantInt>(Operands[0]);4152    auto *Op1 = dyn_cast<ConstantInt>(Operands[1]);4153    if (Op0 && Op1) {4154      unsigned Lanes = FVTy->getNumElements();4155      uint64_t Base = Op0->getZExtValue();4156      uint64_t Limit = Op1->getZExtValue();4157 4158      SmallVector<Constant *, 16> NCs;4159      for (unsigned i = 0; i < Lanes; i++) {4160        if (Base + i < Limit)4161          NCs.push_back(ConstantInt::getTrue(Ty));4162        else4163          NCs.push_back(ConstantInt::getFalse(Ty));4164      }4165      return ConstantVector::get(NCs);4166    }4167    return nullptr;4168  }4169  case Intrinsic::vector_extract: {4170    auto *Idx = dyn_cast<ConstantInt>(Operands[1]);4171    Constant *Vec = Operands[0];4172    if (!Idx || !isa<FixedVectorType>(Vec->getType()))4173      return nullptr;4174 4175    unsigned NumElements = FVTy->getNumElements();4176    unsigned VecNumElements =4177        cast<FixedVectorType>(Vec->getType())->getNumElements();4178    unsigned StartingIndex = Idx->getZExtValue();4179 4180    // Extracting entire vector is nop4181    if (NumElements == VecNumElements && StartingIndex == 0)4182      return Vec;4183 4184    for (unsigned I = StartingIndex, E = StartingIndex + NumElements; I < E;4185         ++I) {4186      Constant *Elt = Vec->getAggregateElement(I);4187      if (!Elt)4188        return nullptr;4189      Result[I - StartingIndex] = Elt;4190    }4191 4192    return ConstantVector::get(Result);4193  }4194  case Intrinsic::vector_insert: {4195    Constant *Vec = Operands[0];4196    Constant *SubVec = Operands[1];4197    auto *Idx = dyn_cast<ConstantInt>(Operands[2]);4198    if (!Idx || !isa<FixedVectorType>(Vec->getType()))4199      return nullptr;4200 4201    unsigned SubVecNumElements =4202        cast<FixedVectorType>(SubVec->getType())->getNumElements();4203    unsigned VecNumElements =4204        cast<FixedVectorType>(Vec->getType())->getNumElements();4205    unsigned IdxN = Idx->getZExtValue();4206    // Replacing entire vector with a subvec is nop4207    if (SubVecNumElements == VecNumElements && IdxN == 0)4208      return SubVec;4209 4210    for (unsigned I = 0; I < VecNumElements; ++I) {4211      Constant *Elt;4212      if (I < IdxN + SubVecNumElements)4213        Elt = SubVec->getAggregateElement(I - IdxN);4214      else4215        Elt = Vec->getAggregateElement(I);4216      if (!Elt)4217        return nullptr;4218      Result[I] = Elt;4219    }4220    return ConstantVector::get(Result);4221  }4222  case Intrinsic::vector_interleave2:4223  case Intrinsic::vector_interleave3:4224  case Intrinsic::vector_interleave4:4225  case Intrinsic::vector_interleave5:4226  case Intrinsic::vector_interleave6:4227  case Intrinsic::vector_interleave7:4228  case Intrinsic::vector_interleave8: {4229    unsigned NumElements =4230        cast<FixedVectorType>(Operands[0]->getType())->getNumElements();4231    unsigned NumOperands = Operands.size();4232    for (unsigned I = 0; I < NumElements; ++I) {4233      for (unsigned J = 0; J < NumOperands; ++J) {4234        Constant *Elt = Operands[J]->getAggregateElement(I);4235        if (!Elt)4236          return nullptr;4237        Result[NumOperands * I + J] = Elt;4238      }4239    }4240    return ConstantVector::get(Result);4241  }4242  case Intrinsic::wasm_dot: {4243    unsigned NumElements =4244        cast<FixedVectorType>(Operands[0]->getType())->getNumElements();4245 4246    assert(NumElements == 8 && Result.size() == 4 &&4247           "wasm dot takes i16x8 and produces i32x4");4248    assert(Ty->isIntegerTy());4249    int32_t MulVector[8];4250 4251    for (unsigned I = 0; I < NumElements; ++I) {4252      ConstantInt *Elt0 =4253          cast<ConstantInt>(Operands[0]->getAggregateElement(I));4254      ConstantInt *Elt1 =4255          cast<ConstantInt>(Operands[1]->getAggregateElement(I));4256 4257      MulVector[I] = Elt0->getSExtValue() * Elt1->getSExtValue();4258    }4259    for (unsigned I = 0; I < Result.size(); I++) {4260      int64_t IAdd = (int64_t)MulVector[I * 2] + (int64_t)MulVector[I * 2 + 1];4261      Result[I] = ConstantInt::get(Ty, IAdd);4262    }4263 4264    return ConstantVector::get(Result);4265  }4266  default:4267    break;4268  }4269 4270  for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {4271    // Gather a column of constants.4272    for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {4273      // Some intrinsics use a scalar type for certain arguments.4274      if (isVectorIntrinsicWithScalarOpAtArg(IntrinsicID, J, /*TTI=*/nullptr)) {4275        Lane[J] = Operands[J];4276        continue;4277      }4278 4279      Constant *Agg = Operands[J]->getAggregateElement(I);4280      if (!Agg)4281        return nullptr;4282 4283      Lane[J] = Agg;4284    }4285 4286    // Use the regular scalar folding to simplify this column.4287    Constant *Folded =4288        ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI, Call);4289    if (!Folded)4290      return nullptr;4291    Result[I] = Folded;4292  }4293 4294  return ConstantVector::get(Result);4295}4296 4297static Constant *ConstantFoldScalableVectorCall(4298    StringRef Name, Intrinsic::ID IntrinsicID, ScalableVectorType *SVTy,4299    ArrayRef<Constant *> Operands, const DataLayout &DL,4300    const TargetLibraryInfo *TLI, const CallBase *Call) {4301  switch (IntrinsicID) {4302  case Intrinsic::aarch64_sve_convert_from_svbool: {4303    auto *Src = dyn_cast<Constant>(Operands[0]);4304    if (!Src || !Src->isNullValue())4305      break;4306 4307    return ConstantInt::getFalse(SVTy);4308  }4309  case Intrinsic::get_active_lane_mask: {4310    auto *Op0 = dyn_cast<ConstantInt>(Operands[0]);4311    auto *Op1 = dyn_cast<ConstantInt>(Operands[1]);4312    if (Op0 && Op1 && Op0->getValue().uge(Op1->getValue()))4313      return ConstantVector::getNullValue(SVTy);4314    break;4315  }4316  case Intrinsic::vector_interleave2:4317  case Intrinsic::vector_interleave3:4318  case Intrinsic::vector_interleave4:4319  case Intrinsic::vector_interleave5:4320  case Intrinsic::vector_interleave6:4321  case Intrinsic::vector_interleave7:4322  case Intrinsic::vector_interleave8: {4323    Constant *SplatVal = Operands[0]->getSplatValue();4324    if (!SplatVal)4325      return nullptr;4326 4327    if (!llvm::all_equal(Operands))4328      return nullptr;4329 4330    return ConstantVector::getSplat(SVTy->getElementCount(), SplatVal);4331  }4332  default:4333    break;4334  }4335 4336  // If trivially vectorizable, try folding it via the scalar call if all4337  // operands are splats.4338 4339  // TODO: ConstantFoldFixedVectorCall should probably check this too?4340  if (!isTriviallyVectorizable(IntrinsicID))4341    return nullptr;4342 4343  SmallVector<Constant *, 4> SplatOps;4344  for (auto [I, Op] : enumerate(Operands)) {4345    if (isVectorIntrinsicWithScalarOpAtArg(IntrinsicID, I, /*TTI=*/nullptr)) {4346      SplatOps.push_back(Op);4347      continue;4348    }4349    Constant *Splat = Op->getSplatValue();4350    if (!Splat)4351      return nullptr;4352    SplatOps.push_back(Splat);4353  }4354  Constant *Folded = ConstantFoldScalarCall(4355      Name, IntrinsicID, SVTy->getElementType(), SplatOps, TLI, Call);4356  if (!Folded)4357    return nullptr;4358  return ConstantVector::getSplat(SVTy->getElementCount(), Folded);4359}4360 4361static std::pair<Constant *, Constant *>4362ConstantFoldScalarFrexpCall(Constant *Op, Type *IntTy) {4363  if (isa<PoisonValue>(Op))4364    return {Op, PoisonValue::get(IntTy)};4365 4366  auto *ConstFP = dyn_cast<ConstantFP>(Op);4367  if (!ConstFP)4368    return {};4369 4370  const APFloat &U = ConstFP->getValueAPF();4371  int FrexpExp;4372  APFloat FrexpMant = frexp(U, FrexpExp, APFloat::rmNearestTiesToEven);4373  Constant *Result0 = ConstantFP::get(ConstFP->getType(), FrexpMant);4374 4375  // The exponent is an "unspecified value" for inf/nan. We use zero to avoid4376  // using undef.4377  Constant *Result1 = FrexpMant.isFinite()4378                          ? ConstantInt::getSigned(IntTy, FrexpExp)4379                          : ConstantInt::getNullValue(IntTy);4380  return {Result0, Result1};4381}4382 4383/// Handle intrinsics that return tuples, which may be tuples of vectors.4384static Constant *4385ConstantFoldStructCall(StringRef Name, Intrinsic::ID IntrinsicID,4386                       StructType *StTy, ArrayRef<Constant *> Operands,4387                       const DataLayout &DL, const TargetLibraryInfo *TLI,4388                       const CallBase *Call) {4389 4390  switch (IntrinsicID) {4391  case Intrinsic::frexp: {4392    Type *Ty0 = StTy->getContainedType(0);4393    Type *Ty1 = StTy->getContainedType(1)->getScalarType();4394 4395    if (auto *FVTy0 = dyn_cast<FixedVectorType>(Ty0)) {4396      SmallVector<Constant *, 4> Results0(FVTy0->getNumElements());4397      SmallVector<Constant *, 4> Results1(FVTy0->getNumElements());4398 4399      for (unsigned I = 0, E = FVTy0->getNumElements(); I != E; ++I) {4400        Constant *Lane = Operands[0]->getAggregateElement(I);4401        std::tie(Results0[I], Results1[I]) =4402            ConstantFoldScalarFrexpCall(Lane, Ty1);4403        if (!Results0[I])4404          return nullptr;4405      }4406 4407      return ConstantStruct::get(StTy, ConstantVector::get(Results0),4408                                 ConstantVector::get(Results1));4409    }4410 4411    auto [Result0, Result1] = ConstantFoldScalarFrexpCall(Operands[0], Ty1);4412    if (!Result0)4413      return nullptr;4414    return ConstantStruct::get(StTy, Result0, Result1);4415  }4416  case Intrinsic::sincos: {4417    Type *Ty = StTy->getContainedType(0);4418    Type *TyScalar = Ty->getScalarType();4419 4420    auto ConstantFoldScalarSincosCall =4421        [&](Constant *Op) -> std::pair<Constant *, Constant *> {4422      Constant *SinResult =4423          ConstantFoldScalarCall(Name, Intrinsic::sin, TyScalar, Op, TLI, Call);4424      Constant *CosResult =4425          ConstantFoldScalarCall(Name, Intrinsic::cos, TyScalar, Op, TLI, Call);4426      return std::make_pair(SinResult, CosResult);4427    };4428 4429    if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) {4430      SmallVector<Constant *> SinResults(FVTy->getNumElements());4431      SmallVector<Constant *> CosResults(FVTy->getNumElements());4432 4433      for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {4434        Constant *Lane = Operands[0]->getAggregateElement(I);4435        std::tie(SinResults[I], CosResults[I]) =4436            ConstantFoldScalarSincosCall(Lane);4437        if (!SinResults[I] || !CosResults[I])4438          return nullptr;4439      }4440 4441      return ConstantStruct::get(StTy, ConstantVector::get(SinResults),4442                                 ConstantVector::get(CosResults));4443    }4444 4445    auto [SinResult, CosResult] = ConstantFoldScalarSincosCall(Operands[0]);4446    if (!SinResult || !CosResult)4447      return nullptr;4448    return ConstantStruct::get(StTy, SinResult, CosResult);4449  }4450  case Intrinsic::vector_deinterleave2:4451  case Intrinsic::vector_deinterleave3:4452  case Intrinsic::vector_deinterleave4:4453  case Intrinsic::vector_deinterleave5:4454  case Intrinsic::vector_deinterleave6:4455  case Intrinsic::vector_deinterleave7:4456  case Intrinsic::vector_deinterleave8: {4457    unsigned NumResults = StTy->getNumElements();4458    auto *Vec = Operands[0];4459    auto *VecTy = cast<VectorType>(Vec->getType());4460 4461    ElementCount ResultEC =4462        VecTy->getElementCount().divideCoefficientBy(NumResults);4463 4464    if (auto *EltC = Vec->getSplatValue()) {4465      auto *ResultVec = ConstantVector::getSplat(ResultEC, EltC);4466      SmallVector<Constant *, 8> Results(NumResults, ResultVec);4467      return ConstantStruct::get(StTy, Results);4468    }4469 4470    if (!ResultEC.isFixed())4471      return nullptr;4472 4473    unsigned NumElements = ResultEC.getFixedValue();4474    SmallVector<Constant *, 8> Results(NumResults);4475    SmallVector<Constant *> Elements(NumElements);4476    for (unsigned I = 0; I != NumResults; ++I) {4477      for (unsigned J = 0; J != NumElements; ++J) {4478        Constant *Elt = Vec->getAggregateElement(J * NumResults + I);4479        if (!Elt)4480          return nullptr;4481        Elements[J] = Elt;4482      }4483      Results[I] = ConstantVector::get(Elements);4484    }4485    return ConstantStruct::get(StTy, Results);4486  }4487  default:4488    // TODO: Constant folding of vector intrinsics that fall through here does4489    // not work (e.g. overflow intrinsics)4490    return ConstantFoldScalarCall(Name, IntrinsicID, StTy, Operands, TLI, Call);4491  }4492 4493  return nullptr;4494}4495 4496} // end anonymous namespace4497 4498Constant *llvm::ConstantFoldBinaryIntrinsic(Intrinsic::ID ID, Constant *LHS,4499                                            Constant *RHS, Type *Ty,4500                                            Instruction *FMFSource) {4501  auto *Call = dyn_cast_if_present<CallBase>(FMFSource);4502  // Ensure we check flags like StrictFP that might prevent this from getting4503  // folded before generating a result.4504  if (Call && !canConstantFoldCallTo(Call, Call->getCalledFunction()))4505    return nullptr;4506  return ConstantFoldIntrinsicCall2(ID, Ty, {LHS, RHS}, Call);4507}4508 4509Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F,4510                                 ArrayRef<Constant *> Operands,4511                                 const TargetLibraryInfo *TLI,4512                                 bool AllowNonDeterministic) {4513  if (Call->isNoBuiltin())4514    return nullptr;4515  if (!F->hasName())4516    return nullptr;4517 4518  // If this is not an intrinsic and not recognized as a library call, bail out.4519  Intrinsic::ID IID = F->getIntrinsicID();4520  if (IID == Intrinsic::not_intrinsic) {4521    if (!TLI)4522      return nullptr;4523    LibFunc LibF;4524    if (!TLI->getLibFunc(*F, LibF))4525      return nullptr;4526  }4527 4528  // Conservatively assume that floating-point libcalls may be4529  // non-deterministic.4530  Type *Ty = F->getReturnType();4531  if (!AllowNonDeterministic && Ty->isFPOrFPVectorTy())4532    return nullptr;4533 4534  StringRef Name = F->getName();4535  if (auto *FVTy = dyn_cast<FixedVectorType>(Ty))4536    return ConstantFoldFixedVectorCall(4537        Name, IID, FVTy, Operands, F->getDataLayout(), TLI, Call);4538 4539  if (auto *SVTy = dyn_cast<ScalableVectorType>(Ty))4540    return ConstantFoldScalableVectorCall(4541        Name, IID, SVTy, Operands, F->getDataLayout(), TLI, Call);4542 4543  if (auto *StTy = dyn_cast<StructType>(Ty))4544    return ConstantFoldStructCall(Name, IID, StTy, Operands,4545                                  F->getDataLayout(), TLI, Call);4546 4547  // TODO: If this is a library function, we already discovered that above,4548  //       so we should pass the LibFunc, not the name (and it might be better4549  //       still to separate intrinsic handling from libcalls).4550  return ConstantFoldScalarCall(Name, IID, Ty, Operands, TLI, Call);4551}4552 4553bool llvm::isMathLibCallNoop(const CallBase *Call,4554                             const TargetLibraryInfo *TLI) {4555  // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap4556  // (and to some extent ConstantFoldScalarCall).4557  if (Call->isNoBuiltin() || Call->isStrictFP())4558    return false;4559  Function *F = Call->getCalledFunction();4560  if (!F)4561    return false;4562 4563  LibFunc Func;4564  if (!TLI || !TLI->getLibFunc(*F, Func))4565    return false;4566 4567  if (Call->arg_size() == 1) {4568    if (ConstantFP *OpC = dyn_cast<ConstantFP>(Call->getArgOperand(0))) {4569      const APFloat &Op = OpC->getValueAPF();4570      switch (Func) {4571      case LibFunc_logl:4572      case LibFunc_log:4573      case LibFunc_logf:4574      case LibFunc_log2l:4575      case LibFunc_log2:4576      case LibFunc_log2f:4577      case LibFunc_log10l:4578      case LibFunc_log10:4579      case LibFunc_log10f:4580        return Op.isNaN() || (!Op.isZero() && !Op.isNegative());4581 4582      case LibFunc_ilogb:4583        return !Op.isNaN() && !Op.isZero() && !Op.isInfinity();4584 4585      case LibFunc_expl:4586      case LibFunc_exp:4587      case LibFunc_expf:4588        // FIXME: These boundaries are slightly conservative.4589        if (OpC->getType()->isDoubleTy())4590          return !(Op < APFloat(-745.0) || Op > APFloat(709.0));4591        if (OpC->getType()->isFloatTy())4592          return !(Op < APFloat(-103.0f) || Op > APFloat(88.0f));4593        break;4594 4595      case LibFunc_exp2l:4596      case LibFunc_exp2:4597      case LibFunc_exp2f:4598        // FIXME: These boundaries are slightly conservative.4599        if (OpC->getType()->isDoubleTy())4600          return !(Op < APFloat(-1074.0) || Op > APFloat(1023.0));4601        if (OpC->getType()->isFloatTy())4602          return !(Op < APFloat(-149.0f) || Op > APFloat(127.0f));4603        break;4604 4605      case LibFunc_sinl:4606      case LibFunc_sin:4607      case LibFunc_sinf:4608      case LibFunc_cosl:4609      case LibFunc_cos:4610      case LibFunc_cosf:4611        return !Op.isInfinity();4612 4613      case LibFunc_tanl:4614      case LibFunc_tan:4615      case LibFunc_tanf: {4616        // FIXME: Stop using the host math library.4617        // FIXME: The computation isn't done in the right precision.4618        Type *Ty = OpC->getType();4619        if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy())4620          return ConstantFoldFP(tan, OpC->getValueAPF(), Ty) != nullptr;4621        break;4622      }4623 4624      case LibFunc_atan:4625      case LibFunc_atanf:4626      case LibFunc_atanl:4627        // Per POSIX, this MAY fail if Op is denormal. We choose not failing.4628        return true;4629 4630      case LibFunc_asinl:4631      case LibFunc_asin:4632      case LibFunc_asinf:4633      case LibFunc_acosl:4634      case LibFunc_acos:4635      case LibFunc_acosf:4636        return !(Op < APFloat::getOne(Op.getSemantics(), true) ||4637                 Op > APFloat::getOne(Op.getSemantics()));4638 4639      case LibFunc_sinh:4640      case LibFunc_cosh:4641      case LibFunc_sinhf:4642      case LibFunc_coshf:4643      case LibFunc_sinhl:4644      case LibFunc_coshl:4645        // FIXME: These boundaries are slightly conservative.4646        if (OpC->getType()->isDoubleTy())4647          return !(Op < APFloat(-710.0) || Op > APFloat(710.0));4648        if (OpC->getType()->isFloatTy())4649          return !(Op < APFloat(-89.0f) || Op > APFloat(89.0f));4650        break;4651 4652      case LibFunc_sqrtl:4653      case LibFunc_sqrt:4654      case LibFunc_sqrtf:4655        return Op.isNaN() || Op.isZero() || !Op.isNegative();4656 4657      // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,4658      // maybe others?4659      default:4660        break;4661      }4662    }4663  }4664 4665  if (Call->arg_size() == 2) {4666    ConstantFP *Op0C = dyn_cast<ConstantFP>(Call->getArgOperand(0));4667    ConstantFP *Op1C = dyn_cast<ConstantFP>(Call->getArgOperand(1));4668    if (Op0C && Op1C) {4669      const APFloat &Op0 = Op0C->getValueAPF();4670      const APFloat &Op1 = Op1C->getValueAPF();4671 4672      switch (Func) {4673      case LibFunc_powl:4674      case LibFunc_pow:4675      case LibFunc_powf: {4676        // FIXME: Stop using the host math library.4677        // FIXME: The computation isn't done in the right precision.4678        Type *Ty = Op0C->getType();4679        if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {4680          if (Ty == Op1C->getType())4681            return ConstantFoldBinaryFP(pow, Op0, Op1, Ty) != nullptr;4682        }4683        break;4684      }4685 4686      case LibFunc_fmodl:4687      case LibFunc_fmod:4688      case LibFunc_fmodf:4689      case LibFunc_remainderl:4690      case LibFunc_remainder:4691      case LibFunc_remainderf:4692        return Op0.isNaN() || Op1.isNaN() ||4693               (!Op0.isInfinity() && !Op1.isZero());4694 4695      case LibFunc_atan2:4696      case LibFunc_atan2f:4697      case LibFunc_atan2l:4698        // Although IEEE-754 says atan2(+/-0.0, +/-0.0) are well-defined, and4699        // GLIBC and MSVC do not appear to raise an error on those, we4700        // cannot rely on that behavior. POSIX and C11 say that a domain error4701        // may occur, so allow for that possibility.4702        return !Op0.isZero() || !Op1.isZero();4703 4704      default:4705        break;4706      }4707    }4708  }4709 4710  return false;4711}4712 4713Constant *llvm::getLosslessInvCast(Constant *C, Type *InvCastTo,4714                                   unsigned CastOp, const DataLayout &DL,4715                                   PreservedCastFlags *Flags) {4716  switch (CastOp) {4717  case Instruction::BitCast:4718    // Bitcast is always lossless.4719    return ConstantFoldCastOperand(Instruction::BitCast, C, InvCastTo, DL);4720  case Instruction::Trunc: {4721    auto *ZExtC = ConstantFoldCastOperand(Instruction::ZExt, C, InvCastTo, DL);4722    if (Flags) {4723      // Truncation back on ZExt value is always NUW.4724      Flags->NUW = true;4725      // Test positivity of C.4726      auto *SExtC =4727          ConstantFoldCastOperand(Instruction::SExt, C, InvCastTo, DL);4728      Flags->NSW = ZExtC == SExtC;4729    }4730    return ZExtC;4731  }4732  case Instruction::SExt:4733  case Instruction::ZExt: {4734    auto *InvC = ConstantExpr::getTrunc(C, InvCastTo);4735    auto *CastInvC = ConstantFoldCastOperand(CastOp, InvC, C->getType(), DL);4736    // Must satisfy CastOp(InvC) == C.4737    if (!CastInvC || CastInvC != C)4738      return nullptr;4739    if (Flags && CastOp == Instruction::ZExt) {4740      auto *SExtInvC =4741          ConstantFoldCastOperand(Instruction::SExt, InvC, C->getType(), DL);4742      // Test positivity of InvC.4743      Flags->NNeg = CastInvC == SExtInvC;4744    }4745    return InvC;4746  }4747  default:4748    return nullptr;4749  }4750}4751 4752Constant *llvm::getLosslessUnsignedTrunc(Constant *C, Type *DestTy,4753                                         const DataLayout &DL,4754                                         PreservedCastFlags *Flags) {4755  return getLosslessInvCast(C, DestTy, Instruction::ZExt, DL, Flags);4756}4757 4758Constant *llvm::getLosslessSignedTrunc(Constant *C, Type *DestTy,4759                                       const DataLayout &DL,4760                                       PreservedCastFlags *Flags) {4761  return getLosslessInvCast(C, DestTy, Instruction::SExt, DL, Flags);4762}4763 4764void TargetFolder::anchor() {}4765