brintos / llvm-project-archived public Read only

1//===- InstCombineAndOrXor.cpp --------------------------------------------===//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 implements the visitAnd, visitOr, and visitXor functions.10//11//===----------------------------------------------------------------------===//12 13#include "InstCombineInternal.h"14#include "llvm/ADT/SmallBitVector.h"15#include "llvm/Analysis/CmpInstAnalysis.h"16#include "llvm/Analysis/FloatingPointPredicateUtils.h"17#include "llvm/Analysis/InstructionSimplify.h"18#include "llvm/IR/ConstantRange.h"19#include "llvm/IR/DerivedTypes.h"20#include "llvm/IR/Instructions.h"21#include "llvm/IR/Intrinsics.h"22#include "llvm/IR/PatternMatch.h"23#include "llvm/Transforms/InstCombine/InstCombiner.h"24#include "llvm/Transforms/Utils/Local.h"25 26using namespace llvm;27using namespace PatternMatch;28 29#define DEBUG_TYPE "instcombine"30 31namespace llvm {32extern cl::opt<bool> ProfcheckDisableMetadataFixes;33}34 35/// This is the complement of getICmpCode, which turns an opcode and two36/// operands into either a constant true or false, or a brand new ICmp37/// instruction. The sign is passed in to determine which kind of predicate to38/// use in the new icmp instruction.39static Value *getNewICmpValue(unsigned Code, bool Sign, Value *LHS, Value *RHS,40                              InstCombiner::BuilderTy &Builder) {41  ICmpInst::Predicate NewPred;42  if (Constant *TorF = getPredForICmpCode(Code, Sign, LHS->getType(), NewPred))43    return TorF;44  return Builder.CreateICmp(NewPred, LHS, RHS);45}46 47/// This is the complement of getFCmpCode, which turns an opcode and two48/// operands into either a FCmp instruction, or a true/false constant.49static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS,50                           InstCombiner::BuilderTy &Builder, FMFSource FMF) {51  FCmpInst::Predicate NewPred;52  if (Constant *TorF = getPredForFCmpCode(Code, LHS->getType(), NewPred))53    return TorF;54  return Builder.CreateFCmpFMF(NewPred, LHS, RHS, FMF);55}56 57/// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise58/// (V < Lo || V >= Hi). This method expects that Lo < Hi. IsSigned indicates59/// whether to treat V, Lo, and Hi as signed or not.60Value *InstCombinerImpl::insertRangeTest(Value *V, const APInt &Lo,61                                         const APInt &Hi, bool isSigned,62                                         bool Inside) {63  assert((isSigned ? Lo.slt(Hi) : Lo.ult(Hi)) &&64         "Lo is not < Hi in range emission code!");65 66  Type *Ty = V->getType();67 68  // V >= Min && V <  Hi --> V <  Hi69  // V <  Min || V >= Hi --> V >= Hi70  ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;71  if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) {72    Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred;73    return Builder.CreateICmp(Pred, V, ConstantInt::get(Ty, Hi));74  }75 76  // V >= Lo && V <  Hi --> V - Lo u<  Hi - Lo77  // V <  Lo || V >= Hi --> V - Lo u>= Hi - Lo78  Value *VMinusLo =79      Builder.CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off");80  Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo);81  return Builder.CreateICmp(Pred, VMinusLo, HiMinusLo);82}83 84/// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns85/// that can be simplified.86/// One of A and B is considered the mask. The other is the value. This is87/// described as the "AMask" or "BMask" part of the enum. If the enum contains88/// only "Mask", then both A and B can be considered masks. If A is the mask,89/// then it was proven that (A & C) == C. This is trivial if C == A or C == 0.90/// If both A and C are constants, this proof is also easy.91/// For the following explanations, we assume that A is the mask.92///93/// "AllOnes" declares that the comparison is true only if (A & B) == A or all94/// bits of A are set in B.95///   Example: (icmp eq (A & 3), 3) -> AMask_AllOnes96///97/// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all98/// bits of A are cleared in B.99///   Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes100///101/// "Mixed" declares that (A & B) == C and C might or might not contain any102/// number of one bits and zero bits.103///   Example: (icmp eq (A & 3), 1) -> AMask_Mixed104///105/// "Not" means that in above descriptions "==" should be replaced by "!=".106///   Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes107///108/// If the mask A contains a single bit, then the following is equivalent:109///    (icmp eq (A & B), A) equals (icmp ne (A & B), 0)110///    (icmp ne (A & B), A) equals (icmp eq (A & B), 0)111enum MaskedICmpType {112  AMask_AllOnes           =     1,113  AMask_NotAllOnes        =     2,114  BMask_AllOnes           =     4,115  BMask_NotAllOnes        =     8,116  Mask_AllZeros           =    16,117  Mask_NotAllZeros        =    32,118  AMask_Mixed             =    64,119  AMask_NotMixed          =   128,120  BMask_Mixed             =   256,121  BMask_NotMixed          =   512122};123 124/// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C)125/// satisfies.126static unsigned getMaskedICmpType(Value *A, Value *B, Value *C,127                                  ICmpInst::Predicate Pred) {128  const APInt *ConstA = nullptr, *ConstB = nullptr, *ConstC = nullptr;129  match(A, m_APInt(ConstA));130  match(B, m_APInt(ConstB));131  match(C, m_APInt(ConstC));132  bool IsEq = (Pred == ICmpInst::ICMP_EQ);133  bool IsAPow2 = ConstA && ConstA->isPowerOf2();134  bool IsBPow2 = ConstB && ConstB->isPowerOf2();135  unsigned MaskVal = 0;136  if (ConstC && ConstC->isZero()) {137    // if C is zero, then both A and B qualify as mask138    MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed)139                     : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed));140    if (IsAPow2)141      MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed)142                       : (AMask_AllOnes | AMask_Mixed));143    if (IsBPow2)144      MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed)145                       : (BMask_AllOnes | BMask_Mixed));146    return MaskVal;147  }148 149  if (A == C) {150    MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed)151                     : (AMask_NotAllOnes | AMask_NotMixed));152    if (IsAPow2)153      MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed)154                       : (Mask_AllZeros | AMask_Mixed));155  } else if (ConstA && ConstC && ConstC->isSubsetOf(*ConstA)) {156    MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed);157  }158 159  if (B == C) {160    MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed)161                     : (BMask_NotAllOnes | BMask_NotMixed));162    if (IsBPow2)163      MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed)164                       : (Mask_AllZeros | BMask_Mixed));165  } else if (ConstB && ConstC && ConstC->isSubsetOf(*ConstB)) {166    MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed);167  }168 169  return MaskVal;170}171 172/// Convert an analysis of a masked ICmp into its equivalent if all boolean173/// operations had the opposite sense. Since each "NotXXX" flag (recording !=)174/// is adjacent to the corresponding normal flag (recording ==), this just175/// involves swapping those bits over.176static unsigned conjugateICmpMask(unsigned Mask) {177  unsigned NewMask;178  NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros |179                     AMask_Mixed | BMask_Mixed))180            << 1;181 182  NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros |183                      AMask_NotMixed | BMask_NotMixed))184             >> 1;185 186  return NewMask;187}188 189// Adapts the external decomposeBitTestICmp for local use.190static bool decomposeBitTestICmp(Value *Cond, CmpInst::Predicate &Pred,191                                 Value *&X, Value *&Y, Value *&Z) {192  auto Res = llvm::decomposeBitTest(Cond, /*LookThroughTrunc=*/true,193                                    /*AllowNonZeroC=*/true);194  if (!Res)195    return false;196 197  Pred = Res->Pred;198  X = Res->X;199  Y = ConstantInt::get(X->getType(), Res->Mask);200  Z = ConstantInt::get(X->getType(), Res->C);201  return true;202}203 204/// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E).205/// Return the pattern classes (from MaskedICmpType) for the left hand side and206/// the right hand side as a pair.207/// LHS and RHS are the left hand side and the right hand side ICmps and PredL208/// and PredR are their predicates, respectively.209static std::optional<std::pair<unsigned, unsigned>>210getMaskedTypeForICmpPair(Value *&A, Value *&B, Value *&C, Value *&D, Value *&E,211                         Value *LHS, Value *RHS, ICmpInst::Predicate &PredL,212                         ICmpInst::Predicate &PredR) {213 214  // Here comes the tricky part:215  // LHS might be of the form L11 & L12 == X, X == L21 & L22,216  // and L11 & L12 == L21 & L22. The same goes for RHS.217  // Now we must find those components L** and R**, that are equal, so218  // that we can extract the parameters A, B, C, D, and E for the canonical219  // above.220 221  // Check whether the icmp can be decomposed into a bit test.222  Value *L1, *L11, *L12, *L2, *L21, *L22;223  if (decomposeBitTestICmp(LHS, PredL, L11, L12, L2)) {224    L21 = L22 = L1 = nullptr;225  } else {226    auto *LHSCMP = dyn_cast<ICmpInst>(LHS);227    if (!LHSCMP)228      return std::nullopt;229 230    // Don't allow pointers. Splat vectors are fine.231    if (!LHSCMP->getOperand(0)->getType()->isIntOrIntVectorTy())232      return std::nullopt;233 234    PredL = LHSCMP->getPredicate();235    L1 = LHSCMP->getOperand(0);236    L2 = LHSCMP->getOperand(1);237    // Look for ANDs in the LHS icmp.238    if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {239      // Any icmp can be viewed as being trivially masked; if it allows us to240      // remove one, it's worth it.241      L11 = L1;242      L12 = Constant::getAllOnesValue(L1->getType());243    }244 245    if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {246      L21 = L2;247      L22 = Constant::getAllOnesValue(L2->getType());248    }249  }250 251  // Bail if LHS was a icmp that can't be decomposed into an equality.252  if (!ICmpInst::isEquality(PredL))253    return std::nullopt;254 255  Value *R11, *R12, *R2;256  if (decomposeBitTestICmp(RHS, PredR, R11, R12, R2)) {257    if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {258      A = R11;259      D = R12;260    } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {261      A = R12;262      D = R11;263    } else {264      return std::nullopt;265    }266    E = R2;267  } else {268    auto *RHSCMP = dyn_cast<ICmpInst>(RHS);269    if (!RHSCMP)270      return std::nullopt;271    // Don't allow pointers. Splat vectors are fine.272    if (!RHSCMP->getOperand(0)->getType()->isIntOrIntVectorTy())273      return std::nullopt;274 275    PredR = RHSCMP->getPredicate();276 277    Value *R1 = RHSCMP->getOperand(0);278    R2 = RHSCMP->getOperand(1);279    bool Ok = false;280    if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {281      // As before, model no mask as a trivial mask if it'll let us do an282      // optimization.283      R11 = R1;284      R12 = Constant::getAllOnesValue(R1->getType());285    }286 287    if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {288      A = R11;289      D = R12;290      E = R2;291      Ok = true;292    } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {293      A = R12;294      D = R11;295      E = R2;296      Ok = true;297    }298 299    // Avoid matching against the -1 value we created for unmasked operand.300    if (Ok && match(A, m_AllOnes()))301      Ok = false;302 303    // Look for ANDs on the right side of the RHS icmp.304    if (!Ok) {305      if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {306        R11 = R2;307        R12 = Constant::getAllOnesValue(R2->getType());308      }309 310      if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {311        A = R11;312        D = R12;313        E = R1;314      } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {315        A = R12;316        D = R11;317        E = R1;318      } else {319        return std::nullopt;320      }321    }322  }323 324  // Bail if RHS was a icmp that can't be decomposed into an equality.325  if (!ICmpInst::isEquality(PredR))326    return std::nullopt;327 328  if (L11 == A) {329    B = L12;330    C = L2;331  } else if (L12 == A) {332    B = L11;333    C = L2;334  } else if (L21 == A) {335    B = L22;336    C = L1;337  } else if (L22 == A) {338    B = L21;339    C = L1;340  }341 342  unsigned LeftType = getMaskedICmpType(A, B, C, PredL);343  unsigned RightType = getMaskedICmpType(A, D, E, PredR);344  return std::optional<std::pair<unsigned, unsigned>>(345      std::make_pair(LeftType, RightType));346}347 348/// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single349/// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros350/// and the right hand side is of type BMask_Mixed. For example,351/// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8).352/// Also used for logical and/or, must be poison safe.353static Value *foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(354    Value *LHS, Value *RHS, bool IsAnd, Value *A, Value *B, Value *D, Value *E,355    ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,356    InstCombiner::BuilderTy &Builder) {357  // We are given the canonical form:358  //   (icmp ne (A & B), 0) & (icmp eq (A & D), E).359  // where D & E == E.360  //361  // If IsAnd is false, we get it in negated form:362  //   (icmp eq (A & B), 0) | (icmp ne (A & D), E) ->363  //      !((icmp ne (A & B), 0) & (icmp eq (A & D), E)).364  //365  // We currently handle the case of B, C, D, E are constant.366  //367  const APInt *BCst, *DCst, *OrigECst;368  if (!match(B, m_APInt(BCst)) || !match(D, m_APInt(DCst)) ||369      !match(E, m_APInt(OrigECst)))370    return nullptr;371 372  ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;373 374  // Update E to the canonical form when D is a power of two and RHS is375  // canonicalized as,376  // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or377  // (icmp ne (A & D), D) -> (icmp eq (A & D), 0).378  APInt ECst = *OrigECst;379  if (PredR != NewCC)380    ECst ^= *DCst;381 382  // If B or D is zero, skip because if LHS or RHS can be trivially folded by383  // other folding rules and this pattern won't apply any more.384  if (*BCst == 0 || *DCst == 0)385    return nullptr;386 387  // If B and D don't intersect, ie. (B & D) == 0, try to fold isNaN idiom:388  // (icmp ne (A & FractionBits), 0) & (icmp eq (A & ExpBits), ExpBits)389  // -> isNaN(A)390  // Otherwise, we cannot deduce anything from it.391  if (!BCst->intersects(*DCst)) {392    Value *Src;393    if (*DCst == ECst && match(A, m_ElementWiseBitCast(m_Value(Src))) &&394        !Builder.GetInsertBlock()->getParent()->hasFnAttribute(395            Attribute::StrictFP)) {396      Type *Ty = Src->getType()->getScalarType();397      if (!Ty->isIEEELikeFPTy())398        return nullptr;399 400      APInt ExpBits = APFloat::getInf(Ty->getFltSemantics()).bitcastToAPInt();401      if (ECst != ExpBits)402        return nullptr;403      APInt FractionBits = ~ExpBits;404      FractionBits.clearSignBit();405      if (*BCst != FractionBits)406        return nullptr;407 408      return Builder.CreateFCmp(IsAnd ? FCmpInst::FCMP_UNO : FCmpInst::FCMP_ORD,409                                Src, ConstantFP::getZero(Src->getType()));410    }411    return nullptr;412  }413 414  // If the following two conditions are met:415  //416  // 1. mask B covers only a single bit that's not covered by mask D, that is,417  // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of418  // B and D has only one bit set) and,419  //420  // 2. RHS (and E) indicates that the rest of B's bits are zero (in other421  // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0422  //423  // then that single bit in B must be one and thus the whole expression can be424  // folded to425  //   (A & (B | D)) == (B & (B ^ D)) | E.426  //427  // For example,428  // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9)429  // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8)430  if ((((*BCst & *DCst) & ECst) == 0) &&431      (*BCst & (*BCst ^ *DCst)).isPowerOf2()) {432    APInt BorD = *BCst | *DCst;433    APInt BandBxorDorE = (*BCst & (*BCst ^ *DCst)) | ECst;434    Value *NewMask = ConstantInt::get(A->getType(), BorD);435    Value *NewMaskedValue = ConstantInt::get(A->getType(), BandBxorDorE);436    Value *NewAnd = Builder.CreateAnd(A, NewMask);437    return Builder.CreateICmp(NewCC, NewAnd, NewMaskedValue);438  }439 440  auto IsSubSetOrEqual = [](const APInt *C1, const APInt *C2) {441    return (*C1 & *C2) == *C1;442  };443  auto IsSuperSetOrEqual = [](const APInt *C1, const APInt *C2) {444    return (*C1 & *C2) == *C2;445  };446 447  // In the following, we consider only the cases where B is a superset of D, B448  // is a subset of D, or B == D because otherwise there's at least one bit449  // covered by B but not D, in which case we can't deduce much from it, so450  // no folding (aside from the single must-be-one bit case right above.)451  // For example,452  // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding.453  if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst))454    return nullptr;455 456  // At this point, either B is a superset of D, B is a subset of D or B == D.457 458  // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict459  // and the whole expression becomes false (or true if negated), otherwise, no460  // folding.461  // For example,462  // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false.463  // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding.464  if (ECst.isZero()) {465    if (IsSubSetOrEqual(BCst, DCst))466      return ConstantInt::get(LHS->getType(), !IsAnd);467    return nullptr;468  }469 470  // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B ==471  // D. If B is a superset of (or equal to) D, since E is not zero, LHS is472  // subsumed by RHS (RHS implies LHS.) So the whole expression becomes473  // RHS. For example,474  // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).475  // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).476  if (IsSuperSetOrEqual(BCst, DCst)) {477    // We can't guarantee that samesign hold after this fold.478    if (auto *ICmp = dyn_cast<ICmpInst>(RHS))479      ICmp->setSameSign(false);480    return RHS;481  }482  // Otherwise, B is a subset of D. If B and E have a common bit set,483  // ie. (B & E) != 0, then LHS is subsumed by RHS. For example.484  // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).485  assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code");486  if ((*BCst & ECst) != 0) {487    // We can't guarantee that samesign hold after this fold.488    if (auto *ICmp = dyn_cast<ICmpInst>(RHS))489      ICmp->setSameSign(false);490    return RHS;491  }492  // Otherwise, LHS and RHS contradict and the whole expression becomes false493  // (or true if negated.) For example,494  // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false.495  // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false.496  return ConstantInt::get(LHS->getType(), !IsAnd);497}498 499/// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single500/// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side501/// aren't of the common mask pattern type.502/// Also used for logical and/or, must be poison safe.503static Value *foldLogOpOfMaskedICmpsAsymmetric(504    Value *LHS, Value *RHS, bool IsAnd, Value *A, Value *B, Value *C, Value *D,505    Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,506    unsigned LHSMask, unsigned RHSMask, InstCombiner::BuilderTy &Builder) {507  assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&508         "Expected equality predicates for masked type of icmps.");509  // Handle Mask_NotAllZeros-BMask_Mixed cases.510  // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or511  // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E)512  //    which gets swapped to513  //    (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C).514  if (!IsAnd) {515    LHSMask = conjugateICmpMask(LHSMask);516    RHSMask = conjugateICmpMask(RHSMask);517  }518  if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) {519    if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(520            LHS, RHS, IsAnd, A, B, D, E, PredL, PredR, Builder)) {521      return V;522    }523  } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) {524    if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(525            RHS, LHS, IsAnd, A, D, B, C, PredR, PredL, Builder)) {526      return V;527    }528  }529  return nullptr;530}531 532/// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)533/// into a single (icmp(A & X) ==/!= Y).534static Value *foldLogOpOfMaskedICmps(Value *LHS, Value *RHS, bool IsAnd,535                                     bool IsLogical,536                                     InstCombiner::BuilderTy &Builder,537                                     const SimplifyQuery &Q) {538  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;539  ICmpInst::Predicate PredL, PredR;540  std::optional<std::pair<unsigned, unsigned>> MaskPair =541      getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR);542  if (!MaskPair)543    return nullptr;544  assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&545         "Expected equality predicates for masked type of icmps.");546  unsigned LHSMask = MaskPair->first;547  unsigned RHSMask = MaskPair->second;548  unsigned Mask = LHSMask & RHSMask;549  if (Mask == 0) {550    // Even if the two sides don't share a common pattern, check if folding can551    // still happen.552    if (Value *V = foldLogOpOfMaskedICmpsAsymmetric(553            LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask,554            Builder))555      return V;556    return nullptr;557  }558 559  // In full generality:560  //     (icmp (A & B) Op C) | (icmp (A & D) Op E)561  // ==  ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]562  //563  // If the latter can be converted into (icmp (A & X) Op Y) then the former is564  // equivalent to (icmp (A & X) !Op Y).565  //566  // Therefore, we can pretend for the rest of this function that we're dealing567  // with the conjunction, provided we flip the sense of any comparisons (both568  // input and output).569 570  // In most cases we're going to produce an EQ for the "&&" case.571  ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;572  if (!IsAnd) {573    // Convert the masking analysis into its equivalent with negated574    // comparisons.575    Mask = conjugateICmpMask(Mask);576  }577 578  if (Mask & Mask_AllZeros) {579    // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)580    // -> (icmp eq (A & (B|D)), 0)581    if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D))582      return nullptr; // TODO: Use freeze?583    Value *NewOr = Builder.CreateOr(B, D);584    Value *NewAnd = Builder.CreateAnd(A, NewOr);585    // We can't use C as zero because we might actually handle586    //   (icmp ne (A & B), B) & (icmp ne (A & D), D)587    // with B and D, having a single bit set.588    Value *Zero = Constant::getNullValue(A->getType());589    return Builder.CreateICmp(NewCC, NewAnd, Zero);590  }591  if (Mask & BMask_AllOnes) {592    // (icmp eq (A & B), B) & (icmp eq (A & D), D)593    // -> (icmp eq (A & (B|D)), (B|D))594    if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D))595      return nullptr; // TODO: Use freeze?596    Value *NewOr = Builder.CreateOr(B, D);597    Value *NewAnd = Builder.CreateAnd(A, NewOr);598    return Builder.CreateICmp(NewCC, NewAnd, NewOr);599  }600  if (Mask & AMask_AllOnes) {601    // (icmp eq (A & B), A) & (icmp eq (A & D), A)602    // -> (icmp eq (A & (B&D)), A)603    if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D))604      return nullptr; // TODO: Use freeze?605    Value *NewAnd1 = Builder.CreateAnd(B, D);606    Value *NewAnd2 = Builder.CreateAnd(A, NewAnd1);607    return Builder.CreateICmp(NewCC, NewAnd2, A);608  }609 610  const APInt *ConstB, *ConstD;611  if (match(B, m_APInt(ConstB)) && match(D, m_APInt(ConstD))) {612    if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) {613      // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and614      // (icmp ne (A & B), B) & (icmp ne (A & D), D)615      //     -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)616      // Only valid if one of the masks is a superset of the other (check "B&D"617      // is the same as either B or D).618      APInt NewMask = *ConstB & *ConstD;619      if (NewMask == *ConstB)620        return LHS;621      if (NewMask == *ConstD) {622        if (IsLogical) {623          if (auto *RHSI = dyn_cast<Instruction>(RHS))624            RHSI->dropPoisonGeneratingFlags();625        }626        return RHS;627      }628    }629 630    if (Mask & AMask_NotAllOnes) {631      // (icmp ne (A & B), B) & (icmp ne (A & D), D)632      //     -> (icmp ne (A & B), A) or (icmp ne (A & D), A)633      // Only valid if one of the masks is a superset of the other (check "B|D"634      // is the same as either B or D).635      APInt NewMask = *ConstB | *ConstD;636      if (NewMask == *ConstB)637        return LHS;638      if (NewMask == *ConstD)639        return RHS;640    }641 642    if (Mask & (BMask_Mixed | BMask_NotMixed)) {643      // Mixed:644      // (icmp eq (A & B), C) & (icmp eq (A & D), E)645      // We already know that B & C == C && D & E == E.646      // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of647      // C and E, which are shared by both the mask B and the mask D, don't648      // contradict, then we can transform to649      // -> (icmp eq (A & (B|D)), (C|E))650      // Currently, we only handle the case of B, C, D, and E being constant.651      // We can't simply use C and E because we might actually handle652      //   (icmp ne (A & B), B) & (icmp eq (A & D), D)653      // with B and D, having a single bit set.654 655      // NotMixed:656      // (icmp ne (A & B), C) & (icmp ne (A & D), E)657      // -> (icmp ne (A & (B & D)), (C & E))658      // Check the intersection (B & D) for inequality.659      // Assume that (B & D) == B || (B & D) == D, i.e B/D is a subset of D/B660      // and (B & D) & (C ^ E) == 0, bits of C and E, which are shared by both661      // the B and the D, don't contradict. Note that we can assume (~B & C) ==662      // 0 && (~D & E) == 0, previous operation should delete these icmps if it663      // hadn't been met.664 665      const APInt *OldConstC, *OldConstE;666      if (!match(C, m_APInt(OldConstC)) || !match(E, m_APInt(OldConstE)))667        return nullptr;668 669      auto FoldBMixed = [&](ICmpInst::Predicate CC, bool IsNot) -> Value * {670        CC = IsNot ? CmpInst::getInversePredicate(CC) : CC;671        const APInt ConstC = PredL != CC ? *ConstB ^ *OldConstC : *OldConstC;672        const APInt ConstE = PredR != CC ? *ConstD ^ *OldConstE : *OldConstE;673 674        if (((*ConstB & *ConstD) & (ConstC ^ ConstE)).getBoolValue())675          return IsNot ? nullptr : ConstantInt::get(LHS->getType(), !IsAnd);676 677        if (IsNot && !ConstB->isSubsetOf(*ConstD) &&678            !ConstD->isSubsetOf(*ConstB))679          return nullptr;680 681        APInt BD, CE;682        if (IsNot) {683          BD = *ConstB & *ConstD;684          CE = ConstC & ConstE;685        } else {686          BD = *ConstB | *ConstD;687          CE = ConstC | ConstE;688        }689        Value *NewAnd = Builder.CreateAnd(A, BD);690        Value *CEVal = ConstantInt::get(A->getType(), CE);691        return Builder.CreateICmp(CC, NewAnd, CEVal);692      };693 694      if (Mask & BMask_Mixed)695        return FoldBMixed(NewCC, false);696      if (Mask & BMask_NotMixed) // can be else also697        return FoldBMixed(NewCC, true);698    }699  }700 701  // (icmp eq (A & B), 0) | (icmp eq (A & D), 0)702  // -> (icmp ne (A & (B|D)), (B|D))703  // (icmp ne (A & B), 0) & (icmp ne (A & D), 0)704  // -> (icmp eq (A & (B|D)), (B|D))705  // iff B and D is known to be a power of two706  if (Mask & Mask_NotAllZeros &&707      isKnownToBeAPowerOfTwo(B, /*OrZero=*/false, Q) &&708      isKnownToBeAPowerOfTwo(D, /*OrZero=*/false, Q)) {709    // If this is a logical and/or, then we must prevent propagation of a710    // poison value from the RHS by inserting freeze.711    if (IsLogical)712      D = Builder.CreateFreeze(D);713    Value *Mask = Builder.CreateOr(B, D);714    Value *Masked = Builder.CreateAnd(A, Mask);715    return Builder.CreateICmp(NewCC, Masked, Mask);716  }717  return nullptr;718}719 720/// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.721/// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n722/// If \p Inverted is true then the check is for the inverted range, e.g.723/// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n724Value *InstCombinerImpl::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,725                                            bool Inverted) {726  // Check the lower range comparison, e.g. x >= 0727  // InstCombine already ensured that if there is a constant it's on the RHS.728  ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));729  if (!RangeStart)730    return nullptr;731 732  ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :733                               Cmp0->getPredicate());734 735  // Accept x > -1 or x >= 0 (after potentially inverting the predicate).736  if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||737        (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))738    return nullptr;739 740  ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :741                               Cmp1->getPredicate());742 743  Value *Input = Cmp0->getOperand(0);744  Value *Cmp1Op0 = Cmp1->getOperand(0);745  Value *Cmp1Op1 = Cmp1->getOperand(1);746  Value *RangeEnd;747  if (match(Cmp1Op0, m_SExtOrSelf(m_Specific(Input)))) {748    // For the upper range compare we have: icmp x, n749    Input = Cmp1Op0;750    RangeEnd = Cmp1Op1;751  } else if (match(Cmp1Op1, m_SExtOrSelf(m_Specific(Input)))) {752    // For the upper range compare we have: icmp n, x753    Input = Cmp1Op1;754    RangeEnd = Cmp1Op0;755    Pred1 = ICmpInst::getSwappedPredicate(Pred1);756  } else {757    return nullptr;758  }759 760  // Check the upper range comparison, e.g. x < n761  ICmpInst::Predicate NewPred;762  switch (Pred1) {763    case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;764    case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;765    default: return nullptr;766  }767 768  // This simplification is only valid if the upper range is not negative.769  KnownBits Known = computeKnownBits(RangeEnd, Cmp1);770  if (!Known.isNonNegative())771    return nullptr;772 773  if (Inverted)774    NewPred = ICmpInst::getInversePredicate(NewPred);775 776  return Builder.CreateICmp(NewPred, Input, RangeEnd);777}778 779// (or (icmp eq X, 0), (icmp eq X, Pow2OrZero))780//      -> (icmp eq (and X, Pow2OrZero), X)781// (and (icmp ne X, 0), (icmp ne X, Pow2OrZero))782//      -> (icmp ne (and X, Pow2OrZero), X)783static Value *784foldAndOrOfICmpsWithPow2AndWithZero(InstCombiner::BuilderTy &Builder,785                                    ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,786                                    const SimplifyQuery &Q) {787  CmpPredicate Pred = IsAnd ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;788  // Make sure we have right compares for our op.789  if (LHS->getPredicate() != Pred || RHS->getPredicate() != Pred)790    return nullptr;791 792  // Make it so we can match LHS against the (icmp eq/ne X, 0) just for793  // simplicity.794  if (match(RHS->getOperand(1), m_Zero()))795    std::swap(LHS, RHS);796 797  Value *Pow2, *Op;798  // Match the desired pattern:799  // LHS: (icmp eq/ne X, 0)800  // RHS: (icmp eq/ne X, Pow2OrZero)801  // Skip if Pow2OrZero is 1. Either way it gets folded to (icmp ugt X, 1) but802  // this form ends up slightly less canonical.803  // We could potentially be more sophisticated than requiring LHS/RHS804  // be one-use. We don't create additional instructions if only one805  // of them is one-use. So cases where one is one-use and the other806  // is two-use might be profitable.807  if (!match(LHS, m_OneUse(m_ICmp(Pred, m_Value(Op), m_Zero()))) ||808      !match(RHS, m_OneUse(m_c_ICmp(Pred, m_Specific(Op), m_Value(Pow2)))) ||809      match(Pow2, m_One()) ||810      !isKnownToBeAPowerOfTwo(Pow2, Q.DL, /*OrZero=*/true, Q.AC, Q.CxtI, Q.DT))811    return nullptr;812 813  Value *And = Builder.CreateAnd(Op, Pow2);814  return Builder.CreateICmp(Pred, And, Op);815}816 817/// General pattern:818///   X & Y819///820/// Where Y is checking that all the high bits (covered by a mask 4294967168)821/// are uniform, i.e.  %arg & 4294967168  can be either  4294967168  or  0822/// Pattern can be one of:823///   %t = add        i32 %arg,    128824///   %r = icmp   ult i32 %t,      256825/// Or826///   %t0 = shl       i32 %arg,    24827///   %t1 = ashr      i32 %t0,     24828///   %r  = icmp  eq  i32 %t1,     %arg829/// Or830///   %t0 = trunc     i32 %arg  to i8831///   %t1 = sext      i8  %t0   to i32832///   %r  = icmp  eq  i32 %t1,     %arg833/// This pattern is a signed truncation check.834///835/// And X is checking that some bit in that same mask is zero.836/// I.e. can be one of:837///   %r = icmp sgt i32   %arg,    -1838/// Or839///   %t = and      i32   %arg,    2147483648840///   %r = icmp eq  i32   %t,      0841///842/// Since we are checking that all the bits in that mask are the same,843/// and a particular bit is zero, what we are really checking is that all the844/// masked bits are zero.845/// So this should be transformed to:846///   %r = icmp ult i32 %arg, 128847static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1,848                                        Instruction &CxtI,849                                        InstCombiner::BuilderTy &Builder) {850  assert(CxtI.getOpcode() == Instruction::And);851 852  // Match  icmp ult (add %arg, C01), C1   (C1 == C01 << 1; powers of two)853  auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X,854                                            APInt &SignBitMask) -> bool {855    const APInt *I01, *I1; // powers of two; I1 == I01 << 1856    if (!(match(ICmp, m_SpecificICmp(ICmpInst::ICMP_ULT,857                                     m_Add(m_Value(X), m_Power2(I01)),858                                     m_Power2(I1))) &&859          I1->ugt(*I01) && I01->shl(1) == *I1))860      return false;861    // Which bit is the new sign bit as per the 'signed truncation' pattern?862    SignBitMask = *I01;863    return true;864  };865 866  // One icmp needs to be 'signed truncation check'.867  // We need to match this first, else we will mismatch commutative cases.868  Value *X1;869  APInt HighestBit;870  ICmpInst *OtherICmp;871  if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit))872    OtherICmp = ICmp0;873  else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit))874    OtherICmp = ICmp1;875  else876    return nullptr;877 878  assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)");879 880  // Try to match/decompose into:  icmp eq (X & Mask), 0881  auto tryToDecompose = [](ICmpInst *ICmp, Value *&X,882                           APInt &UnsetBitsMask) -> bool {883    CmpPredicate Pred = ICmp->getPredicate();884    // Can it be decomposed into  icmp eq (X & Mask), 0  ?885    auto Res = llvm::decomposeBitTestICmp(886        ICmp->getOperand(0), ICmp->getOperand(1), Pred,887        /*LookThroughTrunc=*/false, /*AllowNonZeroC=*/false,888        /*DecomposeAnd=*/true);889    if (Res && Res->Pred == ICmpInst::ICMP_EQ) {890      X = Res->X;891      UnsetBitsMask = Res->Mask;892      return true;893    }894 895    return false;896  };897 898  // And the other icmp needs to be decomposable into a bit test.899  Value *X0;900  APInt UnsetBitsMask;901  if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask))902    return nullptr;903 904  assert(!UnsetBitsMask.isZero() && "empty mask makes no sense.");905 906  // Are they working on the same value?907  Value *X;908  if (X1 == X0) {909    // Ok as is.910    X = X1;911  } else if (match(X0, m_Trunc(m_Specific(X1)))) {912    UnsetBitsMask = UnsetBitsMask.zext(X1->getType()->getScalarSizeInBits());913    X = X1;914  } else915    return nullptr;916 917  // So which bits should be uniform as per the 'signed truncation check'?918  // (all the bits starting with (i.e. including) HighestBit)919  APInt SignBitsMask = ~(HighestBit - 1U);920 921  // UnsetBitsMask must have some common bits with SignBitsMask,922  if (!UnsetBitsMask.intersects(SignBitsMask))923    return nullptr;924 925  // Does UnsetBitsMask contain any bits outside of SignBitsMask?926  if (!UnsetBitsMask.isSubsetOf(SignBitsMask)) {927    APInt OtherHighestBit = (~UnsetBitsMask) + 1U;928    if (!OtherHighestBit.isPowerOf2())929      return nullptr;930    HighestBit = APIntOps::umin(HighestBit, OtherHighestBit);931  }932  // Else, if it does not, then all is ok as-is.933 934  // %r = icmp ult %X, SignBit935  return Builder.CreateICmpULT(X, ConstantInt::get(X->getType(), HighestBit),936                               CxtI.getName() + ".simplified");937}938 939/// Fold (icmp eq ctpop(X) 1) | (icmp eq X 0) into (icmp ult ctpop(X) 2) and940/// fold (icmp ne ctpop(X) 1) & (icmp ne X 0) into (icmp ugt ctpop(X) 1).941/// Also used for logical and/or, must be poison safe if range attributes are942/// dropped.943static Value *foldIsPowerOf2OrZero(ICmpInst *Cmp0, ICmpInst *Cmp1, bool IsAnd,944                                   InstCombiner::BuilderTy &Builder,945                                   InstCombinerImpl &IC) {946  CmpPredicate Pred0, Pred1;947  Value *X;948  if (!match(Cmp0, m_ICmp(Pred0, m_Intrinsic<Intrinsic::ctpop>(m_Value(X)),949                          m_SpecificInt(1))) ||950      !match(Cmp1, m_ICmp(Pred1, m_Specific(X), m_ZeroInt())))951    return nullptr;952 953  auto *CtPop = cast<Instruction>(Cmp0->getOperand(0));954  if (IsAnd && Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_NE) {955    // Drop range attributes and re-infer them in the next iteration.956    CtPop->dropPoisonGeneratingAnnotations();957    IC.addToWorklist(CtPop);958    return Builder.CreateICmpUGT(CtPop, ConstantInt::get(CtPop->getType(), 1));959  }960  if (!IsAnd && Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_EQ) {961    // Drop range attributes and re-infer them in the next iteration.962    CtPop->dropPoisonGeneratingAnnotations();963    IC.addToWorklist(CtPop);964    return Builder.CreateICmpULT(CtPop, ConstantInt::get(CtPop->getType(), 2));965  }966 967  return nullptr;968}969 970/// Reduce a pair of compares that check if a value has exactly 1 bit set.971/// Also used for logical and/or, must be poison safe if range attributes are972/// dropped.973static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd,974                             InstCombiner::BuilderTy &Builder,975                             InstCombinerImpl &IC) {976  // Handle 'and' / 'or' commutation: make the equality check the first operand.977  if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE)978    std::swap(Cmp0, Cmp1);979  else if (!JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_EQ)980    std::swap(Cmp0, Cmp1);981 982  // (X != 0) && (ctpop(X) u< 2) --> ctpop(X) == 1983  Value *X;984  if (JoinedByAnd &&985      match(Cmp0, m_SpecificICmp(ICmpInst::ICMP_NE, m_Value(X), m_ZeroInt())) &&986      match(Cmp1, m_SpecificICmp(ICmpInst::ICMP_ULT,987                                 m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),988                                 m_SpecificInt(2)))) {989    auto *CtPop = cast<Instruction>(Cmp1->getOperand(0));990    // Drop range attributes and re-infer them in the next iteration.991    CtPop->dropPoisonGeneratingAnnotations();992    IC.addToWorklist(CtPop);993    return Builder.CreateICmpEQ(CtPop, ConstantInt::get(CtPop->getType(), 1));994  }995  // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1996  if (!JoinedByAnd &&997      match(Cmp0, m_SpecificICmp(ICmpInst::ICMP_EQ, m_Value(X), m_ZeroInt())) &&998      match(Cmp1, m_SpecificICmp(ICmpInst::ICMP_UGT,999                                 m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),1000                                 m_SpecificInt(1)))) {1001    auto *CtPop = cast<Instruction>(Cmp1->getOperand(0));1002    // Drop range attributes and re-infer them in the next iteration.1003    CtPop->dropPoisonGeneratingAnnotations();1004    IC.addToWorklist(CtPop);1005    return Builder.CreateICmpNE(CtPop, ConstantInt::get(CtPop->getType(), 1));1006  }1007  return nullptr;1008}1009 1010/// Try to fold (icmp(A & B) == 0) & (icmp(A & D) != E) into (icmp A u< D) iff1011/// B is a contiguous set of ones starting from the most significant bit1012/// (negative power of 2), D and E are equal, and D is a contiguous set of ones1013/// starting at the most significant zero bit in B. Parameter B supports masking1014/// using undef/poison in either scalar or vector values.1015static Value *foldNegativePower2AndShiftedMask(1016    Value *A, Value *B, Value *D, Value *E, ICmpInst::Predicate PredL,1017    ICmpInst::Predicate PredR, InstCombiner::BuilderTy &Builder) {1018  assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&1019         "Expected equality predicates for masked type of icmps.");1020  if (PredL != ICmpInst::ICMP_EQ || PredR != ICmpInst::ICMP_NE)1021    return nullptr;1022 1023  if (!match(B, m_NegatedPower2()) || !match(D, m_ShiftedMask()) ||1024      !match(E, m_ShiftedMask()))1025    return nullptr;1026 1027  // Test scalar arguments for conversion. B has been validated earlier to be a1028  // negative power of two and thus is guaranteed to have one or more contiguous1029  // ones starting from the MSB followed by zero or more contiguous zeros. D has1030  // been validated earlier to be a shifted set of one or more contiguous ones.1031  // In order to match, B leading ones and D leading zeros should be equal. The1032  // predicate that B be a negative power of 2 prevents the condition of there1033  // ever being zero leading ones. Thus 0 == 0 cannot occur. The predicate that1034  // D always be a shifted mask prevents the condition of D equaling 0. This1035  // prevents matching the condition where B contains the maximum number of1036  // leading one bits (-1) and D contains the maximum number of leading zero1037  // bits (0).1038  auto isReducible = [](const Value *B, const Value *D, const Value *E) {1039    const APInt *BCst, *DCst, *ECst;1040    return match(B, m_APIntAllowPoison(BCst)) && match(D, m_APInt(DCst)) &&1041           match(E, m_APInt(ECst)) && *DCst == *ECst &&1042           (isa<PoisonValue>(B) ||1043            (BCst->countLeadingOnes() == DCst->countLeadingZeros()));1044  };1045 1046  // Test vector type arguments for conversion.1047  if (const auto *BVTy = dyn_cast<VectorType>(B->getType())) {1048    const auto *BFVTy = dyn_cast<FixedVectorType>(BVTy);1049    const auto *BConst = dyn_cast<Constant>(B);1050    const auto *DConst = dyn_cast<Constant>(D);1051    const auto *EConst = dyn_cast<Constant>(E);1052 1053    if (!BFVTy || !BConst || !DConst || !EConst)1054      return nullptr;1055 1056    for (unsigned I = 0; I != BFVTy->getNumElements(); ++I) {1057      const auto *BElt = BConst->getAggregateElement(I);1058      const auto *DElt = DConst->getAggregateElement(I);1059      const auto *EElt = EConst->getAggregateElement(I);1060 1061      if (!BElt || !DElt || !EElt)1062        return nullptr;1063      if (!isReducible(BElt, DElt, EElt))1064        return nullptr;1065    }1066  } else {1067    // Test scalar type arguments for conversion.1068    if (!isReducible(B, D, E))1069      return nullptr;1070  }1071  return Builder.CreateICmp(ICmpInst::ICMP_ULT, A, D);1072}1073 1074/// Try to fold ((icmp X u< P) & (icmp(X & M) != M)) or ((icmp X s> -1) &1075/// (icmp(X & M) != M)) into (icmp X u< M). Where P is a power of 2, M < P, and1076/// M is a contiguous shifted mask starting at the right most significant zero1077/// bit in P. SGT is supported as when P is the largest representable power of1078/// 2, an earlier optimization converts the expression into (icmp X s> -1).1079/// Parameter P supports masking using undef/poison in either scalar or vector1080/// values.1081static Value *foldPowerOf2AndShiftedMask(ICmpInst *Cmp0, ICmpInst *Cmp1,1082                                         bool JoinedByAnd,1083                                         InstCombiner::BuilderTy &Builder) {1084  if (!JoinedByAnd)1085    return nullptr;1086  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;1087  ICmpInst::Predicate CmpPred0, CmpPred1;1088  // Assuming P is a 2^n, getMaskedTypeForICmpPair will normalize (icmp X u<1089  // 2^n) into (icmp (X & ~(2^n-1)) == 0) and (icmp X s> -1) into (icmp (X &1090  // SignMask) == 0).1091  std::optional<std::pair<unsigned, unsigned>> MaskPair =1092      getMaskedTypeForICmpPair(A, B, C, D, E, Cmp0, Cmp1, CmpPred0, CmpPred1);1093  if (!MaskPair)1094    return nullptr;1095 1096  const auto compareBMask = BMask_NotMixed | BMask_NotAllOnes;1097  unsigned CmpMask0 = MaskPair->first;1098  unsigned CmpMask1 = MaskPair->second;1099  if ((CmpMask0 & Mask_AllZeros) && (CmpMask1 == compareBMask)) {1100    if (Value *V = foldNegativePower2AndShiftedMask(A, B, D, E, CmpPred0,1101                                                    CmpPred1, Builder))1102      return V;1103  } else if ((CmpMask0 == compareBMask) && (CmpMask1 & Mask_AllZeros)) {1104    if (Value *V = foldNegativePower2AndShiftedMask(A, D, B, C, CmpPred1,1105                                                    CmpPred0, Builder))1106      return V;1107  }1108  return nullptr;1109}1110 1111/// Commuted variants are assumed to be handled by calling this function again1112/// with the parameters swapped.1113static Value *foldUnsignedUnderflowCheck(ICmpInst *ZeroICmp,1114                                         ICmpInst *UnsignedICmp, bool IsAnd,1115                                         const SimplifyQuery &Q,1116                                         InstCombiner::BuilderTy &Builder) {1117  Value *ZeroCmpOp;1118  CmpPredicate EqPred;1119  if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(ZeroCmpOp), m_Zero())) ||1120      !ICmpInst::isEquality(EqPred))1121    return nullptr;1122 1123  CmpPredicate UnsignedPred;1124 1125  Value *A, *B;1126  if (match(UnsignedICmp,1127            m_c_ICmp(UnsignedPred, m_Specific(ZeroCmpOp), m_Value(A))) &&1128      match(ZeroCmpOp, m_c_Add(m_Specific(A), m_Value(B))) &&1129      (ZeroICmp->hasOneUse() || UnsignedICmp->hasOneUse())) {1130    auto GetKnownNonZeroAndOther = [&](Value *&NonZero, Value *&Other) {1131      if (!isKnownNonZero(NonZero, Q))1132        std::swap(NonZero, Other);1133      return isKnownNonZero(NonZero, Q);1134    };1135 1136    // Given  ZeroCmpOp = (A + B)1137    //   ZeroCmpOp <  A && ZeroCmpOp != 0  -->  (0-X) <  Y  iff1138    //   ZeroCmpOp >= A || ZeroCmpOp == 0  -->  (0-X) >= Y  iff1139    //     with X being the value (A/B) that is known to be non-zero,1140    //     and Y being remaining value.1141    if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE &&1142        IsAnd && GetKnownNonZeroAndOther(B, A))1143      return Builder.CreateICmpULT(Builder.CreateNeg(B), A);1144    if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ &&1145        !IsAnd && GetKnownNonZeroAndOther(B, A))1146      return Builder.CreateICmpUGE(Builder.CreateNeg(B), A);1147  }1148 1149  return nullptr;1150}1151 1152struct IntPart {1153  Value *From;1154  unsigned StartBit;1155  unsigned NumBits;1156};1157 1158/// Match an extraction of bits from an integer.1159static std::optional<IntPart> matchIntPart(Value *V) {1160  Value *X;1161  if (!match(V, m_OneUse(m_Trunc(m_Value(X)))))1162    return std::nullopt;1163 1164  unsigned NumOriginalBits = X->getType()->getScalarSizeInBits();1165  unsigned NumExtractedBits = V->getType()->getScalarSizeInBits();1166  Value *Y;1167  const APInt *Shift;1168  // For a trunc(lshr Y, Shift) pattern, make sure we're only extracting bits1169  // from Y, not any shifted-in zeroes.1170  if (match(X, m_OneUse(m_LShr(m_Value(Y), m_APInt(Shift)))) &&1171      Shift->ule(NumOriginalBits - NumExtractedBits))1172    return {{Y, (unsigned)Shift->getZExtValue(), NumExtractedBits}};1173  return {{X, 0, NumExtractedBits}};1174}1175 1176/// Materialize an extraction of bits from an integer in IR.1177static Value *extractIntPart(const IntPart &P, IRBuilderBase &Builder) {1178  Value *V = P.From;1179  if (P.StartBit)1180    V = Builder.CreateLShr(V, P.StartBit);1181  Type *TruncTy = V->getType()->getWithNewBitWidth(P.NumBits);1182  if (TruncTy != V->getType())1183    V = Builder.CreateTrunc(V, TruncTy);1184  return V;1185}1186 1187/// (icmp eq X0, Y0) & (icmp eq X1, Y1) -> icmp eq X01, Y011188/// (icmp ne X0, Y0) | (icmp ne X1, Y1) -> icmp ne X01, Y011189/// where X0, X1 and Y0, Y1 are adjacent parts extracted from an integer.1190Value *InstCombinerImpl::foldEqOfParts(Value *Cmp0, Value *Cmp1, bool IsAnd) {1191  if (!Cmp0->hasOneUse() || !Cmp1->hasOneUse())1192    return nullptr;1193 1194  CmpInst::Predicate Pred = IsAnd ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;1195  auto GetMatchPart = [&](Value *CmpV,1196                          unsigned OpNo) -> std::optional<IntPart> {1197    assert(CmpV->getType()->isIntOrIntVectorTy(1) && "Must be bool");1198 1199    Value *X, *Y;1200    // icmp ne (and x, 1), (and y, 1) <=> trunc (xor x, y) to i11201    // icmp eq (and x, 1), (and y, 1) <=> not (trunc (xor x, y) to i1)1202    if (Pred == CmpInst::ICMP_NE1203            ? match(CmpV, m_Trunc(m_Xor(m_Value(X), m_Value(Y))))1204            : match(CmpV, m_Not(m_Trunc(m_Xor(m_Value(X), m_Value(Y))))))1205      return {{OpNo == 0 ? X : Y, 0, 1}};1206 1207    auto *Cmp = dyn_cast<ICmpInst>(CmpV);1208    if (!Cmp)1209      return std::nullopt;1210 1211    if (Pred == Cmp->getPredicate())1212      return matchIntPart(Cmp->getOperand(OpNo));1213 1214    const APInt *C;1215    // (icmp eq (lshr x, C), (lshr y, C)) gets optimized to:1216    // (icmp ult (xor x, y), 1 << C) so also look for that.1217    if (Pred == CmpInst::ICMP_EQ && Cmp->getPredicate() == CmpInst::ICMP_ULT) {1218      if (!match(Cmp->getOperand(1), m_Power2(C)) ||1219          !match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value())))1220        return std::nullopt;1221    }1222 1223    // (icmp ne (lshr x, C), (lshr y, C)) gets optimized to:1224    // (icmp ugt (xor x, y), (1 << C) - 1) so also look for that.1225    else if (Pred == CmpInst::ICMP_NE &&1226             Cmp->getPredicate() == CmpInst::ICMP_UGT) {1227      if (!match(Cmp->getOperand(1), m_LowBitMask(C)) ||1228          !match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value())))1229        return std::nullopt;1230    } else {1231      return std::nullopt;1232    }1233 1234    unsigned From = Pred == CmpInst::ICMP_NE ? C->popcount() : C->countr_zero();1235    Instruction *I = cast<Instruction>(Cmp->getOperand(0));1236    return {{I->getOperand(OpNo), From, C->getBitWidth() - From}};1237  };1238 1239  std::optional<IntPart> L0 = GetMatchPart(Cmp0, 0);1240  std::optional<IntPart> R0 = GetMatchPart(Cmp0, 1);1241  std::optional<IntPart> L1 = GetMatchPart(Cmp1, 0);1242  std::optional<IntPart> R1 = GetMatchPart(Cmp1, 1);1243  if (!L0 || !R0 || !L1 || !R1)1244    return nullptr;1245 1246  // Make sure the LHS/RHS compare a part of the same value, possibly after1247  // an operand swap.1248  if (L0->From != L1->From || R0->From != R1->From) {1249    if (L0->From != R1->From || R0->From != L1->From)1250      return nullptr;1251    std::swap(L1, R1);1252  }1253 1254  // Make sure the extracted parts are adjacent, canonicalizing to L0/R0 being1255  // the low part and L1/R1 being the high part.1256  if (L0->StartBit + L0->NumBits != L1->StartBit ||1257      R0->StartBit + R0->NumBits != R1->StartBit) {1258    if (L1->StartBit + L1->NumBits != L0->StartBit ||1259        R1->StartBit + R1->NumBits != R0->StartBit)1260      return nullptr;1261    std::swap(L0, L1);1262    std::swap(R0, R1);1263  }1264 1265  // We can simplify to a comparison of these larger parts of the integers.1266  IntPart L = {L0->From, L0->StartBit, L0->NumBits + L1->NumBits};1267  IntPart R = {R0->From, R0->StartBit, R0->NumBits + R1->NumBits};1268  Value *LValue = extractIntPart(L, Builder);1269  Value *RValue = extractIntPart(R, Builder);1270  return Builder.CreateICmp(Pred, LValue, RValue);1271}1272 1273/// Reduce logic-of-compares with equality to a constant by substituting a1274/// common operand with the constant. Callers are expected to call this with1275/// Cmp0/Cmp1 switched to handle logic op commutativity.1276static Value *foldAndOrOfICmpsWithConstEq(ICmpInst *Cmp0, ICmpInst *Cmp1,1277                                          bool IsAnd, bool IsLogical,1278                                          InstCombiner::BuilderTy &Builder,1279                                          const SimplifyQuery &Q,1280                                          Instruction &I) {1281  // Match an equality compare with a non-poison constant as Cmp0.1282  // Also, give up if the compare can be constant-folded to avoid looping.1283  CmpPredicate Pred0;1284  Value *X;1285  Constant *C;1286  if (!match(Cmp0, m_ICmp(Pred0, m_Value(X), m_Constant(C))) ||1287      !isGuaranteedNotToBeUndefOrPoison(C) || isa<Constant>(X))1288    return nullptr;1289  if ((IsAnd && Pred0 != ICmpInst::ICMP_EQ) ||1290      (!IsAnd && Pred0 != ICmpInst::ICMP_NE))1291    return nullptr;1292 1293  // The other compare must include a common operand (X). Canonicalize the1294  // common operand as operand 1 (Pred1 is swapped if the common operand was1295  // operand 0).1296  Value *Y;1297  CmpPredicate Pred1;1298  if (!match(Cmp1, m_c_ICmp(Pred1, m_Value(Y), m_Specific(X))))1299    return nullptr;1300 1301  // Replace variable with constant value equivalence to remove a variable use:1302  // (X == C) && (Y Pred1 X) --> (X == C) && (Y Pred1 C)1303  // (X != C) || (Y Pred1 X) --> (X != C) || (Y Pred1 C)1304  // Can think of the 'or' substitution with the 'and' bool equivalent:1305  // A || B --> A || (!A && B)1306  Value *SubstituteCmp = simplifyICmpInst(Pred1, Y, C, Q);1307  if (!SubstituteCmp) {1308    // If we need to create a new instruction, require that the old compare can1309    // be removed.1310    if (!Cmp1->hasOneUse())1311      return nullptr;1312    SubstituteCmp = Builder.CreateICmp(Pred1, Y, C);1313  }1314  if (IsLogical) {1315    Instruction *MDFrom =1316        ProfcheckDisableMetadataFixes && isa<SelectInst>(I) ? nullptr : &I;1317    return IsAnd ? Builder.CreateLogicalAnd(Cmp0, SubstituteCmp, "", MDFrom)1318                 : Builder.CreateLogicalOr(Cmp0, SubstituteCmp, "", MDFrom);1319  }1320  return Builder.CreateBinOp(IsAnd ? Instruction::And : Instruction::Or, Cmp0,1321                             SubstituteCmp);1322}1323 1324/// Fold (icmp Pred1 V1, C1) & (icmp Pred2 V2, C2)1325/// or   (icmp Pred1 V1, C1) | (icmp Pred2 V2, C2)1326/// into a single comparison using range-based reasoning.1327/// NOTE: This is also used for logical and/or, must be poison-safe!1328Value *InstCombinerImpl::foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1,1329                                                     ICmpInst *ICmp2,1330                                                     bool IsAnd) {1331  // Return (V, CR) for a range check idiom V in CR.1332  auto MatchExactRangeCheck =1333      [](ICmpInst *ICmp) -> std::optional<std::pair<Value *, ConstantRange>> {1334    const APInt *C;1335    if (!match(ICmp->getOperand(1), m_APInt(C)))1336      return std::nullopt;1337    Value *LHS = ICmp->getOperand(0);1338    CmpPredicate Pred = ICmp->getPredicate();1339    Value *X;1340    // Match (x & NegPow2) ==/!= C1341    const APInt *Mask;1342    if (ICmpInst::isEquality(Pred) &&1343        match(LHS, m_OneUse(m_And(m_Value(X), m_NegatedPower2(Mask)))) &&1344        C->countr_zero() >= Mask->countr_zero()) {1345      ConstantRange CR(*C, *C - *Mask);1346      if (Pred == ICmpInst::ICMP_NE)1347        CR = CR.inverse();1348      return std::make_pair(X, CR);1349    }1350    ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred, *C);1351    // Match (add X, C1) pred C1352    // TODO: investigate whether we should apply the one-use check on m_AddLike.1353    const APInt *C1;1354    if (match(LHS, m_AddLike(m_Value(X), m_APInt(C1))))1355      return std::make_pair(X, CR.subtract(*C1));1356    return std::make_pair(LHS, CR);1357  };1358 1359  auto RC1 = MatchExactRangeCheck(ICmp1);1360  if (!RC1)1361    return nullptr;1362 1363  auto RC2 = MatchExactRangeCheck(ICmp2);1364  if (!RC2)1365    return nullptr;1366 1367  auto &[V1, CR1] = *RC1;1368  auto &[V2, CR2] = *RC2;1369  if (V1 != V2)1370    return nullptr;1371 1372  // For 'and', we use the De Morgan's Laws to simplify the implementation.1373  if (IsAnd) {1374    CR1 = CR1.inverse();1375    CR2 = CR2.inverse();1376  }1377 1378  Type *Ty = V1->getType();1379  Value *NewV = V1;1380  std::optional<ConstantRange> CR = CR1.exactUnionWith(CR2);1381  if (!CR) {1382    if (!(ICmp1->hasOneUse() && ICmp2->hasOneUse()) || CR1.isWrappedSet() ||1383        CR2.isWrappedSet())1384      return nullptr;1385 1386    // Check whether we have equal-size ranges that only differ by one bit.1387    // In that case we can apply a mask to map one range onto the other.1388    APInt LowerDiff = CR1.getLower() ^ CR2.getLower();1389    APInt UpperDiff = (CR1.getUpper() - 1) ^ (CR2.getUpper() - 1);1390    APInt CR1Size = CR1.getUpper() - CR1.getLower();1391    if (!LowerDiff.isPowerOf2() || LowerDiff != UpperDiff ||1392        CR1Size != CR2.getUpper() - CR2.getLower())1393      return nullptr;1394 1395    CR = CR1.getLower().ult(CR2.getLower()) ? CR1 : CR2;1396    NewV = Builder.CreateAnd(NewV, ConstantInt::get(Ty, ~LowerDiff));1397  }1398 1399  if (IsAnd)1400    CR = CR->inverse();1401 1402  CmpInst::Predicate NewPred;1403  APInt NewC, Offset;1404  CR->getEquivalentICmp(NewPred, NewC, Offset);1405 1406  if (Offset != 0)1407    NewV = Builder.CreateAdd(NewV, ConstantInt::get(Ty, Offset));1408  return Builder.CreateICmp(NewPred, NewV, ConstantInt::get(Ty, NewC));1409}1410 1411/// Ignore all operations which only change the sign of a value, returning the1412/// underlying magnitude value.1413static Value *stripSignOnlyFPOps(Value *Val) {1414  match(Val, m_FNeg(m_Value(Val)));1415  match(Val, m_FAbs(m_Value(Val)));1416  match(Val, m_CopySign(m_Value(Val), m_Value()));1417  return Val;1418}1419 1420/// Matches canonical form of isnan, fcmp ord x, 01421static bool matchIsNotNaN(FCmpInst::Predicate P, Value *LHS, Value *RHS) {1422  return P == FCmpInst::FCMP_ORD && match(RHS, m_AnyZeroFP());1423}1424 1425/// Matches fcmp u__ x, +/-inf1426static bool matchUnorderedInfCompare(FCmpInst::Predicate P, Value *LHS,1427                                     Value *RHS) {1428  return FCmpInst::isUnordered(P) && match(RHS, m_Inf());1429}1430 1431/// and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf1432///1433/// Clang emits this pattern for doing an isfinite check in __builtin_isnormal.1434static Value *matchIsFiniteTest(InstCombiner::BuilderTy &Builder, FCmpInst *LHS,1435                                FCmpInst *RHS) {1436  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);1437  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);1438  FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();1439 1440  if (!matchIsNotNaN(PredL, LHS0, LHS1) ||1441      !matchUnorderedInfCompare(PredR, RHS0, RHS1))1442    return nullptr;1443 1444  return Builder.CreateFCmpFMF(FCmpInst::getOrderedPredicate(PredR), RHS0, RHS1,1445                               FMFSource::intersect(LHS, RHS));1446}1447 1448Value *InstCombinerImpl::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS,1449                                          bool IsAnd, bool IsLogicalSelect) {1450  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);1451  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);1452  FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();1453 1454  if (LHS0 == RHS1 && RHS0 == LHS1) {1455    // Swap RHS operands to match LHS.1456    PredR = FCmpInst::getSwappedPredicate(PredR);1457    std::swap(RHS0, RHS1);1458  }1459 1460  // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).1461  // Suppose the relation between x and y is R, where R is one of1462  // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for1463  // testing the desired relations.1464  //1465  // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:1466  //    bool(R & CC0) && bool(R & CC1)1467  //  = bool((R & CC0) & (R & CC1))1468  //  = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency1469  //1470  // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:1471  //    bool(R & CC0) || bool(R & CC1)1472  //  = bool((R & CC0) | (R & CC1))1473  //  = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)1474  if (LHS0 == RHS0 && LHS1 == RHS1) {1475    unsigned FCmpCodeL = getFCmpCode(PredL);1476    unsigned FCmpCodeR = getFCmpCode(PredR);1477    unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR;1478 1479    // Intersect the fast math flags.1480    // TODO: We can union the fast math flags unless this is a logical select.1481    return getFCmpValue(NewPred, LHS0, LHS1, Builder,1482                        FMFSource::intersect(LHS, RHS));1483  }1484 1485  // This transform is not valid for a logical select.1486  if (!IsLogicalSelect &&1487      ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||1488       (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO &&1489        !IsAnd))) {1490    if (LHS0->getType() != RHS0->getType())1491      return nullptr;1492 1493    // FCmp canonicalization ensures that (fcmp ord/uno X, X) and1494    // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0).1495    if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP())) {1496      // Ignore the constants because they are obviously not NANs:1497      // (fcmp ord x, 0.0) & (fcmp ord y, 0.0)  -> (fcmp ord x, y)1498      // (fcmp uno x, 0.0) | (fcmp uno y, 0.0)  -> (fcmp uno x, y)1499      return Builder.CreateFCmpFMF(PredL, LHS0, RHS0,1500                                   FMFSource::intersect(LHS, RHS));1501    }1502  }1503 1504  // This transform is not valid for a logical select.1505  if (!IsLogicalSelect && IsAnd &&1506      stripSignOnlyFPOps(LHS0) == stripSignOnlyFPOps(RHS0)) {1507    // and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf1508    // and (fcmp ord x, 0), (fcmp u* fabs(x), inf) -> fcmp o* x, inf1509    if (Value *Left = matchIsFiniteTest(Builder, LHS, RHS))1510      return Left;1511    if (Value *Right = matchIsFiniteTest(Builder, RHS, LHS))1512      return Right;1513  }1514 1515  // Turn at least two fcmps with constants into llvm.is.fpclass.1516  //1517  // If we can represent a combined value test with one class call, we can1518  // potentially eliminate 4-6 instructions. If we can represent a test with a1519  // single fcmp with fneg and fabs, that's likely a better canonical form.1520  if (LHS->hasOneUse() && RHS->hasOneUse()) {1521    auto [ClassValRHS, ClassMaskRHS] =1522        fcmpToClassTest(PredR, *RHS->getFunction(), RHS0, RHS1);1523    if (ClassValRHS) {1524      auto [ClassValLHS, ClassMaskLHS] =1525          fcmpToClassTest(PredL, *LHS->getFunction(), LHS0, LHS1);1526      if (ClassValLHS == ClassValRHS) {1527        unsigned CombinedMask = IsAnd ? (ClassMaskLHS & ClassMaskRHS)1528                                      : (ClassMaskLHS | ClassMaskRHS);1529        return Builder.CreateIntrinsic(1530            Intrinsic::is_fpclass, {ClassValLHS->getType()},1531            {ClassValLHS, Builder.getInt32(CombinedMask)});1532      }1533    }1534  }1535 1536  // Canonicalize the range check idiom:1537  // and (fcmp olt/ole/ult/ule x, C), (fcmp ogt/oge/ugt/uge x, -C)1538  // --> fabs(x) olt/ole/ult/ule C1539  // or  (fcmp ogt/oge/ugt/uge x, C), (fcmp olt/ole/ult/ule x, -C)1540  // --> fabs(x) ogt/oge/ugt/uge C1541  // TODO: Generalize to handle a negated variable operand?1542  const APFloat *LHSC, *RHSC;1543  if (LHS0 == RHS0 && LHS->hasOneUse() && RHS->hasOneUse() &&1544      FCmpInst::getSwappedPredicate(PredL) == PredR &&1545      match(LHS1, m_APFloatAllowPoison(LHSC)) &&1546      match(RHS1, m_APFloatAllowPoison(RHSC)) &&1547      LHSC->bitwiseIsEqual(neg(*RHSC))) {1548    auto IsLessThanOrLessEqual = [](FCmpInst::Predicate Pred) {1549      switch (Pred) {1550      case FCmpInst::FCMP_OLT:1551      case FCmpInst::FCMP_OLE:1552      case FCmpInst::FCMP_ULT:1553      case FCmpInst::FCMP_ULE:1554        return true;1555      default:1556        return false;1557      }1558    };1559    if (IsLessThanOrLessEqual(IsAnd ? PredR : PredL)) {1560      std::swap(LHSC, RHSC);1561      std::swap(PredL, PredR);1562    }1563    if (IsLessThanOrLessEqual(IsAnd ? PredL : PredR)) {1564      FastMathFlags NewFlag = LHS->getFastMathFlags();1565      if (!IsLogicalSelect)1566        NewFlag |= RHS->getFastMathFlags();1567 1568      Value *FAbs =1569          Builder.CreateUnaryIntrinsic(Intrinsic::fabs, LHS0, NewFlag);1570      return Builder.CreateFCmpFMF(1571          PredL, FAbs, ConstantFP::get(LHS0->getType(), *LHSC), NewFlag);1572    }1573  }1574 1575  return nullptr;1576}1577 1578/// Match an fcmp against a special value that performs a test possible by1579/// llvm.is.fpclass.1580static bool matchIsFPClassLikeFCmp(Value *Op, Value *&ClassVal,1581                                   uint64_t &ClassMask) {1582  auto *FCmp = dyn_cast<FCmpInst>(Op);1583  if (!FCmp || !FCmp->hasOneUse())1584    return false;1585 1586  std::tie(ClassVal, ClassMask) =1587      fcmpToClassTest(FCmp->getPredicate(), *FCmp->getParent()->getParent(),1588                      FCmp->getOperand(0), FCmp->getOperand(1));1589  return ClassVal != nullptr;1590}1591 1592/// or (is_fpclass x, mask0), (is_fpclass x, mask1)1593///     -> is_fpclass x, (mask0 | mask1)1594/// and (is_fpclass x, mask0), (is_fpclass x, mask1)1595///     -> is_fpclass x, (mask0 & mask1)1596/// xor (is_fpclass x, mask0), (is_fpclass x, mask1)1597///     -> is_fpclass x, (mask0 ^ mask1)1598Instruction *InstCombinerImpl::foldLogicOfIsFPClass(BinaryOperator &BO,1599                                                    Value *Op0, Value *Op1) {1600  Value *ClassVal0 = nullptr;1601  Value *ClassVal1 = nullptr;1602  uint64_t ClassMask0, ClassMask1;1603 1604  // Restrict to folding one fcmp into one is.fpclass for now, don't introduce a1605  // new class.1606  //1607  // TODO: Support forming is.fpclass out of 2 separate fcmps when codegen is1608  // better.1609 1610  bool IsLHSClass =1611      match(Op0, m_OneUse(m_Intrinsic<Intrinsic::is_fpclass>(1612                     m_Value(ClassVal0), m_ConstantInt(ClassMask0))));1613  bool IsRHSClass =1614      match(Op1, m_OneUse(m_Intrinsic<Intrinsic::is_fpclass>(1615                     m_Value(ClassVal1), m_ConstantInt(ClassMask1))));1616  if ((((IsLHSClass || matchIsFPClassLikeFCmp(Op0, ClassVal0, ClassMask0)) &&1617        (IsRHSClass || matchIsFPClassLikeFCmp(Op1, ClassVal1, ClassMask1)))) &&1618      ClassVal0 == ClassVal1) {1619    unsigned NewClassMask;1620    switch (BO.getOpcode()) {1621    case Instruction::And:1622      NewClassMask = ClassMask0 & ClassMask1;1623      break;1624    case Instruction::Or:1625      NewClassMask = ClassMask0 | ClassMask1;1626      break;1627    case Instruction::Xor:1628      NewClassMask = ClassMask0 ^ ClassMask1;1629      break;1630    default:1631      llvm_unreachable("not a binary logic operator");1632    }1633 1634    if (IsLHSClass) {1635      auto *II = cast<IntrinsicInst>(Op0);1636      II->setArgOperand(1637          1, ConstantInt::get(II->getArgOperand(1)->getType(), NewClassMask));1638      return replaceInstUsesWith(BO, II);1639    }1640 1641    if (IsRHSClass) {1642      auto *II = cast<IntrinsicInst>(Op1);1643      II->setArgOperand(1644          1, ConstantInt::get(II->getArgOperand(1)->getType(), NewClassMask));1645      return replaceInstUsesWith(BO, II);1646    }1647 1648    CallInst *NewClass =1649        Builder.CreateIntrinsic(Intrinsic::is_fpclass, {ClassVal0->getType()},1650                                {ClassVal0, Builder.getInt32(NewClassMask)});1651    return replaceInstUsesWith(BO, NewClass);1652  }1653 1654  return nullptr;1655}1656 1657/// Look for the pattern that conditionally negates a value via math operations:1658///   cond.splat = sext i1 cond1659///   sub = add cond.splat, x1660///   xor = xor sub, cond.splat1661/// and rewrite it to do the same, but via logical operations:1662///   value.neg = sub 0, value1663///   cond = select i1 neg, value.neg, value1664Instruction *InstCombinerImpl::canonicalizeConditionalNegationViaMathToSelect(1665    BinaryOperator &I) {1666  assert(I.getOpcode() == BinaryOperator::Xor && "Only for xor!");1667  Value *Cond, *X;1668  // As per complexity ordering, `xor` is not commutative here.1669  if (!match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())) ||1670      !match(I.getOperand(1), m_SExt(m_Value(Cond))) ||1671      !Cond->getType()->isIntOrIntVectorTy(1) ||1672      !match(I.getOperand(0), m_c_Add(m_SExt(m_Specific(Cond)), m_Value(X))))1673    return nullptr;1674  return SelectInst::Create(Cond, Builder.CreateNeg(X, X->getName() + ".neg"),1675                            X);1676}1677 1678/// This a limited reassociation for a special case (see above) where we are1679/// checking if two values are either both NAN (unordered) or not-NAN (ordered).1680/// This could be handled more generally in '-reassociation', but it seems like1681/// an unlikely pattern for a large number of logic ops and fcmps.1682static Instruction *reassociateFCmps(BinaryOperator &BO,1683                                     InstCombiner::BuilderTy &Builder) {1684  Instruction::BinaryOps Opcode = BO.getOpcode();1685  assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&1686         "Expecting and/or op for fcmp transform");1687 1688  // There are 4 commuted variants of the pattern. Canonicalize operands of this1689  // logic op so an fcmp is operand 0 and a matching logic op is operand 1.1690  Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1), *X;1691  if (match(Op1, m_FCmp(m_Value(), m_AnyZeroFP())))1692    std::swap(Op0, Op1);1693 1694  // Match inner binop and the predicate for combining 2 NAN checks into 1.1695  Value *BO10, *BO11;1696  FCmpInst::Predicate NanPred = Opcode == Instruction::And ? FCmpInst::FCMP_ORD1697                                                           : FCmpInst::FCMP_UNO;1698  if (!match(Op0, m_SpecificFCmp(NanPred, m_Value(X), m_AnyZeroFP())) ||1699      !match(Op1, m_BinOp(Opcode, m_Value(BO10), m_Value(BO11))))1700    return nullptr;1701 1702  // The inner logic op must have a matching fcmp operand.1703  Value *Y;1704  if (!match(BO10, m_SpecificFCmp(NanPred, m_Value(Y), m_AnyZeroFP())) ||1705      X->getType() != Y->getType())1706    std::swap(BO10, BO11);1707 1708  if (!match(BO10, m_SpecificFCmp(NanPred, m_Value(Y), m_AnyZeroFP())) ||1709      X->getType() != Y->getType())1710    return nullptr;1711 1712  // and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z1713  // or  (fcmp uno X, 0), (or  (fcmp uno Y, 0), Z) --> or  (fcmp uno X, Y), Z1714  // Intersect FMF from the 2 source fcmps.1715  Value *NewFCmp =1716      Builder.CreateFCmpFMF(NanPred, X, Y, FMFSource::intersect(Op0, BO10));1717  return BinaryOperator::Create(Opcode, NewFCmp, BO11);1718}1719 1720/// Match variations of De Morgan's Laws:1721/// (~A & ~B) == (~(A | B))1722/// (~A | ~B) == (~(A & B))1723static Instruction *matchDeMorgansLaws(BinaryOperator &I,1724                                       InstCombiner &IC) {1725  const Instruction::BinaryOps Opcode = I.getOpcode();1726  assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&1727         "Trying to match De Morgan's Laws with something other than and/or");1728 1729  // Flip the logic operation.1730  const Instruction::BinaryOps FlippedOpcode =1731      (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;1732 1733  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);1734  Value *A, *B;1735  if (match(Op0, m_OneUse(m_Not(m_Value(A)))) &&1736      match(Op1, m_OneUse(m_Not(m_Value(B)))) &&1737      !IC.isFreeToInvert(A, A->hasOneUse()) &&1738      !IC.isFreeToInvert(B, B->hasOneUse())) {1739    Value *AndOr =1740        IC.Builder.CreateBinOp(FlippedOpcode, A, B, I.getName() + ".demorgan");1741    return BinaryOperator::CreateNot(AndOr);1742  }1743 1744  // The 'not' ops may require reassociation.1745  // (A & ~B) & ~C --> A & ~(B | C)1746  // (~B & A) & ~C --> A & ~(B | C)1747  // (A | ~B) | ~C --> A | ~(B & C)1748  // (~B | A) | ~C --> A | ~(B & C)1749  Value *C;1750  if (match(Op0, m_OneUse(m_c_BinOp(Opcode, m_Value(A), m_Not(m_Value(B))))) &&1751      match(Op1, m_Not(m_Value(C)))) {1752    Value *FlippedBO = IC.Builder.CreateBinOp(FlippedOpcode, B, C);1753    return BinaryOperator::Create(Opcode, A, IC.Builder.CreateNot(FlippedBO));1754  }1755 1756  return nullptr;1757}1758 1759bool InstCombinerImpl::shouldOptimizeCast(CastInst *CI) {1760  Value *CastSrc = CI->getOperand(0);1761 1762  // Noop casts and casts of constants should be eliminated trivially.1763  if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc))1764    return false;1765 1766  // If this cast is paired with another cast that can be eliminated, we prefer1767  // to have it eliminated.1768  if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc))1769    if (isEliminableCastPair(PrecedingCI, CI))1770      return false;1771 1772  return true;1773}1774 1775/// Fold {and,or,xor} (cast X), C.1776static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast,1777                                          InstCombinerImpl &IC) {1778  Constant *C = dyn_cast<Constant>(Logic.getOperand(1));1779  if (!C)1780    return nullptr;1781 1782  auto LogicOpc = Logic.getOpcode();1783  Type *DestTy = Logic.getType();1784  Type *SrcTy = Cast->getSrcTy();1785 1786  // Move the logic operation ahead of a zext or sext if the constant is1787  // unchanged in the smaller source type. Performing the logic in a smaller1788  // type may provide more information to later folds, and the smaller logic1789  // instruction may be cheaper (particularly in the case of vectors).1790  Value *X;1791  auto &DL = IC.getDataLayout();1792  if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) {1793    PreservedCastFlags Flags;1794    if (Constant *TruncC = getLosslessUnsignedTrunc(C, SrcTy, DL, &Flags)) {1795      // LogicOpc (zext X), C --> zext (LogicOpc X, C)1796      Value *NewOp = IC.Builder.CreateBinOp(LogicOpc, X, TruncC);1797      auto *ZExt = new ZExtInst(NewOp, DestTy);1798      ZExt->setNonNeg(Flags.NNeg);1799      ZExt->andIRFlags(Cast);1800      return ZExt;1801    }1802  }1803 1804  if (match(Cast, m_OneUse(m_SExtLike(m_Value(X))))) {1805    if (Constant *TruncC = getLosslessSignedTrunc(C, SrcTy, DL)) {1806      // LogicOpc (sext X), C --> sext (LogicOpc X, C)1807      Value *NewOp = IC.Builder.CreateBinOp(LogicOpc, X, TruncC);1808      return new SExtInst(NewOp, DestTy);1809    }1810  }1811 1812  return nullptr;1813}1814 1815/// Fold {and,or,xor} (cast X), Y.1816Instruction *InstCombinerImpl::foldCastedBitwiseLogic(BinaryOperator &I) {1817  auto LogicOpc = I.getOpcode();1818  assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding");1819 1820  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);1821 1822  // fold bitwise(A >> BW - 1, zext(icmp))     (BW is the scalar bits of the1823  // type of A)1824  //   -> bitwise(zext(A < 0), zext(icmp))1825  //   -> zext(bitwise(A < 0, icmp))1826  auto FoldBitwiseICmpZeroWithICmp = [&](Value *Op0,1827                                         Value *Op1) -> Instruction * {1828    Value *A;1829    bool IsMatched =1830        match(Op0,1831              m_OneUse(m_LShr(1832                  m_Value(A),1833                  m_SpecificInt(Op0->getType()->getScalarSizeInBits() - 1)))) &&1834        match(Op1, m_OneUse(m_ZExt(m_ICmp(m_Value(), m_Value()))));1835 1836    if (!IsMatched)1837      return nullptr;1838 1839    auto *ICmpL =1840        Builder.CreateICmpSLT(A, Constant::getNullValue(A->getType()));1841    auto *ICmpR = cast<ZExtInst>(Op1)->getOperand(0);1842    auto *BitwiseOp = Builder.CreateBinOp(LogicOpc, ICmpL, ICmpR);1843 1844    return new ZExtInst(BitwiseOp, Op0->getType());1845  };1846 1847  if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op0, Op1))1848    return Ret;1849 1850  if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op1, Op0))1851    return Ret;1852 1853  CastInst *Cast0 = dyn_cast<CastInst>(Op0);1854  if (!Cast0)1855    return nullptr;1856 1857  // This must be a cast from an integer or integer vector source type to allow1858  // transformation of the logic operation to the source type.1859  Type *DestTy = I.getType();1860  Type *SrcTy = Cast0->getSrcTy();1861  if (!SrcTy->isIntOrIntVectorTy())1862    return nullptr;1863 1864  if (Instruction *Ret = foldLogicCastConstant(I, Cast0, *this))1865    return Ret;1866 1867  CastInst *Cast1 = dyn_cast<CastInst>(Op1);1868  if (!Cast1)1869    return nullptr;1870 1871  // Both operands of the logic operation are casts. The casts must be the1872  // same kind for reduction.1873  Instruction::CastOps CastOpcode = Cast0->getOpcode();1874  if (CastOpcode != Cast1->getOpcode())1875    return nullptr;1876 1877  // Can't fold it profitably if no one of casts has one use.1878  if (!Cast0->hasOneUse() && !Cast1->hasOneUse())1879    return nullptr;1880 1881  Value *X, *Y;1882  if (match(Cast0, m_ZExtOrSExt(m_Value(X))) &&1883      match(Cast1, m_ZExtOrSExt(m_Value(Y)))) {1884    // Cast the narrower source to the wider source type.1885    unsigned XNumBits = X->getType()->getScalarSizeInBits();1886    unsigned YNumBits = Y->getType()->getScalarSizeInBits();1887    if (XNumBits != YNumBits) {1888      // Cast the narrower source to the wider source type only if both of casts1889      // have one use to avoid creating an extra instruction.1890      if (!Cast0->hasOneUse() || !Cast1->hasOneUse())1891        return nullptr;1892 1893      // If the source types do not match, but the casts are matching extends,1894      // we can still narrow the logic op.1895      if (XNumBits < YNumBits) {1896        X = Builder.CreateCast(CastOpcode, X, Y->getType());1897      } else if (YNumBits < XNumBits) {1898        Y = Builder.CreateCast(CastOpcode, Y, X->getType());1899      }1900    }1901 1902    // Do the logic op in the intermediate width, then widen more.1903    Value *NarrowLogic = Builder.CreateBinOp(LogicOpc, X, Y, I.getName());1904    auto *Disjoint = dyn_cast<PossiblyDisjointInst>(&I);1905    auto *NewDisjoint = dyn_cast<PossiblyDisjointInst>(NarrowLogic);1906    if (Disjoint && NewDisjoint)1907      NewDisjoint->setIsDisjoint(Disjoint->isDisjoint());1908    return CastInst::Create(CastOpcode, NarrowLogic, DestTy);1909  }1910 1911  // If the src type of casts are different, give up for other cast opcodes.1912  if (SrcTy != Cast1->getSrcTy())1913    return nullptr;1914 1915  Value *Cast0Src = Cast0->getOperand(0);1916  Value *Cast1Src = Cast1->getOperand(0);1917 1918  // fold logic(cast(A), cast(B)) -> cast(logic(A, B))1919  if (shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {1920    Value *NewOp = Builder.CreateBinOp(LogicOpc, Cast0Src, Cast1Src,1921                                       I.getName());1922    return CastInst::Create(CastOpcode, NewOp, DestTy);1923  }1924 1925  return nullptr;1926}1927 1928static Instruction *foldAndToXor(BinaryOperator &I,1929                                 InstCombiner::BuilderTy &Builder) {1930  assert(I.getOpcode() == Instruction::And);1931  Value *Op0 = I.getOperand(0);1932  Value *Op1 = I.getOperand(1);1933  Value *A, *B;1934 1935  // Operand complexity canonicalization guarantees that the 'or' is Op0.1936  // (A | B) & ~(A & B) --> A ^ B1937  // (A | B) & ~(B & A) --> A ^ B1938  if (match(&I, m_BinOp(m_Or(m_Value(A), m_Value(B)),1939                        m_Not(m_c_And(m_Deferred(A), m_Deferred(B))))))1940    return BinaryOperator::CreateXor(A, B);1941 1942  // (A | ~B) & (~A | B) --> ~(A ^ B)1943  // (A | ~B) & (B | ~A) --> ~(A ^ B)1944  // (~B | A) & (~A | B) --> ~(A ^ B)1945  // (~B | A) & (B | ~A) --> ~(A ^ B)1946  if (Op0->hasOneUse() || Op1->hasOneUse())1947    if (match(&I, m_BinOp(m_c_Or(m_Value(A), m_Not(m_Value(B))),1948                          m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))1949      return BinaryOperator::CreateNot(Builder.CreateXor(A, B));1950 1951  return nullptr;1952}1953 1954static Instruction *foldOrToXor(BinaryOperator &I,1955                                InstCombiner::BuilderTy &Builder) {1956  assert(I.getOpcode() == Instruction::Or);1957  Value *Op0 = I.getOperand(0);1958  Value *Op1 = I.getOperand(1);1959  Value *A, *B;1960 1961  // Operand complexity canonicalization guarantees that the 'and' is Op0.1962  // (A & B) | ~(A | B) --> ~(A ^ B)1963  // (A & B) | ~(B | A) --> ~(A ^ B)1964  if (Op0->hasOneUse() || Op1->hasOneUse())1965    if (match(Op0, m_And(m_Value(A), m_Value(B))) &&1966        match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))1967      return BinaryOperator::CreateNot(Builder.CreateXor(A, B));1968 1969  // Operand complexity canonicalization guarantees that the 'xor' is Op0.1970  // (A ^ B) | ~(A | B) --> ~(A & B)1971  // (A ^ B) | ~(B | A) --> ~(A & B)1972  if (Op0->hasOneUse() || Op1->hasOneUse())1973    if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&1974        match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))1975      return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));1976 1977  // (A & ~B) | (~A & B) --> A ^ B1978  // (A & ~B) | (B & ~A) --> A ^ B1979  // (~B & A) | (~A & B) --> A ^ B1980  // (~B & A) | (B & ~A) --> A ^ B1981  if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&1982      match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B))))1983    return BinaryOperator::CreateXor(A, B);1984 1985  return nullptr;1986}1987 1988/// Return true if a constant shift amount is always less than the specified1989/// bit-width. If not, the shift could create poison in the narrower type.1990static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) {1991  APInt Threshold(C->getType()->getScalarSizeInBits(), BitWidth);1992  return match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold));1993}1994 1995/// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and1996/// a common zext operand: and (binop (zext X), C), (zext X).1997Instruction *InstCombinerImpl::narrowMaskedBinOp(BinaryOperator &And) {1998  // This transform could also apply to {or, and, xor}, but there are better1999  // folds for those cases, so we don't expect those patterns here. AShr is not2000  // handled because it should always be transformed to LShr in this sequence.2001  // The subtract transform is different because it has a constant on the left.2002  // Add/mul commute the constant to RHS; sub with constant RHS becomes add.2003  Value *Op0 = And.getOperand(0), *Op1 = And.getOperand(1);2004  Constant *C;2005  if (!match(Op0, m_OneUse(m_Add(m_Specific(Op1), m_Constant(C)))) &&2006      !match(Op0, m_OneUse(m_Mul(m_Specific(Op1), m_Constant(C)))) &&2007      !match(Op0, m_OneUse(m_LShr(m_Specific(Op1), m_Constant(C)))) &&2008      !match(Op0, m_OneUse(m_Shl(m_Specific(Op1), m_Constant(C)))) &&2009      !match(Op0, m_OneUse(m_Sub(m_Constant(C), m_Specific(Op1)))))2010    return nullptr;2011 2012  Value *X;2013  if (!match(Op1, m_ZExt(m_Value(X))) || Op1->hasNUsesOrMore(3))2014    return nullptr;2015 2016  Type *Ty = And.getType();2017  if (!isa<VectorType>(Ty) && !shouldChangeType(Ty, X->getType()))2018    return nullptr;2019 2020  // If we're narrowing a shift, the shift amount must be safe (less than the2021  // width) in the narrower type. If the shift amount is greater, instsimplify2022  // usually handles that case, but we can't guarantee/assert it.2023  Instruction::BinaryOps Opc = cast<BinaryOperator>(Op0)->getOpcode();2024  if (Opc == Instruction::LShr || Opc == Instruction::Shl)2025    if (!canNarrowShiftAmt(C, X->getType()->getScalarSizeInBits()))2026      return nullptr;2027 2028  // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X)2029  // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X)2030  Value *NewC = ConstantExpr::getTrunc(C, X->getType());2031  Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, NewC, X)2032                                         : Builder.CreateBinOp(Opc, X, NewC);2033  return new ZExtInst(Builder.CreateAnd(NewBO, X), Ty);2034}2035 2036/// Try folding relatively complex patterns for both And and Or operations2037/// with all And and Or swapped.2038static Instruction *foldComplexAndOrPatterns(BinaryOperator &I,2039                                             InstCombiner::BuilderTy &Builder) {2040  const Instruction::BinaryOps Opcode = I.getOpcode();2041  assert(Opcode == Instruction::And || Opcode == Instruction::Or);2042 2043  // Flip the logic operation.2044  const Instruction::BinaryOps FlippedOpcode =2045      (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;2046 2047  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);2048  Value *A, *B, *C, *X, *Y, *Dummy;2049 2050  // Match following expressions:2051  // (~(A | B) & C)2052  // (~(A & B) | C)2053  // Captures X = ~(A | B) or ~(A & B)2054  const auto matchNotOrAnd =2055      [Opcode, FlippedOpcode](Value *Op, auto m_A, auto m_B, auto m_C,2056                              Value *&X, bool CountUses = false) -> bool {2057    if (CountUses && !Op->hasOneUse())2058      return false;2059 2060    if (match(Op,2061              m_c_BinOp(FlippedOpcode,2062                        m_Value(X, m_Not(m_c_BinOp(Opcode, m_A, m_B))), m_C)))2063      return !CountUses || X->hasOneUse();2064 2065    return false;2066  };2067 2068  // (~(A | B) & C) | ... --> ...2069  // (~(A & B) | C) & ... --> ...2070  // TODO: One use checks are conservative. We just need to check that a total2071  //       number of multiple used values does not exceed reduction2072  //       in operations.2073  if (matchNotOrAnd(Op0, m_Value(A), m_Value(B), m_Value(C), X)) {2074    // (~(A | B) & C) | (~(A | C) & B) --> (B ^ C) & ~A2075    // (~(A & B) | C) & (~(A & C) | B) --> ~((B ^ C) & A)2076    if (matchNotOrAnd(Op1, m_Specific(A), m_Specific(C), m_Specific(B), Dummy,2077                      true)) {2078      Value *Xor = Builder.CreateXor(B, C);2079      return (Opcode == Instruction::Or)2080                 ? BinaryOperator::CreateAnd(Xor, Builder.CreateNot(A))2081                 : BinaryOperator::CreateNot(Builder.CreateAnd(Xor, A));2082    }2083 2084    // (~(A | B) & C) | (~(B | C) & A) --> (A ^ C) & ~B2085    // (~(A & B) | C) & (~(B & C) | A) --> ~((A ^ C) & B)2086    if (matchNotOrAnd(Op1, m_Specific(B), m_Specific(C), m_Specific(A), Dummy,2087                      true)) {2088      Value *Xor = Builder.CreateXor(A, C);2089      return (Opcode == Instruction::Or)2090                 ? BinaryOperator::CreateAnd(Xor, Builder.CreateNot(B))2091                 : BinaryOperator::CreateNot(Builder.CreateAnd(Xor, B));2092    }2093 2094    // (~(A | B) & C) | ~(A | C) --> ~((B & C) | A)2095    // (~(A & B) | C) & ~(A & C) --> ~((B | C) & A)2096    if (match(Op1, m_OneUse(m_Not(m_OneUse(2097                       m_c_BinOp(Opcode, m_Specific(A), m_Specific(C)))))))2098      return BinaryOperator::CreateNot(Builder.CreateBinOp(2099          Opcode, Builder.CreateBinOp(FlippedOpcode, B, C), A));2100 2101    // (~(A | B) & C) | ~(B | C) --> ~((A & C) | B)2102    // (~(A & B) | C) & ~(B & C) --> ~((A | C) & B)2103    if (match(Op1, m_OneUse(m_Not(m_OneUse(2104                       m_c_BinOp(Opcode, m_Specific(B), m_Specific(C)))))))2105      return BinaryOperator::CreateNot(Builder.CreateBinOp(2106          Opcode, Builder.CreateBinOp(FlippedOpcode, A, C), B));2107 2108    // (~(A | B) & C) | ~(C | (A ^ B)) --> ~((A | B) & (C | (A ^ B)))2109    // Note, the pattern with swapped and/or is not handled because the2110    // result is more undefined than a source:2111    // (~(A & B) | C) & ~(C & (A ^ B)) --> (A ^ B ^ C) | ~(A | C) is invalid.2112    if (Opcode == Instruction::Or && Op0->hasOneUse() &&2113        match(Op1,2114              m_OneUse(m_Not(m_Value(2115                  Y, m_c_BinOp(Opcode, m_Specific(C),2116                               m_c_Xor(m_Specific(A), m_Specific(B)))))))) {2117      // X = ~(A | B)2118      // Y = (C | (A ^ B)2119      Value *Or = cast<BinaryOperator>(X)->getOperand(0);2120      return BinaryOperator::CreateNot(Builder.CreateAnd(Or, Y));2121    }2122  }2123 2124  // (~A & B & C) | ... --> ...2125  // (~A | B | C) | ... --> ...2126  // TODO: One use checks are conservative. We just need to check that a total2127  //       number of multiple used values does not exceed reduction2128  //       in operations.2129  if (match(Op0,2130            m_OneUse(m_c_BinOp(FlippedOpcode,2131                               m_BinOp(FlippedOpcode, m_Value(B), m_Value(C)),2132                               m_Value(X, m_Not(m_Value(A)))))) ||2133      match(Op0, m_OneUse(m_c_BinOp(FlippedOpcode,2134                                    m_c_BinOp(FlippedOpcode, m_Value(C),2135                                              m_Value(X, m_Not(m_Value(A)))),2136                                    m_Value(B))))) {2137    // X = ~A2138    // (~A & B & C) | ~(A | B | C) --> ~(A | (B ^ C))2139    // (~A | B | C) & ~(A & B & C) --> (~A | (B ^ C))2140    if (match(Op1, m_OneUse(m_Not(m_c_BinOp(2141                       Opcode, m_c_BinOp(Opcode, m_Specific(A), m_Specific(B)),2142                       m_Specific(C))))) ||2143        match(Op1, m_OneUse(m_Not(m_c_BinOp(2144                       Opcode, m_c_BinOp(Opcode, m_Specific(B), m_Specific(C)),2145                       m_Specific(A))))) ||2146        match(Op1, m_OneUse(m_Not(m_c_BinOp(2147                       Opcode, m_c_BinOp(Opcode, m_Specific(A), m_Specific(C)),2148                       m_Specific(B)))))) {2149      Value *Xor = Builder.CreateXor(B, C);2150      return (Opcode == Instruction::Or)2151                 ? BinaryOperator::CreateNot(Builder.CreateOr(Xor, A))2152                 : BinaryOperator::CreateOr(Xor, X);2153    }2154 2155    // (~A & B & C) | ~(A | B) --> (C | ~B) & ~A2156    // (~A | B | C) & ~(A & B) --> (C & ~B) | ~A2157    if (match(Op1, m_OneUse(m_Not(m_OneUse(2158                       m_c_BinOp(Opcode, m_Specific(A), m_Specific(B)))))))2159      return BinaryOperator::Create(2160          FlippedOpcode, Builder.CreateBinOp(Opcode, C, Builder.CreateNot(B)),2161          X);2162 2163    // (~A & B & C) | ~(A | C) --> (B | ~C) & ~A2164    // (~A | B | C) & ~(A & C) --> (B & ~C) | ~A2165    if (match(Op1, m_OneUse(m_Not(m_OneUse(2166                       m_c_BinOp(Opcode, m_Specific(A), m_Specific(C)))))))2167      return BinaryOperator::Create(2168          FlippedOpcode, Builder.CreateBinOp(Opcode, B, Builder.CreateNot(C)),2169          X);2170  }2171 2172  return nullptr;2173}2174 2175/// Try to reassociate a pair of binops so that values with one use only are2176/// part of the same instruction. This may enable folds that are limited with2177/// multi-use restrictions and makes it more likely to match other patterns that2178/// are looking for a common operand.2179static Instruction *reassociateForUses(BinaryOperator &BO,2180                                       InstCombinerImpl::BuilderTy &Builder) {2181  Instruction::BinaryOps Opcode = BO.getOpcode();2182  Value *X, *Y, *Z;2183  if (match(&BO,2184            m_c_BinOp(Opcode, m_OneUse(m_BinOp(Opcode, m_Value(X), m_Value(Y))),2185                      m_OneUse(m_Value(Z))))) {2186    if (!isa<Constant>(X) && !isa<Constant>(Y) && !isa<Constant>(Z)) {2187      // (X op Y) op Z --> (Y op Z) op X2188      if (!X->hasOneUse()) {2189        Value *YZ = Builder.CreateBinOp(Opcode, Y, Z);2190        return BinaryOperator::Create(Opcode, YZ, X);2191      }2192      // (X op Y) op Z --> (X op Z) op Y2193      if (!Y->hasOneUse()) {2194        Value *XZ = Builder.CreateBinOp(Opcode, X, Z);2195        return BinaryOperator::Create(Opcode, XZ, Y);2196      }2197    }2198  }2199 2200  return nullptr;2201}2202 2203// Match2204// (X + C2) | C2205// (X + C2) ^ C2206// (X + C2) & C2207// and convert to do the bitwise logic first:2208// (X | C) + C22209// (X ^ C) + C22210// (X & C) + C22211// iff bits affected by logic op are lower than last bit affected by math op2212static Instruction *canonicalizeLogicFirst(BinaryOperator &I,2213                                           InstCombiner::BuilderTy &Builder) {2214  Type *Ty = I.getType();2215  Instruction::BinaryOps OpC = I.getOpcode();2216  Value *Op0 = I.getOperand(0);2217  Value *Op1 = I.getOperand(1);2218  Value *X;2219  const APInt *C, *C2;2220 2221  if (!(match(Op0, m_OneUse(m_Add(m_Value(X), m_APInt(C2)))) &&2222        match(Op1, m_APInt(C))))2223    return nullptr;2224 2225  unsigned Width = Ty->getScalarSizeInBits();2226  unsigned LastOneMath = Width - C2->countr_zero();2227 2228  switch (OpC) {2229  case Instruction::And:2230    if (C->countl_one() < LastOneMath)2231      return nullptr;2232    break;2233  case Instruction::Xor:2234  case Instruction::Or:2235    if (C->countl_zero() < LastOneMath)2236      return nullptr;2237    break;2238  default:2239    llvm_unreachable("Unexpected BinaryOp!");2240  }2241 2242  Value *NewBinOp = Builder.CreateBinOp(OpC, X, ConstantInt::get(Ty, *C));2243  return BinaryOperator::CreateWithCopiedFlags(Instruction::Add, NewBinOp,2244                                               ConstantInt::get(Ty, *C2), Op0);2245}2246 2247// binop(shift(ShiftedC1, ShAmt), shift(ShiftedC2, add(ShAmt, AddC))) ->2248// shift(binop(ShiftedC1, shift(ShiftedC2, AddC)), ShAmt)2249// where both shifts are the same and AddC is a valid shift amount.2250Instruction *InstCombinerImpl::foldBinOpOfDisplacedShifts(BinaryOperator &I) {2251  assert((I.isBitwiseLogicOp() || I.getOpcode() == Instruction::Add) &&2252         "Unexpected opcode");2253 2254  Value *ShAmt;2255  Constant *ShiftedC1, *ShiftedC2, *AddC;2256  Type *Ty = I.getType();2257  unsigned BitWidth = Ty->getScalarSizeInBits();2258  if (!match(&I, m_c_BinOp(m_Shift(m_ImmConstant(ShiftedC1), m_Value(ShAmt)),2259                           m_Shift(m_ImmConstant(ShiftedC2),2260                                   m_AddLike(m_Deferred(ShAmt),2261                                             m_ImmConstant(AddC))))))2262    return nullptr;2263 2264  // Make sure the add constant is a valid shift amount.2265  if (!match(AddC,2266             m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(BitWidth, BitWidth))))2267    return nullptr;2268 2269  // Avoid constant expressions.2270  auto *Op0Inst = dyn_cast<Instruction>(I.getOperand(0));2271  auto *Op1Inst = dyn_cast<Instruction>(I.getOperand(1));2272  if (!Op0Inst || !Op1Inst)2273    return nullptr;2274 2275  // Both shifts must be the same.2276  Instruction::BinaryOps ShiftOp =2277      static_cast<Instruction::BinaryOps>(Op0Inst->getOpcode());2278  if (ShiftOp != Op1Inst->getOpcode())2279    return nullptr;2280 2281  // For adds, only left shifts are supported.2282  if (I.getOpcode() == Instruction::Add && ShiftOp != Instruction::Shl)2283    return nullptr;2284 2285  Value *NewC = Builder.CreateBinOp(2286      I.getOpcode(), ShiftedC1, Builder.CreateBinOp(ShiftOp, ShiftedC2, AddC));2287  return BinaryOperator::Create(ShiftOp, NewC, ShAmt);2288}2289 2290// Fold and/or/xor with two equal intrinsic IDs:2291// bitwise(fshl (A, B, ShAmt), fshl(C, D, ShAmt))2292// -> fshl(bitwise(A, C), bitwise(B, D), ShAmt)2293// bitwise(fshr (A, B, ShAmt), fshr(C, D, ShAmt))2294// -> fshr(bitwise(A, C), bitwise(B, D), ShAmt)2295// bitwise(bswap(A), bswap(B)) -> bswap(bitwise(A, B))2296// bitwise(bswap(A), C) -> bswap(bitwise(A, bswap(C)))2297// bitwise(bitreverse(A), bitreverse(B)) -> bitreverse(bitwise(A, B))2298// bitwise(bitreverse(A), C) -> bitreverse(bitwise(A, bitreverse(C)))2299static Instruction *2300foldBitwiseLogicWithIntrinsics(BinaryOperator &I,2301                               InstCombiner::BuilderTy &Builder) {2302  assert(I.isBitwiseLogicOp() && "Should and/or/xor");2303  if (!I.getOperand(0)->hasOneUse())2304    return nullptr;2305  IntrinsicInst *X = dyn_cast<IntrinsicInst>(I.getOperand(0));2306  if (!X)2307    return nullptr;2308 2309  IntrinsicInst *Y = dyn_cast<IntrinsicInst>(I.getOperand(1));2310  if (Y && (!Y->hasOneUse() || X->getIntrinsicID() != Y->getIntrinsicID()))2311    return nullptr;2312 2313  Intrinsic::ID IID = X->getIntrinsicID();2314  const APInt *RHSC;2315  // Try to match constant RHS.2316  if (!Y && (!(IID == Intrinsic::bswap || IID == Intrinsic::bitreverse) ||2317             !match(I.getOperand(1), m_APInt(RHSC))))2318    return nullptr;2319 2320  switch (IID) {2321  case Intrinsic::fshl:2322  case Intrinsic::fshr: {2323    if (X->getOperand(2) != Y->getOperand(2))2324      return nullptr;2325    Value *NewOp0 =2326        Builder.CreateBinOp(I.getOpcode(), X->getOperand(0), Y->getOperand(0));2327    Value *NewOp1 =2328        Builder.CreateBinOp(I.getOpcode(), X->getOperand(1), Y->getOperand(1));2329    Function *F =2330        Intrinsic::getOrInsertDeclaration(I.getModule(), IID, I.getType());2331    return CallInst::Create(F, {NewOp0, NewOp1, X->getOperand(2)});2332  }2333  case Intrinsic::bswap:2334  case Intrinsic::bitreverse: {2335    Value *NewOp0 = Builder.CreateBinOp(2336        I.getOpcode(), X->getOperand(0),2337        Y ? Y->getOperand(0)2338          : ConstantInt::get(I.getType(), IID == Intrinsic::bswap2339                                              ? RHSC->byteSwap()2340                                              : RHSC->reverseBits()));2341    Function *F =2342        Intrinsic::getOrInsertDeclaration(I.getModule(), IID, I.getType());2343    return CallInst::Create(F, {NewOp0});2344  }2345  default:2346    return nullptr;2347  }2348}2349 2350// Try to simplify V by replacing occurrences of Op with RepOp, but only look2351// through bitwise operations. In particular, for X | Y we try to replace Y with2352// 0 inside X and for X & Y we try to replace Y with -1 inside X.2353// Return the simplified result of X if successful, and nullptr otherwise.2354// If SimplifyOnly is true, no new instructions will be created.2355static Value *simplifyAndOrWithOpReplaced(Value *V, Value *Op, Value *RepOp,2356                                          bool SimplifyOnly,2357                                          InstCombinerImpl &IC,2358                                          unsigned Depth = 0) {2359  if (Op == RepOp)2360    return nullptr;2361 2362  if (V == Op)2363    return RepOp;2364 2365  auto *I = dyn_cast<BinaryOperator>(V);2366  if (!I || !I->isBitwiseLogicOp() || Depth >= 3)2367    return nullptr;2368 2369  if (!I->hasOneUse())2370    SimplifyOnly = true;2371 2372  Value *NewOp0 = simplifyAndOrWithOpReplaced(I->getOperand(0), Op, RepOp,2373                                              SimplifyOnly, IC, Depth + 1);2374  Value *NewOp1 = simplifyAndOrWithOpReplaced(I->getOperand(1), Op, RepOp,2375                                              SimplifyOnly, IC, Depth + 1);2376  if (!NewOp0 && !NewOp1)2377    return nullptr;2378 2379  if (!NewOp0)2380    NewOp0 = I->getOperand(0);2381  if (!NewOp1)2382    NewOp1 = I->getOperand(1);2383 2384  if (Value *Res = simplifyBinOp(I->getOpcode(), NewOp0, NewOp1,2385                                 IC.getSimplifyQuery().getWithInstruction(I)))2386    return Res;2387 2388  if (SimplifyOnly)2389    return nullptr;2390  return IC.Builder.CreateBinOp(I->getOpcode(), NewOp0, NewOp1);2391}2392 2393/// Reassociate and/or expressions to see if we can fold the inner and/or ops.2394/// TODO: Make this recursive; it's a little tricky because an arbitrary2395/// number of and/or instructions might have to be created.2396Value *InstCombinerImpl::reassociateBooleanAndOr(Value *LHS, Value *X, Value *Y,2397                                                 Instruction &I, bool IsAnd,2398                                                 bool RHSIsLogical) {2399  Instruction::BinaryOps Opcode = IsAnd ? Instruction::And : Instruction::Or;2400  // LHS bop (X lop Y) --> (LHS bop X) lop Y2401  // LHS bop (X bop Y) --> (LHS bop X) bop Y2402  if (Value *Res = foldBooleanAndOr(LHS, X, I, IsAnd, /*IsLogical=*/false))2403    return RHSIsLogical ? Builder.CreateLogicalOp(Opcode, Res, Y)2404                        : Builder.CreateBinOp(Opcode, Res, Y);2405  // LHS bop (X bop Y) --> X bop (LHS bop Y)2406  // LHS bop (X lop Y) --> X lop (LHS bop Y)2407  if (Value *Res = foldBooleanAndOr(LHS, Y, I, IsAnd, /*IsLogical=*/false))2408    return RHSIsLogical ? Builder.CreateLogicalOp(Opcode, X, Res)2409                        : Builder.CreateBinOp(Opcode, X, Res);2410  return nullptr;2411}2412 2413// FIXME: We use commutative matchers (m_c_*) for some, but not all, matches2414// here. We should standardize that construct where it is needed or choose some2415// other way to ensure that commutated variants of patterns are not missed.2416Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {2417  Type *Ty = I.getType();2418 2419  if (Value *V = simplifyAndInst(I.getOperand(0), I.getOperand(1),2420                                 SQ.getWithInstruction(&I)))2421    return replaceInstUsesWith(I, V);2422 2423  if (SimplifyAssociativeOrCommutative(I))2424    return &I;2425 2426  if (Instruction *X = foldVectorBinop(I))2427    return X;2428 2429  if (Instruction *Phi = foldBinopWithPhiOperands(I))2430    return Phi;2431 2432  // See if we can simplify any instructions used by the instruction whose sole2433  // purpose is to compute bits we don't care about.2434  if (SimplifyDemandedInstructionBits(I))2435    return &I;2436 2437  // Do this before using distributive laws to catch simple and/or/not patterns.2438  if (Instruction *Xor = foldAndToXor(I, Builder))2439    return Xor;2440 2441  if (Instruction *X = foldComplexAndOrPatterns(I, Builder))2442    return X;2443 2444  // (A|B)&(A|C) -> A|(B&C) etc2445  if (Value *V = foldUsingDistributiveLaws(I))2446    return replaceInstUsesWith(I, V);2447 2448  if (Instruction *R = foldBinOpShiftWithShift(I))2449    return R;2450 2451  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);2452 2453  Value *X, *Y;2454  const APInt *C;2455  if ((match(Op0, m_OneUse(m_LogicalShift(m_One(), m_Value(X)))) ||2456       (match(Op0, m_OneUse(m_Shl(m_APInt(C), m_Value(X)))) && (*C)[0])) &&2457      match(Op1, m_One())) {2458    // (1 >> X) & 1 --> zext(X == 0)2459    // (C << X) & 1 --> zext(X == 0), when C is odd2460    Value *IsZero = Builder.CreateICmpEQ(X, ConstantInt::get(Ty, 0));2461    return new ZExtInst(IsZero, Ty);2462  }2463 2464  // (-(X & 1)) & Y --> (X & 1) == 0 ? 0 : Y2465  Value *Neg;2466  if (match(&I,2467            m_c_And(m_Value(Neg, m_OneUse(m_Neg(m_And(m_Value(), m_One())))),2468                    m_Value(Y)))) {2469    Value *Cmp = Builder.CreateIsNull(Neg);2470    return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Y);2471  }2472 2473  // Canonicalize:2474  // (X +/- Y) & Y --> ~X & Y when Y is a power of 2.2475  if (match(&I, m_c_And(m_Value(Y), m_OneUse(m_CombineOr(2476                                        m_c_Add(m_Value(X), m_Deferred(Y)),2477                                        m_Sub(m_Value(X), m_Deferred(Y)))))) &&2478      isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, &I))2479    return BinaryOperator::CreateAnd(Builder.CreateNot(X), Y);2480 2481  if (match(Op1, m_APInt(C))) {2482    const APInt *XorC;2483    if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_APInt(XorC))))) {2484      // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)2485      Constant *NewC = ConstantInt::get(Ty, *C & *XorC);2486      Value *And = Builder.CreateAnd(X, Op1);2487      And->takeName(Op0);2488      return BinaryOperator::CreateXor(And, NewC);2489    }2490 2491    const APInt *OrC;2492    if (match(Op0, m_OneUse(m_Or(m_Value(X), m_APInt(OrC))))) {2493      // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2)2494      // NOTE: This reduces the number of bits set in the & mask, which2495      // can expose opportunities for store narrowing for scalars.2496      // NOTE: SimplifyDemandedBits should have already removed bits from C12497      // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in2498      // above, but this feels safer.2499      APInt Together = *C & *OrC;2500      Value *And = Builder.CreateAnd(X, ConstantInt::get(Ty, Together ^ *C));2501      And->takeName(Op0);2502      return BinaryOperator::CreateOr(And, ConstantInt::get(Ty, Together));2503    }2504 2505    unsigned Width = Ty->getScalarSizeInBits();2506    const APInt *ShiftC;2507    if (match(Op0, m_OneUse(m_SExt(m_AShr(m_Value(X), m_APInt(ShiftC))))) &&2508        ShiftC->ult(Width)) {2509      if (*C == APInt::getLowBitsSet(Width, Width - ShiftC->getZExtValue())) {2510        // We are clearing high bits that were potentially set by sext+ashr:2511        // and (sext (ashr X, ShiftC)), C --> lshr (sext X), ShiftC2512        Value *Sext = Builder.CreateSExt(X, Ty);2513        Constant *ShAmtC = ConstantInt::get(Ty, ShiftC->zext(Width));2514        return BinaryOperator::CreateLShr(Sext, ShAmtC);2515      }2516    }2517 2518    // If this 'and' clears the sign-bits added by ashr, replace with lshr:2519    // and (ashr X, ShiftC), C --> lshr X, ShiftC2520    if (match(Op0, m_AShr(m_Value(X), m_APInt(ShiftC))) && ShiftC->ult(Width) &&2521        C->isMask(Width - ShiftC->getZExtValue()))2522      return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, *ShiftC));2523 2524    const APInt *AddC;2525    if (match(Op0, m_Add(m_Value(X), m_APInt(AddC)))) {2526      // If we are masking the result of the add down to exactly one bit and2527      // the constant we are adding has no bits set below that bit, then the2528      // add is flipping a single bit. Example:2529      // (X + 4) & 4 --> (X & 4) ^ 42530      if (Op0->hasOneUse() && C->isPowerOf2() && (*AddC & (*C - 1)) == 0) {2531        assert((*C & *AddC) != 0 && "Expected common bit");2532        Value *NewAnd = Builder.CreateAnd(X, Op1);2533        return BinaryOperator::CreateXor(NewAnd, Op1);2534      }2535    }2536 2537    // ((C1 OP zext(X)) & C2) -> zext((C1 OP X) & C2) if C2 fits in the2538    // bitwidth of X and OP behaves well when given trunc(C1) and X.2539    auto isNarrowableBinOpcode = [](BinaryOperator *B) {2540      switch (B->getOpcode()) {2541      case Instruction::Xor:2542      case Instruction::Or:2543      case Instruction::Mul:2544      case Instruction::Add:2545      case Instruction::Sub:2546        return true;2547      default:2548        return false;2549      }2550    };2551    BinaryOperator *BO;2552    if (match(Op0, m_OneUse(m_BinOp(BO))) && isNarrowableBinOpcode(BO)) {2553      Instruction::BinaryOps BOpcode = BO->getOpcode();2554      Value *X;2555      const APInt *C1;2556      // TODO: The one-use restrictions could be relaxed a little if the AND2557      // is going to be removed.2558      // Try to narrow the 'and' and a binop with constant operand:2559      // and (bo (zext X), C1), C --> zext (and (bo X, TruncC1), TruncC)2560      if (match(BO, m_c_BinOp(m_OneUse(m_ZExt(m_Value(X))), m_APInt(C1))) &&2561          C->isIntN(X->getType()->getScalarSizeInBits())) {2562        unsigned XWidth = X->getType()->getScalarSizeInBits();2563        Constant *TruncC1 = ConstantInt::get(X->getType(), C1->trunc(XWidth));2564        Value *BinOp = isa<ZExtInst>(BO->getOperand(0))2565                           ? Builder.CreateBinOp(BOpcode, X, TruncC1)2566                           : Builder.CreateBinOp(BOpcode, TruncC1, X);2567        Constant *TruncC = ConstantInt::get(X->getType(), C->trunc(XWidth));2568        Value *And = Builder.CreateAnd(BinOp, TruncC);2569        return new ZExtInst(And, Ty);2570      }2571 2572      // Similar to above: if the mask matches the zext input width, then the2573      // 'and' can be eliminated, so we can truncate the other variable op:2574      // and (bo (zext X), Y), C --> zext (bo X, (trunc Y))2575      if (isa<Instruction>(BO->getOperand(0)) &&2576          match(BO->getOperand(0), m_OneUse(m_ZExt(m_Value(X)))) &&2577          C->isMask(X->getType()->getScalarSizeInBits())) {2578        Y = BO->getOperand(1);2579        Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr");2580        Value *NewBO =2581            Builder.CreateBinOp(BOpcode, X, TrY, BO->getName() + ".narrow");2582        return new ZExtInst(NewBO, Ty);2583      }2584      // and (bo Y, (zext X)), C --> zext (bo (trunc Y), X)2585      if (isa<Instruction>(BO->getOperand(1)) &&2586          match(BO->getOperand(1), m_OneUse(m_ZExt(m_Value(X)))) &&2587          C->isMask(X->getType()->getScalarSizeInBits())) {2588        Y = BO->getOperand(0);2589        Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr");2590        Value *NewBO =2591            Builder.CreateBinOp(BOpcode, TrY, X, BO->getName() + ".narrow");2592        return new ZExtInst(NewBO, Ty);2593      }2594    }2595 2596    // This is intentionally placed after the narrowing transforms for2597    // efficiency (transform directly to the narrow logic op if possible).2598    // If the mask is only needed on one incoming arm, push the 'and' op up.2599    if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_Value(Y)))) ||2600        match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {2601      APInt NotAndMask(~(*C));2602      BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Op0)->getOpcode();2603      if (MaskedValueIsZero(X, NotAndMask, &I)) {2604        // Not masking anything out for the LHS, move mask to RHS.2605        // and ({x}or X, Y), C --> {x}or X, (and Y, C)2606        Value *NewRHS = Builder.CreateAnd(Y, Op1, Y->getName() + ".masked");2607        return BinaryOperator::Create(BinOp, X, NewRHS);2608      }2609      if (!isa<Constant>(Y) && MaskedValueIsZero(Y, NotAndMask, &I)) {2610        // Not masking anything out for the RHS, move mask to LHS.2611        // and ({x}or X, Y), C --> {x}or (and X, C), Y2612        Value *NewLHS = Builder.CreateAnd(X, Op1, X->getName() + ".masked");2613        return BinaryOperator::Create(BinOp, NewLHS, Y);2614      }2615    }2616 2617    // When the mask is a power-of-2 constant and op0 is a shifted-power-of-22618    // constant, test if the shift amount equals the offset bit index:2619    // (ShiftC << X) & C --> X == (log2(C) - log2(ShiftC)) ? C : 02620    // (ShiftC >> X) & C --> X == (log2(ShiftC) - log2(C)) ? C : 02621    if (C->isPowerOf2() &&2622        match(Op0, m_OneUse(m_LogicalShift(m_Power2(ShiftC), m_Value(X))))) {2623      int Log2ShiftC = ShiftC->exactLogBase2();2624      int Log2C = C->exactLogBase2();2625      bool IsShiftLeft =2626         cast<BinaryOperator>(Op0)->getOpcode() == Instruction::Shl;2627      int BitNum = IsShiftLeft ? Log2C - Log2ShiftC : Log2ShiftC - Log2C;2628      assert(BitNum >= 0 && "Expected demanded bits to handle impossible mask");2629      Value *Cmp = Builder.CreateICmpEQ(X, ConstantInt::get(Ty, BitNum));2630      return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C),2631                                ConstantInt::getNullValue(Ty));2632    }2633 2634    Constant *C1, *C2;2635    const APInt *C3 = C;2636    Value *X;2637    if (C3->isPowerOf2()) {2638      Constant *Log2C3 = ConstantInt::get(Ty, C3->countr_zero());2639      if (match(Op0, m_OneUse(m_LShr(m_Shl(m_ImmConstant(C1), m_Value(X)),2640                                     m_ImmConstant(C2)))) &&2641          match(C1, m_Power2())) {2642        Constant *Log2C1 = ConstantExpr::getExactLogBase2(C1);2643        Constant *LshrC = ConstantExpr::getAdd(C2, Log2C3);2644        KnownBits KnownLShrc = computeKnownBits(LshrC, nullptr);2645        if (KnownLShrc.getMaxValue().ult(Width)) {2646          // iff C1,C3 is pow2 and C2 + cttz(C3) < BitWidth:2647          // ((C1 << X) >> C2) & C3 -> X == (cttz(C3)+C2-cttz(C1)) ? C3 : 02648          Constant *CmpC = ConstantExpr::getSub(LshrC, Log2C1);2649          Value *Cmp = Builder.CreateICmpEQ(X, CmpC);2650          return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C3),2651                                    ConstantInt::getNullValue(Ty));2652        }2653      }2654 2655      if (match(Op0, m_OneUse(m_Shl(m_LShr(m_ImmConstant(C1), m_Value(X)),2656                                    m_ImmConstant(C2)))) &&2657          match(C1, m_Power2())) {2658        Constant *Log2C1 = ConstantExpr::getExactLogBase2(C1);2659        Constant *Cmp =2660            ConstantFoldCompareInstOperands(ICmpInst::ICMP_ULT, Log2C3, C2, DL);2661        if (Cmp && Cmp->isZeroValue()) {2662          // iff C1,C3 is pow2 and Log2(C3) >= C2:2663          // ((C1 >> X) << C2) & C3 -> X == (cttz(C1)+C2-cttz(C3)) ? C3 : 02664          Constant *ShlC = ConstantExpr::getAdd(C2, Log2C1);2665          Constant *CmpC = ConstantExpr::getSub(ShlC, Log2C3);2666          Value *Cmp = Builder.CreateICmpEQ(X, CmpC);2667          return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C3),2668                                    ConstantInt::getNullValue(Ty));2669        }2670      }2671    }2672  }2673 2674  // If we are clearing the sign bit of a floating-point value, convert this to2675  // fabs, then cast back to integer.2676  //2677  // This is a generous interpretation for noimplicitfloat, this is not a true2678  // floating-point operation.2679  //2680  // Assumes any IEEE-represented type has the sign bit in the high bit.2681  // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt2682  Value *CastOp;2683  if (match(Op0, m_ElementWiseBitCast(m_Value(CastOp))) &&2684      match(Op1, m_MaxSignedValue()) &&2685      !Builder.GetInsertBlock()->getParent()->hasFnAttribute(2686          Attribute::NoImplicitFloat)) {2687    Type *EltTy = CastOp->getType()->getScalarType();2688    if (EltTy->isFloatingPointTy() &&2689        APFloat::hasSignBitInMSB(EltTy->getFltSemantics())) {2690      Value *FAbs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, CastOp);2691      return new BitCastInst(FAbs, I.getType());2692    }2693  }2694 2695  // and(shl(zext(X), Y), SignMask) -> and(sext(X), SignMask)2696  // where Y is a valid shift amount.2697  if (match(&I, m_And(m_OneUse(m_Shl(m_ZExt(m_Value(X)), m_Value(Y))),2698                      m_SignMask())) &&2699      match(Y, m_SpecificInt_ICMP(2700                   ICmpInst::Predicate::ICMP_EQ,2701                   APInt(Ty->getScalarSizeInBits(),2702                         Ty->getScalarSizeInBits() -2703                             X->getType()->getScalarSizeInBits())))) {2704    auto *SExt = Builder.CreateSExt(X, Ty, X->getName() + ".signext");2705    return BinaryOperator::CreateAnd(SExt, Op1);2706  }2707 2708  if (Instruction *Z = narrowMaskedBinOp(I))2709    return Z;2710 2711  if (I.getType()->isIntOrIntVectorTy(1)) {2712    if (auto *SI0 = dyn_cast<SelectInst>(Op0)) {2713      if (auto *R =2714              foldAndOrOfSelectUsingImpliedCond(Op1, *SI0, /* IsAnd */ true))2715        return R;2716    }2717    if (auto *SI1 = dyn_cast<SelectInst>(Op1)) {2718      if (auto *R =2719              foldAndOrOfSelectUsingImpliedCond(Op0, *SI1, /* IsAnd */ true))2720        return R;2721    }2722  }2723 2724  if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))2725    return FoldedLogic;2726 2727  if (Instruction *DeMorgan = matchDeMorgansLaws(I, *this))2728    return DeMorgan;2729 2730  {2731    Value *A, *B, *C;2732    // A & ~(A ^ B) --> A & B2733    if (match(Op1, m_Not(m_c_Xor(m_Specific(Op0), m_Value(B)))))2734      return BinaryOperator::CreateAnd(Op0, B);2735    // ~(A ^ B) & A --> A & B2736    if (match(Op0, m_Not(m_c_Xor(m_Specific(Op1), m_Value(B)))))2737      return BinaryOperator::CreateAnd(Op1, B);2738 2739    // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C2740    if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&2741        match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A)))) {2742      Value *NotC = Op1->hasOneUse()2743                        ? Builder.CreateNot(C)2744                        : getFreelyInverted(C, C->hasOneUse(), &Builder);2745      if (NotC != nullptr)2746        return BinaryOperator::CreateAnd(Op0, NotC);2747    }2748 2749    // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C2750    if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))) &&2751        match(Op1, m_Xor(m_Specific(B), m_Specific(A)))) {2752      Value *NotC = Op0->hasOneUse()2753                        ? Builder.CreateNot(C)2754                        : getFreelyInverted(C, C->hasOneUse(), &Builder);2755      if (NotC != nullptr)2756        return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(C));2757    }2758 2759    // (A | B) & (~A ^ B) -> A & B2760    // (A | B) & (B ^ ~A) -> A & B2761    // (B | A) & (~A ^ B) -> A & B2762    // (B | A) & (B ^ ~A) -> A & B2763    if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&2764        match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))2765      return BinaryOperator::CreateAnd(A, B);2766 2767    // (~A ^ B) & (A | B) -> A & B2768    // (~A ^ B) & (B | A) -> A & B2769    // (B ^ ~A) & (A | B) -> A & B2770    // (B ^ ~A) & (B | A) -> A & B2771    if (match(Op0, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&2772        match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))2773      return BinaryOperator::CreateAnd(A, B);2774 2775    // (~A | B) & (A ^ B) -> ~A & B2776    // (~A | B) & (B ^ A) -> ~A & B2777    // (B | ~A) & (A ^ B) -> ~A & B2778    // (B | ~A) & (B ^ A) -> ~A & B2779    if (match(Op0, m_c_Or(m_Not(m_Value(A)), m_Value(B))) &&2780        match(Op1, m_c_Xor(m_Specific(A), m_Specific(B))))2781      return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);2782 2783    // (A ^ B) & (~A | B) -> ~A & B2784    // (B ^ A) & (~A | B) -> ~A & B2785    // (A ^ B) & (B | ~A) -> ~A & B2786    // (B ^ A) & (B | ~A) -> ~A & B2787    if (match(Op1, m_c_Or(m_Not(m_Value(A)), m_Value(B))) &&2788        match(Op0, m_c_Xor(m_Specific(A), m_Specific(B))))2789      return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);2790  }2791 2792  if (Value *Res =2793          foldBooleanAndOr(Op0, Op1, I, /*IsAnd=*/true, /*IsLogical=*/false))2794    return replaceInstUsesWith(I, Res);2795 2796  if (match(Op1, m_OneUse(m_LogicalAnd(m_Value(X), m_Value(Y))))) {2797    bool IsLogical = isa<SelectInst>(Op1);2798    if (auto *V = reassociateBooleanAndOr(Op0, X, Y, I, /*IsAnd=*/true,2799                                          /*RHSIsLogical=*/IsLogical))2800      return replaceInstUsesWith(I, V);2801  }2802  if (match(Op0, m_OneUse(m_LogicalAnd(m_Value(X), m_Value(Y))))) {2803    bool IsLogical = isa<SelectInst>(Op0);2804    if (auto *V = reassociateBooleanAndOr(Op1, X, Y, I, /*IsAnd=*/true,2805                                          /*RHSIsLogical=*/IsLogical))2806      return replaceInstUsesWith(I, V);2807  }2808 2809  if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))2810    return FoldedFCmps;2811 2812  if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))2813    return CastedAnd;2814 2815  if (Instruction *Sel = foldBinopOfSextBoolToSelect(I))2816    return Sel;2817 2818  // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>.2819  // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold2820  //       with binop identity constant. But creating a select with non-constant2821  //       arm may not be reversible due to poison semantics. Is that a good2822  //       canonicalization?2823  Value *A, *B;2824  if (match(&I, m_c_And(m_SExt(m_Value(A)), m_Value(B))) &&2825      A->getType()->isIntOrIntVectorTy(1))2826    return SelectInst::Create(A, B, Constant::getNullValue(Ty));2827 2828  // Similarly, a 'not' of the bool translates to a swap of the select arms:2829  // ~sext(A) & B / B & ~sext(A) --> A ? 0 : B2830  if (match(&I, m_c_And(m_Not(m_SExt(m_Value(A))), m_Value(B))) &&2831      A->getType()->isIntOrIntVectorTy(1))2832    return SelectInst::Create(A, Constant::getNullValue(Ty), B);2833 2834  // and(zext(A), B) -> A ? (B & 1) : 02835  if (match(&I, m_c_And(m_OneUse(m_ZExt(m_Value(A))), m_Value(B))) &&2836      A->getType()->isIntOrIntVectorTy(1))2837    return SelectInst::Create(A, Builder.CreateAnd(B, ConstantInt::get(Ty, 1)),2838                              Constant::getNullValue(Ty));2839 2840  // (-1 + A) & B --> A ? 0 : B where A is 0/1.2841  if (match(&I, m_c_And(m_OneUse(m_Add(m_ZExtOrSelf(m_Value(A)), m_AllOnes())),2842                        m_Value(B)))) {2843    if (A->getType()->isIntOrIntVectorTy(1))2844      return SelectInst::Create(A, Constant::getNullValue(Ty), B);2845    if (computeKnownBits(A, &I).countMaxActiveBits() <= 1) {2846      return SelectInst::Create(2847          Builder.CreateICmpEQ(A, Constant::getNullValue(A->getType())), B,2848          Constant::getNullValue(Ty));2849    }2850  }2851 2852  // (iN X s>> (N-1)) & Y --> (X s< 0) ? Y : 0 -- with optional sext2853  if (match(&I, m_c_And(m_OneUse(m_SExtOrSelf(2854                            m_AShr(m_Value(X), m_APIntAllowPoison(C)))),2855                        m_Value(Y))) &&2856      *C == X->getType()->getScalarSizeInBits() - 1) {2857    Value *IsNeg = Builder.CreateIsNeg(X, "isneg");2858    return SelectInst::Create(IsNeg, Y, ConstantInt::getNullValue(Ty));2859  }2860  // If there's a 'not' of the shifted value, swap the select operands:2861  // ~(iN X s>> (N-1)) & Y --> (X s< 0) ? 0 : Y -- with optional sext2862  if (match(&I, m_c_And(m_OneUse(m_SExtOrSelf(2863                            m_Not(m_AShr(m_Value(X), m_APIntAllowPoison(C))))),2864                        m_Value(Y))) &&2865      *C == X->getType()->getScalarSizeInBits() - 1) {2866    Value *IsNeg = Builder.CreateIsNeg(X, "isneg");2867    return SelectInst::Create(IsNeg, ConstantInt::getNullValue(Ty), Y);2868  }2869 2870  // (~x) & y  -->  ~(x | (~y))  iff that gets rid of inversions2871  if (sinkNotIntoOtherHandOfLogicalOp(I))2872    return &I;2873 2874  // An and recurrence w/loop invariant step is equivelent to (and start, step)2875  PHINode *PN = nullptr;2876  Value *Start = nullptr, *Step = nullptr;2877  if (matchSimpleRecurrence(&I, PN, Start, Step) && DT.dominates(Step, PN))2878    return replaceInstUsesWith(I, Builder.CreateAnd(Start, Step));2879 2880  if (Instruction *R = reassociateForUses(I, Builder))2881    return R;2882 2883  if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))2884    return Canonicalized;2885 2886  if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1))2887    return Folded;2888 2889  if (Instruction *Res = foldBinOpOfDisplacedShifts(I))2890    return Res;2891 2892  if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))2893    return Res;2894 2895  if (Value *V =2896          simplifyAndOrWithOpReplaced(Op0, Op1, Constant::getAllOnesValue(Ty),2897                                      /*SimplifyOnly*/ false, *this))2898    return BinaryOperator::CreateAnd(V, Op1);2899  if (Value *V =2900          simplifyAndOrWithOpReplaced(Op1, Op0, Constant::getAllOnesValue(Ty),2901                                      /*SimplifyOnly*/ false, *this))2902    return BinaryOperator::CreateAnd(Op0, V);2903 2904  return nullptr;2905}2906 2907Instruction *InstCombinerImpl::matchBSwapOrBitReverse(Instruction &I,2908                                                      bool MatchBSwaps,2909                                                      bool MatchBitReversals) {2910  SmallVector<Instruction *, 4> Insts;2911  if (!recognizeBSwapOrBitReverseIdiom(&I, MatchBSwaps, MatchBitReversals,2912                                       Insts))2913    return nullptr;2914  Instruction *LastInst = Insts.pop_back_val();2915  LastInst->removeFromParent();2916 2917  for (auto *Inst : Insts) {2918    Inst->setDebugLoc(I.getDebugLoc());2919    Worklist.push(Inst);2920  }2921  return LastInst;2922}2923 2924std::optional<std::pair<Intrinsic::ID, SmallVector<Value *, 3>>>2925InstCombinerImpl::convertOrOfShiftsToFunnelShift(Instruction &Or) {2926  // TODO: Can we reduce the code duplication between this and the related2927  // rotate matching code under visitSelect and visitTrunc?2928  assert(Or.getOpcode() == BinaryOperator::Or && "Expecting or instruction");2929 2930  unsigned Width = Or.getType()->getScalarSizeInBits();2931 2932  Instruction *Or0, *Or1;2933  if (!match(Or.getOperand(0), m_Instruction(Or0)) ||2934      !match(Or.getOperand(1), m_Instruction(Or1)))2935    return std::nullopt;2936 2937  bool IsFshl = true; // Sub on LSHR.2938  SmallVector<Value *, 3> FShiftArgs;2939 2940  // First, find an or'd pair of opposite shifts:2941  // or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1)2942  if (isa<BinaryOperator>(Or0) && isa<BinaryOperator>(Or1)) {2943    Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1;2944    if (!match(Or0,2945               m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) ||2946        !match(Or1,2947               m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) ||2948        Or0->getOpcode() == Or1->getOpcode())2949      return std::nullopt;2950 2951    // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)).2952    if (Or0->getOpcode() == BinaryOperator::LShr) {2953      std::swap(Or0, Or1);2954      std::swap(ShVal0, ShVal1);2955      std::swap(ShAmt0, ShAmt1);2956    }2957    assert(Or0->getOpcode() == BinaryOperator::Shl &&2958           Or1->getOpcode() == BinaryOperator::LShr &&2959           "Illegal or(shift,shift) pair");2960 2961    // Match the shift amount operands for a funnel shift pattern. This always2962    // matches a subtraction on the R operand.2963    auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * {2964      // Check for constant shift amounts that sum to the bitwidth.2965      const APInt *LI, *RI;2966      if (match(L, m_APIntAllowPoison(LI)) && match(R, m_APIntAllowPoison(RI)))2967        if (LI->ult(Width) && RI->ult(Width) && (*LI + *RI) == Width)2968          return ConstantInt::get(L->getType(), *LI);2969 2970      Constant *LC, *RC;2971      if (match(L, m_Constant(LC)) && match(R, m_Constant(RC)) &&2972          match(L,2973                m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) &&2974          match(R,2975                m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) &&2976          match(ConstantExpr::getAdd(LC, RC), m_SpecificIntAllowPoison(Width)))2977        return ConstantExpr::mergeUndefsWith(LC, RC);2978 2979      // (shl ShVal, X) | (lshr ShVal, (Width - x)) iff X < Width.2980      // We limit this to X < Width in case the backend re-expands the2981      // intrinsic, and has to reintroduce a shift modulo operation (InstCombine2982      // might remove it after this fold). This still doesn't guarantee that the2983      // final codegen will match this original pattern.2984      if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L))))) {2985        KnownBits KnownL = computeKnownBits(L, &Or);2986        return KnownL.getMaxValue().ult(Width) ? L : nullptr;2987      }2988 2989      // For non-constant cases, the following patterns currently only work for2990      // rotation patterns.2991      // TODO: Add general funnel-shift compatible patterns.2992      if (ShVal0 != ShVal1)2993        return nullptr;2994 2995      // For non-constant cases we don't support non-pow2 shift masks.2996      // TODO: Is it worth matching urem as well?2997      if (!isPowerOf2_32(Width))2998        return nullptr;2999 3000      // The shift amount may be masked with negation:3001      // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1)))3002      Value *X;3003      unsigned Mask = Width - 1;3004      if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&3005          match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))3006        return X;3007 3008      // (shl ShVal, X) | (lshr ShVal, ((-X) & (Width - 1)))3009      if (match(R, m_And(m_Neg(m_Specific(L)), m_SpecificInt(Mask))))3010        return L;3011 3012      // Similar to above, but the shift amount may be extended after masking,3013      // so return the extended value as the parameter for the intrinsic.3014      if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&3015          match(R,3016                m_And(m_Neg(m_ZExt(m_And(m_Specific(X), m_SpecificInt(Mask)))),3017                      m_SpecificInt(Mask))))3018        return L;3019 3020      if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&3021          match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))))3022        return L;3023 3024      return nullptr;3025    };3026 3027    Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width);3028    if (!ShAmt) {3029      ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width);3030      IsFshl = false; // Sub on SHL.3031    }3032    if (!ShAmt)3033      return std::nullopt;3034 3035    FShiftArgs = {ShVal0, ShVal1, ShAmt};3036  } else if (isa<ZExtInst>(Or0) || isa<ZExtInst>(Or1)) {3037    // If there are two 'or' instructions concat variables in opposite order:3038    //3039    // Slot1 and Slot2 are all zero bits.3040    // | Slot1 | Low | Slot2 | High |3041    // LowHigh = or (shl (zext Low), ZextLowShlAmt), (zext High)3042    // | Slot2 | High | Slot1 | Low |3043    // HighLow = or (shl (zext High), ZextHighShlAmt), (zext Low)3044    //3045    // the latter 'or' can be safely convert to3046    // -> HighLow = fshl LowHigh, LowHigh, ZextHighShlAmt3047    // if ZextLowShlAmt + ZextHighShlAmt == Width.3048    if (!isa<ZExtInst>(Or1))3049      std::swap(Or0, Or1);3050 3051    Value *High, *ZextHigh, *Low;3052    const APInt *ZextHighShlAmt;3053    if (!match(Or0,3054               m_OneUse(m_Shl(m_Value(ZextHigh), m_APInt(ZextHighShlAmt)))))3055      return std::nullopt;3056 3057    if (!match(Or1, m_ZExt(m_Value(Low))) ||3058        !match(ZextHigh, m_ZExt(m_Value(High))))3059      return std::nullopt;3060 3061    unsigned HighSize = High->getType()->getScalarSizeInBits();3062    unsigned LowSize = Low->getType()->getScalarSizeInBits();3063    // Make sure High does not overlap with Low and most significant bits of3064    // High aren't shifted out.3065    if (ZextHighShlAmt->ult(LowSize) || ZextHighShlAmt->ugt(Width - HighSize))3066      return std::nullopt;3067 3068    for (User *U : ZextHigh->users()) {3069      Value *X, *Y;3070      if (!match(U, m_Or(m_Value(X), m_Value(Y))))3071        continue;3072 3073      if (!isa<ZExtInst>(Y))3074        std::swap(X, Y);3075 3076      const APInt *ZextLowShlAmt;3077      if (!match(X, m_Shl(m_Specific(Or1), m_APInt(ZextLowShlAmt))) ||3078          !match(Y, m_Specific(ZextHigh)) || !DT.dominates(U, &Or))3079        continue;3080 3081      // HighLow is good concat. If sum of two shifts amount equals to Width,3082      // LowHigh must also be a good concat.3083      if (*ZextLowShlAmt + *ZextHighShlAmt != Width)3084        continue;3085 3086      // Low must not overlap with High and most significant bits of Low must3087      // not be shifted out.3088      assert(ZextLowShlAmt->uge(HighSize) &&3089             ZextLowShlAmt->ule(Width - LowSize) && "Invalid concat");3090 3091      // We cannot reuse the result if it may produce poison.3092      // Drop poison generating flags in the expression tree.3093      // Or3094      cast<Instruction>(U)->dropPoisonGeneratingFlags();3095      // Shl3096      cast<Instruction>(X)->dropPoisonGeneratingFlags();3097 3098      FShiftArgs = {U, U, ConstantInt::get(Or0->getType(), *ZextHighShlAmt)};3099      break;3100    }3101  }3102 3103  if (FShiftArgs.empty())3104    return std::nullopt;3105 3106  Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;3107  return std::make_pair(IID, FShiftArgs);3108}3109 3110/// Match UB-safe variants of the funnel shift intrinsic.3111static Instruction *matchFunnelShift(Instruction &Or, InstCombinerImpl &IC) {3112  if (auto Opt = IC.convertOrOfShiftsToFunnelShift(Or)) {3113    auto [IID, FShiftArgs] = *Opt;3114    Function *F =3115        Intrinsic::getOrInsertDeclaration(Or.getModule(), IID, Or.getType());3116    return CallInst::Create(F, FShiftArgs);3117  }3118 3119  return nullptr;3120}3121 3122/// Attempt to combine or(zext(x),shl(zext(y),bw/2) concat packing patterns.3123static Value *matchOrConcat(Instruction &Or, InstCombiner::BuilderTy &Builder) {3124  assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'");3125  Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1);3126  Type *Ty = Or.getType();3127 3128  unsigned Width = Ty->getScalarSizeInBits();3129  if ((Width & 1) != 0)3130    return nullptr;3131  unsigned HalfWidth = Width / 2;3132 3133  // Canonicalize zext (lower half) to LHS.3134  if (!isa<ZExtInst>(Op0))3135    std::swap(Op0, Op1);3136 3137  // Find lower/upper half.3138  Value *LowerSrc, *ShlVal, *UpperSrc;3139  const APInt *C;3140  if (!match(Op0, m_OneUse(m_ZExt(m_Value(LowerSrc)))) ||3141      !match(Op1, m_OneUse(m_Shl(m_Value(ShlVal), m_APInt(C)))) ||3142      !match(ShlVal, m_OneUse(m_ZExt(m_Value(UpperSrc)))))3143    return nullptr;3144  if (*C != HalfWidth || LowerSrc->getType() != UpperSrc->getType() ||3145      LowerSrc->getType()->getScalarSizeInBits() != HalfWidth)3146    return nullptr;3147 3148  auto ConcatIntrinsicCalls = [&](Intrinsic::ID id, Value *Lo, Value *Hi) {3149    Value *NewLower = Builder.CreateZExt(Lo, Ty);3150    Value *NewUpper = Builder.CreateZExt(Hi, Ty);3151    NewUpper = Builder.CreateShl(NewUpper, HalfWidth);3152    Value *BinOp = Builder.CreateOr(NewLower, NewUpper);3153    return Builder.CreateIntrinsic(id, Ty, BinOp);3154  };3155 3156  // BSWAP: Push the concat down, swapping the lower/upper sources.3157  // concat(bswap(x),bswap(y)) -> bswap(concat(x,y))3158  Value *LowerBSwap, *UpperBSwap;3159  if (match(LowerSrc, m_BSwap(m_Value(LowerBSwap))) &&3160      match(UpperSrc, m_BSwap(m_Value(UpperBSwap))))3161    return ConcatIntrinsicCalls(Intrinsic::bswap, UpperBSwap, LowerBSwap);3162 3163  // BITREVERSE: Push the concat down, swapping the lower/upper sources.3164  // concat(bitreverse(x),bitreverse(y)) -> bitreverse(concat(x,y))3165  Value *LowerBRev, *UpperBRev;3166  if (match(LowerSrc, m_BitReverse(m_Value(LowerBRev))) &&3167      match(UpperSrc, m_BitReverse(m_Value(UpperBRev))))3168    return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev);3169 3170  // iX ext split: extending or(zext(x),shl(zext(y),bw/2) pattern3171  // to consume sext/ashr:3172  // or(zext(sext(x)),shl(zext(sext(ashr(x,xbw-1))),bw/2)3173  // or(zext(x),shl(zext(ashr(x,xbw-1)),bw/2)3174  Value *X;3175  if (match(LowerSrc, m_SExtOrSelf(m_Value(X))) &&3176      match(UpperSrc,3177            m_SExtOrSelf(m_AShr(3178                m_Specific(X),3179                m_SpecificInt(X->getType()->getScalarSizeInBits() - 1)))))3180    return Builder.CreateSExt(X, Ty);3181 3182  return nullptr;3183}3184 3185/// If all elements of two constant vectors are 0/-1 and inverses, return true.3186static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {3187  unsigned NumElts = cast<FixedVectorType>(C1->getType())->getNumElements();3188  for (unsigned i = 0; i != NumElts; ++i) {3189    Constant *EltC1 = C1->getAggregateElement(i);3190    Constant *EltC2 = C2->getAggregateElement(i);3191    if (!EltC1 || !EltC2)3192      return false;3193 3194    // One element must be all ones, and the other must be all zeros.3195    if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) ||3196          (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes()))))3197      return false;3198  }3199  return true;3200}3201 3202/// We have an expression of the form (A & C) | (B & D). If A is a scalar or3203/// vector composed of all-zeros or all-ones values and is the bitwise 'not' of3204/// B, it can be used as the condition operand of a select instruction.3205/// We will detect (A & C) | ~(B | D) when the flag ABIsTheSame enabled.3206Value *InstCombinerImpl::getSelectCondition(Value *A, Value *B,3207                                            bool ABIsTheSame) {3208  // We may have peeked through bitcasts in the caller.3209  // Exit immediately if we don't have (vector) integer types.3210  Type *Ty = A->getType();3211  if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy())3212    return nullptr;3213 3214  // If A is the 'not' operand of B and has enough signbits, we have our answer.3215  if (ABIsTheSame ? (A == B) : match(B, m_Not(m_Specific(A)))) {3216    // If these are scalars or vectors of i1, A can be used directly.3217    if (Ty->isIntOrIntVectorTy(1))3218      return A;3219 3220    // If we look through a vector bitcast, the caller will bitcast the operands3221    // to match the condition's number of bits (N x i1).3222    // To make this poison-safe, disallow bitcast from wide element to narrow3223    // element. That could allow poison in lanes where it was not present in the3224    // original code.3225    A = peekThroughBitcast(A);3226    if (A->getType()->isIntOrIntVectorTy()) {3227      unsigned NumSignBits = ComputeNumSignBits(A);3228      if (NumSignBits == A->getType()->getScalarSizeInBits() &&3229          NumSignBits <= Ty->getScalarSizeInBits())3230        return Builder.CreateTrunc(A, CmpInst::makeCmpResultType(A->getType()));3231    }3232    return nullptr;3233  }3234 3235  // TODO: add support for sext and constant case3236  if (ABIsTheSame)3237    return nullptr;3238 3239  // If both operands are constants, see if the constants are inverse bitmasks.3240  Constant *AConst, *BConst;3241  if (match(A, m_Constant(AConst)) && match(B, m_Constant(BConst)))3242    if (AConst == ConstantExpr::getNot(BConst) &&3243        ComputeNumSignBits(A) == Ty->getScalarSizeInBits())3244      return Builder.CreateZExtOrTrunc(A, CmpInst::makeCmpResultType(Ty));3245 3246  // Look for more complex patterns. The 'not' op may be hidden behind various3247  // casts. Look through sexts and bitcasts to find the booleans.3248  Value *Cond;3249  Value *NotB;3250  if (match(A, m_SExt(m_Value(Cond))) &&3251      Cond->getType()->isIntOrIntVectorTy(1)) {3252    // A = sext i1 Cond; B = sext (not (i1 Cond))3253    if (match(B, m_SExt(m_Not(m_Specific(Cond)))))3254      return Cond;3255 3256    // A = sext i1 Cond; B = not ({bitcast} (sext (i1 Cond)))3257    // TODO: The one-use checks are unnecessary or misplaced. If the caller3258    //       checked for uses on logic ops/casts, that should be enough to3259    //       make this transform worthwhile.3260    if (match(B, m_OneUse(m_Not(m_Value(NotB))))) {3261      NotB = peekThroughBitcast(NotB, true);3262      if (match(NotB, m_SExt(m_Specific(Cond))))3263        return Cond;3264    }3265  }3266 3267  // All scalar (and most vector) possibilities should be handled now.3268  // Try more matches that only apply to non-splat constant vectors.3269  if (!Ty->isVectorTy())3270    return nullptr;3271 3272  // If both operands are xor'd with constants using the same sexted boolean3273  // operand, see if the constants are inverse bitmasks.3274  // TODO: Use ConstantExpr::getNot()?3275  if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AConst)))) &&3276      match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BConst)))) &&3277      Cond->getType()->isIntOrIntVectorTy(1) &&3278      areInverseVectorBitmasks(AConst, BConst)) {3279    AConst = ConstantExpr::getTrunc(AConst, CmpInst::makeCmpResultType(Ty));3280    return Builder.CreateXor(Cond, AConst);3281  }3282  return nullptr;3283}3284 3285/// We have an expression of the form (A & B) | (C & D). Try to simplify this3286/// to "A' ? B : D", where A' is a boolean or vector of booleans.3287/// When InvertFalseVal is set to true, we try to match the pattern3288/// where we have peeked through a 'not' op and A and C are the same:3289/// (A & B) | ~(A | D) --> (A & B) | (~A & ~D) --> A' ? B : ~D3290Value *InstCombinerImpl::matchSelectFromAndOr(Value *A, Value *B, Value *C,3291                                              Value *D, bool InvertFalseVal) {3292  // The potential condition of the select may be bitcasted. In that case, look3293  // through its bitcast and the corresponding bitcast of the 'not' condition.3294  Type *OrigType = A->getType();3295  A = peekThroughBitcast(A, true);3296  C = peekThroughBitcast(C, true);3297  if (Value *Cond = getSelectCondition(A, C, InvertFalseVal)) {3298    // ((bc Cond) & B) | ((bc ~Cond) & D) --> bc (select Cond, (bc B), (bc D))3299    // If this is a vector, we may need to cast to match the condition's length.3300    // The bitcasts will either all exist or all not exist. The builder will3301    // not create unnecessary casts if the types already match.3302    Type *SelTy = A->getType();3303    if (auto *VecTy = dyn_cast<VectorType>(Cond->getType())) {3304      // For a fixed or scalable vector get N from <{vscale x} N x iM>3305      unsigned Elts = VecTy->getElementCount().getKnownMinValue();3306      // For a fixed or scalable vector, get the size in bits of N x iM; for a3307      // scalar this is just M.3308      unsigned SelEltSize = SelTy->getPrimitiveSizeInBits().getKnownMinValue();3309      Type *EltTy = Builder.getIntNTy(SelEltSize / Elts);3310      SelTy = VectorType::get(EltTy, VecTy->getElementCount());3311    }3312    Value *BitcastB = Builder.CreateBitCast(B, SelTy);3313    if (InvertFalseVal)3314      D = Builder.CreateNot(D);3315    Value *BitcastD = Builder.CreateBitCast(D, SelTy);3316    Value *Select = Builder.CreateSelect(Cond, BitcastB, BitcastD);3317    return Builder.CreateBitCast(Select, OrigType);3318  }3319 3320  return nullptr;3321}3322 3323// (icmp eq X, C) | (icmp ult Other, (X - C)) -> (icmp ule Other, (X - (C + 1)))3324// (icmp ne X, C) & (icmp uge Other, (X - C)) -> (icmp ugt Other, (X - (C + 1)))3325static Value *foldAndOrOfICmpEqConstantAndICmp(ICmpInst *LHS, ICmpInst *RHS,3326                                               bool IsAnd, bool IsLogical,3327                                               IRBuilderBase &Builder) {3328  Value *LHS0 = LHS->getOperand(0);3329  Value *RHS0 = RHS->getOperand(0);3330  Value *RHS1 = RHS->getOperand(1);3331 3332  ICmpInst::Predicate LPred =3333      IsAnd ? LHS->getInversePredicate() : LHS->getPredicate();3334  ICmpInst::Predicate RPred =3335      IsAnd ? RHS->getInversePredicate() : RHS->getPredicate();3336 3337  const APInt *CInt;3338  if (LPred != ICmpInst::ICMP_EQ ||3339      !match(LHS->getOperand(1), m_APIntAllowPoison(CInt)) ||3340      !LHS0->getType()->isIntOrIntVectorTy() ||3341      !(LHS->hasOneUse() || RHS->hasOneUse()))3342    return nullptr;3343 3344  auto MatchRHSOp = [LHS0, CInt](const Value *RHSOp) {3345    return match(RHSOp,3346                 m_Add(m_Specific(LHS0), m_SpecificIntAllowPoison(-*CInt))) ||3347           (CInt->isZero() && RHSOp == LHS0);3348  };3349 3350  Value *Other;3351  if (RPred == ICmpInst::ICMP_ULT && MatchRHSOp(RHS1))3352    Other = RHS0;3353  else if (RPred == ICmpInst::ICMP_UGT && MatchRHSOp(RHS0))3354    Other = RHS1;3355  else3356    return nullptr;3357 3358  if (IsLogical)3359    Other = Builder.CreateFreeze(Other);3360 3361  return Builder.CreateICmp(3362      IsAnd ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE,3363      Builder.CreateSub(LHS0, ConstantInt::get(LHS0->getType(), *CInt + 1)),3364      Other);3365}3366 3367/// Fold (icmp)&(icmp) or (icmp)|(icmp) if possible.3368/// If IsLogical is true, then the and/or is in select form and the transform3369/// must be poison-safe.3370Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,3371                                          Instruction &I, bool IsAnd,3372                                          bool IsLogical) {3373  const SimplifyQuery Q = SQ.getWithInstruction(&I);3374 3375  ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();3376  Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);3377  Value *LHS1 = LHS->getOperand(1), *RHS1 = RHS->getOperand(1);3378 3379  const APInt *LHSC = nullptr, *RHSC = nullptr;3380  match(LHS1, m_APInt(LHSC));3381  match(RHS1, m_APInt(RHSC));3382 3383  // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)3384  // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)3385  if (predicatesFoldable(PredL, PredR)) {3386    if (LHS0 == RHS1 && LHS1 == RHS0) {3387      PredL = ICmpInst::getSwappedPredicate(PredL);3388      std::swap(LHS0, LHS1);3389    }3390    if (LHS0 == RHS0 && LHS1 == RHS1) {3391      unsigned Code = IsAnd ? getICmpCode(PredL) & getICmpCode(PredR)3392                            : getICmpCode(PredL) | getICmpCode(PredR);3393      bool IsSigned = LHS->isSigned() || RHS->isSigned();3394      return getNewICmpValue(Code, IsSigned, LHS0, LHS1, Builder);3395    }3396  }3397 3398  if (Value *V =3399          foldAndOrOfICmpEqConstantAndICmp(LHS, RHS, IsAnd, IsLogical, Builder))3400    return V;3401  // We can treat logical like bitwise here, because both operands are used on3402  // the LHS, and as such poison from both will propagate.3403  if (Value *V = foldAndOrOfICmpEqConstantAndICmp(RHS, LHS, IsAnd,3404                                                  /*IsLogical*/ false, Builder))3405    return V;3406 3407  if (Value *V = foldAndOrOfICmpsWithConstEq(LHS, RHS, IsAnd, IsLogical,3408                                             Builder, Q, I))3409    return V;3410  // We can convert this case to bitwise and, because both operands are used3411  // on the LHS, and as such poison from both will propagate.3412  if (Value *V = foldAndOrOfICmpsWithConstEq(3413          RHS, LHS, IsAnd, /*IsLogical=*/false, Builder, Q, I)) {3414    // If RHS is still used, we should drop samesign flag.3415    if (IsLogical && RHS->hasSameSign() && !RHS->use_empty()) {3416      RHS->setSameSign(false);3417      addToWorklist(RHS);3418    }3419    return V;3420  }3421 3422  if (Value *V = foldIsPowerOf2OrZero(LHS, RHS, IsAnd, Builder, *this))3423    return V;3424  if (Value *V = foldIsPowerOf2OrZero(RHS, LHS, IsAnd, Builder, *this))3425    return V;3426 3427  // TODO: One of these directions is fine with logical and/or, the other could3428  // be supported by inserting freeze.3429  if (!IsLogical) {3430    // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n3431    // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n3432    if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/!IsAnd))3433      return V;3434 3435    // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n3436    // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n3437    if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/!IsAnd))3438      return V;3439  }3440 3441  // TODO: Add conjugated or fold, check whether it is safe for logical and/or.3442  if (IsAnd && !IsLogical)3443    if (Value *V = foldSignedTruncationCheck(LHS, RHS, I, Builder))3444      return V;3445 3446  if (Value *V = foldIsPowerOf2(LHS, RHS, IsAnd, Builder, *this))3447    return V;3448 3449  if (Value *V = foldPowerOf2AndShiftedMask(LHS, RHS, IsAnd, Builder))3450    return V;3451 3452  // TODO: Verify whether this is safe for logical and/or.3453  if (!IsLogical) {3454    if (Value *X = foldUnsignedUnderflowCheck(LHS, RHS, IsAnd, Q, Builder))3455      return X;3456    if (Value *X = foldUnsignedUnderflowCheck(RHS, LHS, IsAnd, Q, Builder))3457      return X;3458  }3459 3460  // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)3461  // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)3462  // TODO: Remove this and below when foldLogOpOfMaskedICmps can handle undefs.3463  if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&3464      PredL == PredR && match(LHS1, m_ZeroInt()) && match(RHS1, m_ZeroInt()) &&3465      LHS0->getType() == RHS0->getType() &&3466      (!IsLogical || isGuaranteedNotToBePoison(RHS0))) {3467    Value *NewOr = Builder.CreateOr(LHS0, RHS0);3468    return Builder.CreateICmp(PredL, NewOr,3469                              Constant::getNullValue(NewOr->getType()));3470  }3471 3472  // (icmp ne A, -1) | (icmp ne B, -1) --> (icmp ne (A&B), -1)3473  // (icmp eq A, -1) & (icmp eq B, -1) --> (icmp eq (A&B), -1)3474  if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&3475      PredL == PredR && match(LHS1, m_AllOnes()) && match(RHS1, m_AllOnes()) &&3476      LHS0->getType() == RHS0->getType() &&3477      (!IsLogical || isGuaranteedNotToBePoison(RHS0))) {3478    Value *NewAnd = Builder.CreateAnd(LHS0, RHS0);3479    return Builder.CreateICmp(PredL, NewAnd,3480                              Constant::getAllOnesValue(LHS0->getType()));3481  }3482 3483  if (!IsLogical)3484    if (Value *V =3485            foldAndOrOfICmpsWithPow2AndWithZero(Builder, LHS, RHS, IsAnd, Q))3486      return V;3487 3488  // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).3489  if (!LHSC || !RHSC)3490    return nullptr;3491 3492  // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C23493  // (trunc x) != C1 | (and x, CA) != C2 -> (and x, CA|CMAX) != C1|C23494  // where CMAX is the all ones value for the truncated type,3495  // iff the lower bits of C2 and CA are zero.3496  if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&3497      PredL == PredR && LHS->hasOneUse() && RHS->hasOneUse()) {3498    Value *V;3499    const APInt *AndC, *SmallC = nullptr, *BigC = nullptr;3500 3501    // (trunc x) == C1 & (and x, CA) == C23502    // (and x, CA) == C2 & (trunc x) == C13503    if (match(RHS0, m_Trunc(m_Value(V))) &&3504        match(LHS0, m_And(m_Specific(V), m_APInt(AndC)))) {3505      SmallC = RHSC;3506      BigC = LHSC;3507    } else if (match(LHS0, m_Trunc(m_Value(V))) &&3508               match(RHS0, m_And(m_Specific(V), m_APInt(AndC)))) {3509      SmallC = LHSC;3510      BigC = RHSC;3511    }3512 3513    if (SmallC && BigC) {3514      unsigned BigBitSize = BigC->getBitWidth();3515      unsigned SmallBitSize = SmallC->getBitWidth();3516 3517      // Check that the low bits are zero.3518      APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);3519      if ((Low & *AndC).isZero() && (Low & *BigC).isZero()) {3520        Value *NewAnd = Builder.CreateAnd(V, Low | *AndC);3521        APInt N = SmallC->zext(BigBitSize) | *BigC;3522        Value *NewVal = ConstantInt::get(NewAnd->getType(), N);3523        return Builder.CreateICmp(PredL, NewAnd, NewVal);3524      }3525    }3526  }3527 3528  // Match naive pattern (and its inverted form) for checking if two values3529  // share same sign. An example of the pattern:3530  // (icmp slt (X & Y), 0) | (icmp sgt (X | Y), -1) -> (icmp sgt (X ^ Y), -1)3531  // Inverted form (example):3532  // (icmp slt (X | Y), 0) & (icmp sgt (X & Y), -1) -> (icmp slt (X ^ Y), 0)3533  bool TrueIfSignedL, TrueIfSignedR;3534  if (isSignBitCheck(PredL, *LHSC, TrueIfSignedL) &&3535      isSignBitCheck(PredR, *RHSC, TrueIfSignedR) &&3536      (RHS->hasOneUse() || LHS->hasOneUse())) {3537    Value *X, *Y;3538    if (IsAnd) {3539      if ((TrueIfSignedL && !TrueIfSignedR &&3540           match(LHS0, m_Or(m_Value(X), m_Value(Y))) &&3541           match(RHS0, m_c_And(m_Specific(X), m_Specific(Y)))) ||3542          (!TrueIfSignedL && TrueIfSignedR &&3543           match(LHS0, m_And(m_Value(X), m_Value(Y))) &&3544           match(RHS0, m_c_Or(m_Specific(X), m_Specific(Y))))) {3545        Value *NewXor = Builder.CreateXor(X, Y);3546        return Builder.CreateIsNeg(NewXor);3547      }3548    } else {3549      if ((TrueIfSignedL && !TrueIfSignedR &&3550            match(LHS0, m_And(m_Value(X), m_Value(Y))) &&3551            match(RHS0, m_c_Or(m_Specific(X), m_Specific(Y)))) ||3552          (!TrueIfSignedL && TrueIfSignedR &&3553           match(LHS0, m_Or(m_Value(X), m_Value(Y))) &&3554           match(RHS0, m_c_And(m_Specific(X), m_Specific(Y))))) {3555        Value *NewXor = Builder.CreateXor(X, Y);3556        return Builder.CreateIsNotNeg(NewXor);3557      }3558    }3559  }3560 3561  // (X & ExpMask) != 0 && (X & ExpMask) != ExpMask -> isnormal(X)3562  // (X & ExpMask) == 0 || (X & ExpMask) == ExpMask -> !isnormal(X)3563  Value *X;3564  const APInt *MaskC;3565  if (LHS0 == RHS0 && PredL == PredR &&3566      PredL == (IsAnd ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ) &&3567      !I.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&3568      LHS->hasOneUse() && RHS->hasOneUse() &&3569      match(LHS0, m_And(m_ElementWiseBitCast(m_Value(X)), m_APInt(MaskC))) &&3570      X->getType()->getScalarType()->isIEEELikeFPTy() &&3571      APFloat(X->getType()->getScalarType()->getFltSemantics(), *MaskC)3572          .isPosInfinity() &&3573      ((LHSC->isZero() && *RHSC == *MaskC) ||3574       (RHSC->isZero() && *LHSC == *MaskC)))3575    return Builder.createIsFPClass(X, IsAnd ? FPClassTest::fcNormal3576                                            : ~FPClassTest::fcNormal);3577 3578  return foldAndOrOfICmpsUsingRanges(LHS, RHS, IsAnd);3579}3580 3581/// If IsLogical is true, then the and/or is in select form and the transform3582/// must be poison-safe.3583Value *InstCombinerImpl::foldBooleanAndOr(Value *LHS, Value *RHS,3584                                          Instruction &I, bool IsAnd,3585                                          bool IsLogical) {3586  if (!LHS->getType()->isIntOrIntVectorTy(1))3587    return nullptr;3588 3589  // handle (roughly):3590  // (icmp ne (A & B), C) | (icmp ne (A & D), E)3591  // (icmp eq (A & B), C) & (icmp eq (A & D), E)3592  if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, IsAnd, IsLogical, Builder,3593                                        SQ.getWithInstruction(&I)))3594    return V;3595 3596  if (auto *LHSCmp = dyn_cast<ICmpInst>(LHS))3597    if (auto *RHSCmp = dyn_cast<ICmpInst>(RHS))3598      if (Value *Res = foldAndOrOfICmps(LHSCmp, RHSCmp, I, IsAnd, IsLogical))3599        return Res;3600 3601  if (auto *LHSCmp = dyn_cast<FCmpInst>(LHS))3602    if (auto *RHSCmp = dyn_cast<FCmpInst>(RHS))3603      if (Value *Res = foldLogicOfFCmps(LHSCmp, RHSCmp, IsAnd, IsLogical))3604        return Res;3605 3606  if (Value *Res = foldEqOfParts(LHS, RHS, IsAnd))3607    return Res;3608 3609  return nullptr;3610}3611 3612static Value *foldOrOfInversions(BinaryOperator &I,3613                                 InstCombiner::BuilderTy &Builder) {3614  assert(I.getOpcode() == Instruction::Or &&3615         "Simplification only supports or at the moment.");3616 3617  Value *Cmp1, *Cmp2, *Cmp3, *Cmp4;3618  if (!match(I.getOperand(0), m_And(m_Value(Cmp1), m_Value(Cmp2))) ||3619      !match(I.getOperand(1), m_And(m_Value(Cmp3), m_Value(Cmp4))))3620    return nullptr;3621 3622  // Check if any two pairs of the and operations are inversions of each other.3623  if (isKnownInversion(Cmp1, Cmp3) && isKnownInversion(Cmp2, Cmp4))3624    return Builder.CreateXor(Cmp1, Cmp4);3625  if (isKnownInversion(Cmp1, Cmp4) && isKnownInversion(Cmp2, Cmp3))3626    return Builder.CreateXor(Cmp1, Cmp3);3627 3628  return nullptr;3629}3630 3631/// Match \p V as "shufflevector -> bitcast" or "extractelement -> zext -> shl"3632/// patterns, which extract vector elements and pack them in the same relative3633/// positions.3634///3635/// \p Vec is the underlying vector being extracted from.3636/// \p Mask is a bitmask identifying which packed elements are obtained from the3637/// vector.3638/// \p VecOffset is the vector element corresponding to index 0 of the3639/// mask.3640static bool matchSubIntegerPackFromVector(Value *V, Value *&Vec,3641                                          int64_t &VecOffset,3642                                          SmallBitVector &Mask,3643                                          const DataLayout &DL) {3644  // First try to match extractelement -> zext -> shl3645  uint64_t VecIdx, ShlAmt;3646  if (match(V, m_ShlOrSelf(m_ZExtOrSelf(m_ExtractElt(m_Value(Vec),3647                                                     m_ConstantInt(VecIdx))),3648                           ShlAmt))) {3649    auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType());3650    if (!VecTy)3651      return false;3652    auto *EltTy = dyn_cast<IntegerType>(VecTy->getElementType());3653    if (!EltTy)3654      return false;3655 3656    const unsigned EltBitWidth = EltTy->getBitWidth();3657    const unsigned TargetBitWidth = V->getType()->getIntegerBitWidth();3658    if (TargetBitWidth % EltBitWidth != 0 || ShlAmt % EltBitWidth != 0)3659      return false;3660    const unsigned TargetEltWidth = TargetBitWidth / EltBitWidth;3661    const unsigned ShlEltAmt = ShlAmt / EltBitWidth;3662 3663    const unsigned MaskIdx =3664        DL.isLittleEndian() ? ShlEltAmt : TargetEltWidth - ShlEltAmt - 1;3665 3666    VecOffset = static_cast<int64_t>(VecIdx) - static_cast<int64_t>(MaskIdx);3667    Mask.resize(TargetEltWidth);3668    Mask.set(MaskIdx);3669    return true;3670  }3671 3672  // Now try to match a bitcasted subvector.3673  Instruction *SrcVecI;3674  if (!match(V, m_BitCast(m_Instruction(SrcVecI))))3675    return false;3676 3677  auto *SrcTy = dyn_cast<FixedVectorType>(SrcVecI->getType());3678  if (!SrcTy)3679    return false;3680 3681  Mask.resize(SrcTy->getNumElements());3682 3683  // First check for a subvector obtained from a shufflevector.3684  if (isa<ShuffleVectorInst>(SrcVecI)) {3685    Constant *ConstVec;3686    ArrayRef<int> ShuffleMask;3687    if (!match(SrcVecI, m_Shuffle(m_Value(Vec), m_Constant(ConstVec),3688                                  m_Mask(ShuffleMask))))3689      return false;3690 3691    auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType());3692    if (!VecTy)3693      return false;3694 3695    const unsigned NumVecElts = VecTy->getNumElements();3696    bool FoundVecOffset = false;3697    for (unsigned Idx = 0; Idx < ShuffleMask.size(); ++Idx) {3698      if (ShuffleMask[Idx] == PoisonMaskElem)3699        return false;3700      const unsigned ShuffleIdx = ShuffleMask[Idx];3701      if (ShuffleIdx >= NumVecElts) {3702        const unsigned ConstIdx = ShuffleIdx - NumVecElts;3703        auto *ConstElt =3704            dyn_cast<ConstantInt>(ConstVec->getAggregateElement(ConstIdx));3705        if (!ConstElt || !ConstElt->isNullValue())3706          return false;3707        continue;3708      }3709 3710      if (FoundVecOffset) {3711        if (VecOffset + Idx != ShuffleIdx)3712          return false;3713      } else {3714        if (ShuffleIdx < Idx)3715          return false;3716        VecOffset = ShuffleIdx - Idx;3717        FoundVecOffset = true;3718      }3719      Mask.set(Idx);3720    }3721    return FoundVecOffset;3722  }3723 3724  // Check for a subvector obtained as an (insertelement V, 0, idx)3725  uint64_t InsertIdx;3726  if (!match(SrcVecI,3727             m_InsertElt(m_Value(Vec), m_Zero(), m_ConstantInt(InsertIdx))))3728    return false;3729 3730  auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType());3731  if (!VecTy)3732    return false;3733  VecOffset = 0;3734  bool AlreadyInsertedMaskedElt = Mask.test(InsertIdx);3735  Mask.set();3736  if (!AlreadyInsertedMaskedElt)3737    Mask.reset(InsertIdx);3738  return true;3739}3740 3741/// Try to fold the join of two scalar integers whose contents are packed3742/// elements of the same vector.3743static Instruction *foldIntegerPackFromVector(Instruction &I,3744                                              InstCombiner::BuilderTy &Builder,3745                                              const DataLayout &DL) {3746  assert(I.getOpcode() == Instruction::Or);3747  Value *LhsVec, *RhsVec;3748  int64_t LhsVecOffset, RhsVecOffset;3749  SmallBitVector Mask;3750  if (!matchSubIntegerPackFromVector(I.getOperand(0), LhsVec, LhsVecOffset,3751                                     Mask, DL))3752    return nullptr;3753  if (!matchSubIntegerPackFromVector(I.getOperand(1), RhsVec, RhsVecOffset,3754                                     Mask, DL))3755    return nullptr;3756  if (LhsVec != RhsVec || LhsVecOffset != RhsVecOffset)3757    return nullptr;3758 3759  // Convert into shufflevector -> bitcast;3760  const unsigned ZeroVecIdx =3761      cast<FixedVectorType>(LhsVec->getType())->getNumElements();3762  SmallVector<int> ShuffleMask(Mask.size(), ZeroVecIdx);3763  for (unsigned Idx : Mask.set_bits()) {3764    assert(LhsVecOffset + Idx >= 0);3765    ShuffleMask[Idx] = LhsVecOffset + Idx;3766  }3767 3768  Value *MaskedVec = Builder.CreateShuffleVector(3769      LhsVec, Constant::getNullValue(LhsVec->getType()), ShuffleMask,3770      I.getName() + ".v");3771  return CastInst::Create(Instruction::BitCast, MaskedVec, I.getType());3772}3773 3774/// Match \p V as "lshr -> mask -> zext -> shl".3775///3776/// \p Int is the underlying integer being extracted from.3777/// \p Mask is a bitmask identifying which bits of the integer are being3778/// extracted. \p Offset identifies which bit of the result \p V corresponds to3779/// the least significant bit of \p Int3780static bool matchZExtedSubInteger(Value *V, Value *&Int, APInt &Mask,3781                                  uint64_t &Offset, bool &IsShlNUW,3782                                  bool &IsShlNSW) {3783  Value *ShlOp0;3784  uint64_t ShlAmt = 0;3785  if (!match(V, m_OneUse(m_Shl(m_Value(ShlOp0), m_ConstantInt(ShlAmt)))))3786    return false;3787 3788  IsShlNUW = cast<BinaryOperator>(V)->hasNoUnsignedWrap();3789  IsShlNSW = cast<BinaryOperator>(V)->hasNoSignedWrap();3790 3791  Value *ZExtOp0;3792  if (!match(ShlOp0, m_OneUse(m_ZExt(m_Value(ZExtOp0)))))3793    return false;3794 3795  Value *MaskedOp0;3796  const APInt *ShiftedMaskConst = nullptr;3797  if (!match(ZExtOp0, m_CombineOr(m_OneUse(m_And(m_Value(MaskedOp0),3798                                                 m_APInt(ShiftedMaskConst))),3799                                  m_Value(MaskedOp0))))3800    return false;3801 3802  uint64_t LShrAmt = 0;3803  if (!match(MaskedOp0,3804             m_CombineOr(m_OneUse(m_LShr(m_Value(Int), m_ConstantInt(LShrAmt))),3805                         m_Value(Int))))3806    return false;3807 3808  if (LShrAmt > ShlAmt)3809    return false;3810  Offset = ShlAmt - LShrAmt;3811 3812  Mask = ShiftedMaskConst ? ShiftedMaskConst->shl(LShrAmt)3813                          : APInt::getBitsSetFrom(3814                                Int->getType()->getScalarSizeInBits(), LShrAmt);3815 3816  return true;3817}3818 3819/// Try to fold the join of two scalar integers whose bits are unpacked and3820/// zexted from the same source integer.3821static Value *foldIntegerRepackThroughZExt(Value *Lhs, Value *Rhs,3822                                           InstCombiner::BuilderTy &Builder) {3823 3824  Value *LhsInt, *RhsInt;3825  APInt LhsMask, RhsMask;3826  uint64_t LhsOffset, RhsOffset;3827  bool IsLhsShlNUW, IsLhsShlNSW, IsRhsShlNUW, IsRhsShlNSW;3828  if (!matchZExtedSubInteger(Lhs, LhsInt, LhsMask, LhsOffset, IsLhsShlNUW,3829                             IsLhsShlNSW))3830    return nullptr;3831  if (!matchZExtedSubInteger(Rhs, RhsInt, RhsMask, RhsOffset, IsRhsShlNUW,3832                             IsRhsShlNSW))3833    return nullptr;3834  if (LhsInt != RhsInt || LhsOffset != RhsOffset)3835    return nullptr;3836 3837  APInt Mask = LhsMask | RhsMask;3838 3839  Type *DestTy = Lhs->getType();3840  Value *Res = Builder.CreateShl(3841      Builder.CreateZExt(3842          Builder.CreateAnd(LhsInt, Mask, LhsInt->getName() + ".mask"), DestTy,3843          LhsInt->getName() + ".zext"),3844      ConstantInt::get(DestTy, LhsOffset), "", IsLhsShlNUW && IsRhsShlNUW,3845      IsLhsShlNSW && IsRhsShlNSW);3846  Res->takeName(Lhs);3847  return Res;3848}3849 3850// A decomposition of ((X & Mask) * Factor). The NUW / NSW bools3851// track these properities for preservation. Note that we can decompose3852// equivalent select form of this expression (e.g. (!(X & Mask) ? 0 : Mask *3853// Factor))3854struct DecomposedBitMaskMul {3855  Value *X;3856  APInt Factor;3857  APInt Mask;3858  bool NUW;3859  bool NSW;3860 3861  bool isCombineableWith(const DecomposedBitMaskMul Other) {3862    return X == Other.X && !Mask.intersects(Other.Mask) &&3863           Factor == Other.Factor;3864  }3865};3866 3867static std::optional<DecomposedBitMaskMul> matchBitmaskMul(Value *V) {3868  Instruction *Op = dyn_cast<Instruction>(V);3869  if (!Op)3870    return std::nullopt;3871 3872  // Decompose (A & N) * C) into BitMaskMul3873  Value *Original = nullptr;3874  const APInt *Mask = nullptr;3875  const APInt *MulConst = nullptr;3876  if (match(Op, m_Mul(m_And(m_Value(Original), m_APInt(Mask)),3877                      m_APInt(MulConst)))) {3878    if (MulConst->isZero() || Mask->isZero())3879      return std::nullopt;3880 3881    return std::optional<DecomposedBitMaskMul>(3882        {Original, *MulConst, *Mask,3883         cast<BinaryOperator>(Op)->hasNoUnsignedWrap(),3884         cast<BinaryOperator>(Op)->hasNoSignedWrap()});3885  }3886 3887  Value *Cond = nullptr;3888  const APInt *EqZero = nullptr, *NeZero = nullptr;3889 3890  // Decompose ((A & N) ? 0 : N * C) into BitMaskMul3891  if (match(Op, m_Select(m_Value(Cond), m_APInt(EqZero), m_APInt(NeZero)))) {3892    auto ICmpDecompose =3893        decomposeBitTest(Cond, /*LookThruTrunc=*/true,3894                         /*AllowNonZeroC=*/false, /*DecomposeBitMask=*/true);3895    if (!ICmpDecompose.has_value())3896      return std::nullopt;3897 3898    assert(ICmpInst::isEquality(ICmpDecompose->Pred) &&3899           ICmpDecompose->C.isZero());3900 3901    if (ICmpDecompose->Pred == ICmpInst::ICMP_NE)3902      std::swap(EqZero, NeZero);3903 3904    if (!EqZero->isZero() || NeZero->isZero())3905      return std::nullopt;3906 3907    if (!ICmpDecompose->Mask.isPowerOf2() || ICmpDecompose->Mask.isZero() ||3908        NeZero->getBitWidth() != ICmpDecompose->Mask.getBitWidth())3909      return std::nullopt;3910 3911    if (!NeZero->urem(ICmpDecompose->Mask).isZero())3912      return std::nullopt;3913 3914    return std::optional<DecomposedBitMaskMul>(3915        {ICmpDecompose->X, NeZero->udiv(ICmpDecompose->Mask),3916         ICmpDecompose->Mask, /*NUW=*/false, /*NSW=*/false});3917  }3918 3919  return std::nullopt;3920}3921 3922/// (A & N) * C + (A & M) * C -> (A & (N + M)) & C3923/// This also accepts the equivalent select form of (A & N) * C3924/// expressions i.e. !(A & N) ? 0 : N * C)3925static Value *foldBitmaskMul(Value *Op0, Value *Op1,3926                             InstCombiner::BuilderTy &Builder) {3927  auto Decomp1 = matchBitmaskMul(Op1);3928  if (!Decomp1)3929    return nullptr;3930 3931  auto Decomp0 = matchBitmaskMul(Op0);3932  if (!Decomp0)3933    return nullptr;3934 3935  if (Decomp0->isCombineableWith(*Decomp1)) {3936    Value *NewAnd = Builder.CreateAnd(3937        Decomp0->X,3938        ConstantInt::get(Decomp0->X->getType(), Decomp0->Mask + Decomp1->Mask));3939 3940    return Builder.CreateMul(3941        NewAnd, ConstantInt::get(NewAnd->getType(), Decomp1->Factor), "",3942        Decomp0->NUW && Decomp1->NUW, Decomp0->NSW && Decomp1->NSW);3943  }3944 3945  return nullptr;3946}3947 3948Value *InstCombinerImpl::foldDisjointOr(Value *LHS, Value *RHS) {3949  if (Value *Res = foldBitmaskMul(LHS, RHS, Builder))3950    return Res;3951  if (Value *Res = foldIntegerRepackThroughZExt(LHS, RHS, Builder))3952    return Res;3953 3954  return nullptr;3955}3956 3957Value *InstCombinerImpl::reassociateDisjointOr(Value *LHS, Value *RHS) {3958 3959  Value *X, *Y;3960  if (match(RHS, m_OneUse(m_DisjointOr(m_Value(X), m_Value(Y))))) {3961    if (Value *Res = foldDisjointOr(LHS, X))3962      return Builder.CreateOr(Res, Y, "", /*IsDisjoint=*/true);3963    if (Value *Res = foldDisjointOr(LHS, Y))3964      return Builder.CreateOr(Res, X, "", /*IsDisjoint=*/true);3965  }3966 3967  if (match(LHS, m_OneUse(m_DisjointOr(m_Value(X), m_Value(Y))))) {3968    if (Value *Res = foldDisjointOr(X, RHS))3969      return Builder.CreateOr(Res, Y, "", /*IsDisjoint=*/true);3970    if (Value *Res = foldDisjointOr(Y, RHS))3971      return Builder.CreateOr(Res, X, "", /*IsDisjoint=*/true);3972  }3973 3974  return nullptr;3975}3976 3977/// Fold Res, Overflow = (umul.with.overflow x c1); (or Overflow (ugt Res c2))3978/// --> (ugt x (c2/c1)). This code checks whether a multiplication of two3979/// unsigned numbers (one is a constant) is mathematically greater than a3980/// second constant.3981static Value *foldOrUnsignedUMulOverflowICmp(BinaryOperator &I,3982                                             InstCombiner::BuilderTy &Builder,3983                                             const DataLayout &DL) {3984  Value *WOV, *X;3985  const APInt *C1, *C2;3986  if (match(&I,3987            m_c_Or(m_ExtractValue<1>(3988                       m_Value(WOV, m_Intrinsic<Intrinsic::umul_with_overflow>(3989                                        m_Value(X), m_APInt(C1)))),3990                   m_OneUse(m_SpecificCmp(ICmpInst::ICMP_UGT,3991                                          m_ExtractValue<0>(m_Deferred(WOV)),3992                                          m_APInt(C2))))) &&3993      !C1->isZero()) {3994    Constant *NewC = ConstantInt::get(X->getType(), C2->udiv(*C1));3995    return Builder.CreateICmp(ICmpInst::ICMP_UGT, X, NewC);3996  }3997  return nullptr;3998}3999 4000/// Fold select(X >s 0, 0, -X) | smax(X, 0) --> abs(X)4001///      select(X <s 0, -X, 0) | smax(X, 0) --> abs(X)4002static Value *FoldOrOfSelectSmaxToAbs(BinaryOperator &I,4003                                      InstCombiner::BuilderTy &Builder) {4004  Value *X;4005  Value *Sel;4006  if (match(&I,4007            m_c_Or(m_Value(Sel), m_OneUse(m_SMax(m_Value(X), m_ZeroInt()))))) {4008    auto NegX = m_Neg(m_Specific(X));4009    if (match(Sel, m_Select(m_SpecificICmp(ICmpInst::ICMP_SGT, m_Specific(X),4010                                           m_ZeroInt()),4011                            m_ZeroInt(), NegX)) ||4012        match(Sel, m_Select(m_SpecificICmp(ICmpInst::ICMP_SLT, m_Specific(X),4013                                           m_ZeroInt()),4014                            NegX, m_ZeroInt())))4015      return Builder.CreateBinaryIntrinsic(Intrinsic::abs, X,4016                                           Builder.getFalse());4017  }4018  return nullptr;4019}4020 4021// FIXME: We use commutative matchers (m_c_*) for some, but not all, matches4022// here. We should standardize that construct where it is needed or choose some4023// other way to ensure that commutated variants of patterns are not missed.4024Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {4025  if (Value *V = simplifyOrInst(I.getOperand(0), I.getOperand(1),4026                                SQ.getWithInstruction(&I)))4027    return replaceInstUsesWith(I, V);4028 4029  if (SimplifyAssociativeOrCommutative(I))4030    return &I;4031 4032  if (Instruction *X = foldVectorBinop(I))4033    return X;4034 4035  if (Instruction *Phi = foldBinopWithPhiOperands(I))4036    return Phi;4037 4038  // See if we can simplify any instructions used by the instruction whose sole4039  // purpose is to compute bits we don't care about.4040  if (SimplifyDemandedInstructionBits(I))4041    return &I;4042 4043  // Do this before using distributive laws to catch simple and/or/not patterns.4044  if (Instruction *Xor = foldOrToXor(I, Builder))4045    return Xor;4046 4047  if (Instruction *X = foldComplexAndOrPatterns(I, Builder))4048    return X;4049 4050  if (Instruction *X = foldIntegerPackFromVector(I, Builder, DL))4051    return X;4052 4053  // (A & B) | (C & D) -> A ^ D where A == ~C && B == ~D4054  // (A & B) | (C & D) -> A ^ C where A == ~D && B == ~C4055  if (Value *V = foldOrOfInversions(I, Builder))4056    return replaceInstUsesWith(I, V);4057 4058  // (A&B)|(A&C) -> A&(B|C) etc4059  if (Value *V = foldUsingDistributiveLaws(I))4060    return replaceInstUsesWith(I, V);4061 4062  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);4063  Type *Ty = I.getType();4064  if (Ty->isIntOrIntVectorTy(1)) {4065    if (auto *SI0 = dyn_cast<SelectInst>(Op0)) {4066      if (auto *R =4067              foldAndOrOfSelectUsingImpliedCond(Op1, *SI0, /* IsAnd */ false))4068        return R;4069    }4070    if (auto *SI1 = dyn_cast<SelectInst>(Op1)) {4071      if (auto *R =4072              foldAndOrOfSelectUsingImpliedCond(Op0, *SI1, /* IsAnd */ false))4073        return R;4074    }4075  }4076 4077  if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))4078    return FoldedLogic;4079 4080  if (Instruction *BitOp = matchBSwapOrBitReverse(I, /*MatchBSwaps*/ true,4081                                                  /*MatchBitReversals*/ true))4082    return BitOp;4083 4084  if (Instruction *Funnel = matchFunnelShift(I, *this))4085    return Funnel;4086 4087  if (Value *Concat = matchOrConcat(I, Builder))4088    return replaceInstUsesWith(I, Concat);4089 4090  if (Instruction *R = foldBinOpShiftWithShift(I))4091    return R;4092 4093  if (Instruction *R = tryFoldInstWithCtpopWithNot(&I))4094    return R;4095 4096  if (cast<PossiblyDisjointInst>(I).isDisjoint()) {4097    if (Instruction *R =4098            foldAddLikeCommutative(I.getOperand(0), I.getOperand(1),4099                                   /*NSW=*/true, /*NUW=*/true))4100      return R;4101    if (Instruction *R =4102            foldAddLikeCommutative(I.getOperand(1), I.getOperand(0),4103                                   /*NSW=*/true, /*NUW=*/true))4104      return R;4105 4106    if (Value *Res = foldDisjointOr(I.getOperand(0), I.getOperand(1)))4107      return replaceInstUsesWith(I, Res);4108 4109    if (Value *Res = reassociateDisjointOr(I.getOperand(0), I.getOperand(1)))4110      return replaceInstUsesWith(I, Res);4111  }4112 4113  Value *X, *Y;4114  const APInt *CV;4115  if (match(&I, m_c_Or(m_OneUse(m_Xor(m_Value(X), m_APInt(CV))), m_Value(Y))) &&4116      !CV->isAllOnes() && MaskedValueIsZero(Y, *CV, &I)) {4117    // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 04118    // The check for a 'not' op is for efficiency (if Y is known zero --> ~X).4119    Value *Or = Builder.CreateOr(X, Y);4120    return BinaryOperator::CreateXor(Or, ConstantInt::get(Ty, *CV));4121  }4122 4123  // If the operands have no common bits set:4124  // or (mul X, Y), X --> add (mul X, Y), X --> mul X, (Y + 1)4125  if (match(&I, m_c_DisjointOr(m_OneUse(m_Mul(m_Value(X), m_Value(Y))),4126                               m_Deferred(X)))) {4127    Value *IncrementY = Builder.CreateAdd(Y, ConstantInt::get(Ty, 1));4128    return BinaryOperator::CreateMul(X, IncrementY);4129  }4130 4131  // (A & C) | (B & D)4132  Value *A, *B, *C, *D;4133  if (match(Op0, m_And(m_Value(A), m_Value(C))) &&4134      match(Op1, m_And(m_Value(B), m_Value(D)))) {4135 4136    // (A & C0) | (B & C1)4137    const APInt *C0, *C1;4138    if (match(C, m_APInt(C0)) && match(D, m_APInt(C1))) {4139      Value *X;4140      if (*C0 == ~*C1) {4141        // ((X | B) & MaskC) | (B & ~MaskC) -> (X & MaskC) | B4142        if (match(A, m_c_Or(m_Value(X), m_Specific(B))))4143          return BinaryOperator::CreateOr(Builder.CreateAnd(X, *C0), B);4144        // (A & MaskC) | ((X | A) & ~MaskC) -> (X & ~MaskC) | A4145        if (match(B, m_c_Or(m_Specific(A), m_Value(X))))4146          return BinaryOperator::CreateOr(Builder.CreateAnd(X, *C1), A);4147 4148        // ((X ^ B) & MaskC) | (B & ~MaskC) -> (X & MaskC) ^ B4149        if (match(A, m_c_Xor(m_Value(X), m_Specific(B))))4150          return BinaryOperator::CreateXor(Builder.CreateAnd(X, *C0), B);4151        // (A & MaskC) | ((X ^ A) & ~MaskC) -> (X & ~MaskC) ^ A4152        if (match(B, m_c_Xor(m_Specific(A), m_Value(X))))4153          return BinaryOperator::CreateXor(Builder.CreateAnd(X, *C1), A);4154      }4155 4156      if ((*C0 & *C1).isZero()) {4157        // ((X | B) & C0) | (B & C1) --> (X | B) & (C0 | C1)4158        // iff (C0 & C1) == 0 and (X & ~C0) == 04159        if (match(A, m_c_Or(m_Value(X), m_Specific(B))) &&4160            MaskedValueIsZero(X, ~*C0, &I)) {4161          Constant *C01 = ConstantInt::get(Ty, *C0 | *C1);4162          return BinaryOperator::CreateAnd(A, C01);4163        }4164        // (A & C0) | ((X | A) & C1) --> (X | A) & (C0 | C1)4165        // iff (C0 & C1) == 0 and (X & ~C1) == 04166        if (match(B, m_c_Or(m_Value(X), m_Specific(A))) &&4167            MaskedValueIsZero(X, ~*C1, &I)) {4168          Constant *C01 = ConstantInt::get(Ty, *C0 | *C1);4169          return BinaryOperator::CreateAnd(B, C01);4170        }4171        // ((X | C2) & C0) | ((X | C3) & C1) --> (X | C2 | C3) & (C0 | C1)4172        // iff (C0 & C1) == 0 and (C2 & ~C0) == 0 and (C3 & ~C1) == 0.4173        const APInt *C2, *C3;4174        if (match(A, m_Or(m_Value(X), m_APInt(C2))) &&4175            match(B, m_Or(m_Specific(X), m_APInt(C3))) &&4176            (*C2 & ~*C0).isZero() && (*C3 & ~*C1).isZero()) {4177          Value *Or = Builder.CreateOr(X, *C2 | *C3, "bitfield");4178          Constant *C01 = ConstantInt::get(Ty, *C0 | *C1);4179          return BinaryOperator::CreateAnd(Or, C01);4180        }4181      }4182    }4183 4184    // Don't try to form a select if it's unlikely that we'll get rid of at4185    // least one of the operands. A select is generally more expensive than the4186    // 'or' that it is replacing.4187    if (Op0->hasOneUse() || Op1->hasOneUse()) {4188      // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.4189      if (Value *V = matchSelectFromAndOr(A, C, B, D))4190        return replaceInstUsesWith(I, V);4191      if (Value *V = matchSelectFromAndOr(A, C, D, B))4192        return replaceInstUsesWith(I, V);4193      if (Value *V = matchSelectFromAndOr(C, A, B, D))4194        return replaceInstUsesWith(I, V);4195      if (Value *V = matchSelectFromAndOr(C, A, D, B))4196        return replaceInstUsesWith(I, V);4197      if (Value *V = matchSelectFromAndOr(B, D, A, C))4198        return replaceInstUsesWith(I, V);4199      if (Value *V = matchSelectFromAndOr(B, D, C, A))4200        return replaceInstUsesWith(I, V);4201      if (Value *V = matchSelectFromAndOr(D, B, A, C))4202        return replaceInstUsesWith(I, V);4203      if (Value *V = matchSelectFromAndOr(D, B, C, A))4204        return replaceInstUsesWith(I, V);4205    }4206  }4207 4208  if (match(Op0, m_And(m_Value(A), m_Value(C))) &&4209      match(Op1, m_Not(m_Or(m_Value(B), m_Value(D)))) &&4210      (Op0->hasOneUse() || Op1->hasOneUse())) {4211    // (Cond & C) | ~(Cond | D) -> Cond ? C : ~D4212    if (Value *V = matchSelectFromAndOr(A, C, B, D, true))4213      return replaceInstUsesWith(I, V);4214    if (Value *V = matchSelectFromAndOr(A, C, D, B, true))4215      return replaceInstUsesWith(I, V);4216    if (Value *V = matchSelectFromAndOr(C, A, B, D, true))4217      return replaceInstUsesWith(I, V);4218    if (Value *V = matchSelectFromAndOr(C, A, D, B, true))4219      return replaceInstUsesWith(I, V);4220  }4221 4222  // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C4223  if (match(Op0, m_Xor(m_Value(A), m_Value(B))))4224    if (match(Op1,4225              m_c_Xor(m_c_Xor(m_Specific(B), m_Value(C)), m_Specific(A))) ||4226        match(Op1, m_c_Xor(m_c_Xor(m_Specific(A), m_Value(C)), m_Specific(B))))4227      return BinaryOperator::CreateOr(Op0, C);4228 4229  // ((B ^ C) ^ A) | (A ^ B) -> (A ^ B) | C4230  if (match(Op1, m_Xor(m_Value(A), m_Value(B))))4231    if (match(Op0,4232              m_c_Xor(m_c_Xor(m_Specific(B), m_Value(C)), m_Specific(A))) ||4233        match(Op0, m_c_Xor(m_c_Xor(m_Specific(A), m_Value(C)), m_Specific(B))))4234      return BinaryOperator::CreateOr(Op1, C);4235 4236  if (Instruction *DeMorgan = matchDeMorgansLaws(I, *this))4237    return DeMorgan;4238 4239  // Canonicalize xor to the RHS.4240  bool SwappedForXor = false;4241  if (match(Op0, m_Xor(m_Value(), m_Value()))) {4242    std::swap(Op0, Op1);4243    SwappedForXor = true;4244  }4245 4246  if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {4247    // (A | ?) | (A ^ B) --> (A | ?) | B4248    // (B | ?) | (A ^ B) --> (B | ?) | A4249    if (match(Op0, m_c_Or(m_Specific(A), m_Value())))4250      return BinaryOperator::CreateOr(Op0, B);4251    if (match(Op0, m_c_Or(m_Specific(B), m_Value())))4252      return BinaryOperator::CreateOr(Op0, A);4253 4254    // (A & B) | (A ^ B) --> A | B4255    // (B & A) | (A ^ B) --> A | B4256    if (match(Op0, m_c_And(m_Specific(A), m_Specific(B))))4257      return BinaryOperator::CreateOr(A, B);4258 4259    // ~A | (A ^ B) --> ~(A & B)4260    // ~B | (A ^ B) --> ~(A & B)4261    // The swap above should always make Op0 the 'not'.4262    if ((Op0->hasOneUse() || Op1->hasOneUse()) &&4263        (match(Op0, m_Not(m_Specific(A))) || match(Op0, m_Not(m_Specific(B)))))4264      return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));4265 4266    // Same as above, but peek through an 'and' to the common operand:4267    // ~(A & ?) | (A ^ B) --> ~((A & ?) & B)4268    // ~(B & ?) | (A ^ B) --> ~((B & ?) & A)4269    Instruction *And;4270    if ((Op0->hasOneUse() || Op1->hasOneUse()) &&4271        match(Op0,4272              m_Not(m_Instruction(And, m_c_And(m_Specific(A), m_Value())))))4273      return BinaryOperator::CreateNot(Builder.CreateAnd(And, B));4274    if ((Op0->hasOneUse() || Op1->hasOneUse()) &&4275        match(Op0,4276              m_Not(m_Instruction(And, m_c_And(m_Specific(B), m_Value())))))4277      return BinaryOperator::CreateNot(Builder.CreateAnd(And, A));4278 4279    // (~A | C) | (A ^ B) --> ~(A & B) | C4280    // (~B | C) | (A ^ B) --> ~(A & B) | C4281    if (Op0->hasOneUse() && Op1->hasOneUse() &&4282        (match(Op0, m_c_Or(m_Not(m_Specific(A)), m_Value(C))) ||4283         match(Op0, m_c_Or(m_Not(m_Specific(B)), m_Value(C))))) {4284      Value *Nand = Builder.CreateNot(Builder.CreateAnd(A, B), "nand");4285      return BinaryOperator::CreateOr(Nand, C);4286    }4287  }4288 4289  if (SwappedForXor)4290    std::swap(Op0, Op1);4291 4292  if (Value *Res =4293          foldBooleanAndOr(Op0, Op1, I, /*IsAnd=*/false, /*IsLogical=*/false))4294    return replaceInstUsesWith(I, Res);4295 4296  if (match(Op1, m_OneUse(m_LogicalOr(m_Value(X), m_Value(Y))))) {4297    bool IsLogical = isa<SelectInst>(Op1);4298    if (auto *V = reassociateBooleanAndOr(Op0, X, Y, I, /*IsAnd=*/false,4299                                          /*RHSIsLogical=*/IsLogical))4300      return replaceInstUsesWith(I, V);4301  }4302  if (match(Op0, m_OneUse(m_LogicalOr(m_Value(X), m_Value(Y))))) {4303    bool IsLogical = isa<SelectInst>(Op0);4304    if (auto *V = reassociateBooleanAndOr(Op1, X, Y, I, /*IsAnd=*/false,4305                                          /*RHSIsLogical=*/IsLogical))4306      return replaceInstUsesWith(I, V);4307  }4308 4309  if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))4310    return FoldedFCmps;4311 4312  if (Instruction *CastedOr = foldCastedBitwiseLogic(I))4313    return CastedOr;4314 4315  if (Instruction *Sel = foldBinopOfSextBoolToSelect(I))4316    return Sel;4317 4318  // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.4319  // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold4320  //       with binop identity constant. But creating a select with non-constant4321  //       arm may not be reversible due to poison semantics. Is that a good4322  //       canonicalization?4323  if (match(&I, m_c_Or(m_OneUse(m_SExt(m_Value(A))), m_Value(B))) &&4324      A->getType()->isIntOrIntVectorTy(1))4325    return SelectInst::Create(A, ConstantInt::getAllOnesValue(Ty), B);4326 4327  // Note: If we've gotten to the point of visiting the outer OR, then the4328  // inner one couldn't be simplified.  If it was a constant, then it won't4329  // be simplified by a later pass either, so we try swapping the inner/outer4330  // ORs in the hopes that we'll be able to simplify it this way.4331  // (X|C) | V --> (X|V) | C4332  // Pass the disjoint flag in the following two patterns:4333  // 1. or-disjoint (or-disjoint X, C), V -->4334  //    or-disjoint (or-disjoint X, V), C4335  //4336  // 2. or-disjoint (or X, C), V -->4337  //    or (or-disjoint X, V), C4338  ConstantInt *CI;4339  if (Op0->hasOneUse() && !match(Op1, m_ConstantInt()) &&4340      match(Op0, m_Or(m_Value(A), m_ConstantInt(CI)))) {4341    bool IsDisjointOuter = cast<PossiblyDisjointInst>(I).isDisjoint();4342    bool IsDisjointInner = cast<PossiblyDisjointInst>(Op0)->isDisjoint();4343    Value *Inner = Builder.CreateOr(A, Op1);4344    cast<PossiblyDisjointInst>(Inner)->setIsDisjoint(IsDisjointOuter);4345    Inner->takeName(Op0);4346    return IsDisjointOuter && IsDisjointInner4347               ? BinaryOperator::CreateDisjointOr(Inner, CI)4348               : BinaryOperator::CreateOr(Inner, CI);4349  }4350 4351  // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))4352  // Since this OR statement hasn't been optimized further yet, we hope4353  // that this transformation will allow the new ORs to be optimized.4354  {4355    Value *X = nullptr, *Y = nullptr;4356    if (Op0->hasOneUse() && Op1->hasOneUse() &&4357        match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&4358        match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {4359      Value *orTrue = Builder.CreateOr(A, C);4360      Value *orFalse = Builder.CreateOr(B, D);4361      return SelectInst::Create(X, orTrue, orFalse);4362    }4363  }4364 4365  // or(ashr(subNSW(Y, X), ScalarSizeInBits(Y) - 1), X)  --> X s> Y ? -1 : X.4366  {4367    Value *X, *Y;4368    if (match(&I, m_c_Or(m_OneUse(m_AShr(4369                             m_NSWSub(m_Value(Y), m_Value(X)),4370                             m_SpecificInt(Ty->getScalarSizeInBits() - 1))),4371                         m_Deferred(X)))) {4372      Value *NewICmpInst = Builder.CreateICmpSGT(X, Y);4373      Value *AllOnes = ConstantInt::getAllOnesValue(Ty);4374      return SelectInst::Create(NewICmpInst, AllOnes, X);4375    }4376  }4377 4378  {4379    // ((A & B) ^ A) | ((A & B) ^ B) -> A ^ B4380    // (A ^ (A & B)) | (B ^ (A & B)) -> A ^ B4381    // ((A & B) ^ B) | ((A & B) ^ A) -> A ^ B4382    // (B ^ (A & B)) | (A ^ (A & B)) -> A ^ B4383    const auto TryXorOpt = [&](Value *Lhs, Value *Rhs) -> Instruction * {4384      if (match(Lhs, m_c_Xor(m_And(m_Value(A), m_Value(B)), m_Deferred(A))) &&4385          match(Rhs,4386                m_c_Xor(m_And(m_Specific(A), m_Specific(B)), m_Specific(B)))) {4387        return BinaryOperator::CreateXor(A, B);4388      }4389      return nullptr;4390    };4391 4392    if (Instruction *Result = TryXorOpt(Op0, Op1))4393      return Result;4394    if (Instruction *Result = TryXorOpt(Op1, Op0))4395      return Result;4396  }4397 4398  if (Instruction *V =4399          canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))4400    return V;4401 4402  CmpPredicate Pred;4403  Value *Mul, *Ov, *MulIsNotZero, *UMulWithOv;4404  // Check if the OR weakens the overflow condition for umul.with.overflow by4405  // treating any non-zero result as overflow. In that case, we overflow if both4406  // umul.with.overflow operands are != 0, as in that case the result can only4407  // be 0, iff the multiplication overflows.4408  if (match(&I, m_c_Or(m_Value(Ov, m_ExtractValue<1>(m_Value(UMulWithOv))),4409                       m_Value(MulIsNotZero,4410                               m_SpecificICmp(4411                                   ICmpInst::ICMP_NE,4412                                   m_Value(Mul, m_ExtractValue<0>(4413                                                    m_Deferred(UMulWithOv))),4414                                   m_ZeroInt())))) &&4415      (Ov->hasOneUse() || (MulIsNotZero->hasOneUse() && Mul->hasOneUse()))) {4416    Value *A, *B;4417    if (match(UMulWithOv, m_Intrinsic<Intrinsic::umul_with_overflow>(4418                              m_Value(A), m_Value(B)))) {4419      Value *NotNullA = Builder.CreateIsNotNull(A);4420      Value *NotNullB = Builder.CreateIsNotNull(B);4421      return BinaryOperator::CreateAnd(NotNullA, NotNullB);4422    }4423  }4424 4425  /// Res, Overflow = xxx_with_overflow X, C14426  /// Try to canonicalize the pattern "Overflow | icmp pred Res, C2" into4427  /// "Overflow | icmp pred X, C2 +/- C1".4428  const WithOverflowInst *WO;4429  const Value *WOV;4430  const APInt *C1, *C2;4431  if (match(&I, m_c_Or(m_Value(Ov, m_ExtractValue<1>(4432                                       m_Value(WOV, m_WithOverflowInst(WO)))),4433                       m_OneUse(m_ICmp(Pred, m_ExtractValue<0>(m_Deferred(WOV)),4434                                       m_APInt(C2))))) &&4435      (WO->getBinaryOp() == Instruction::Add ||4436       WO->getBinaryOp() == Instruction::Sub) &&4437      (ICmpInst::isEquality(Pred) ||4438       WO->isSigned() == ICmpInst::isSigned(Pred)) &&4439      match(WO->getRHS(), m_APInt(C1))) {4440    bool Overflow;4441    APInt NewC = WO->getBinaryOp() == Instruction::Add4442                     ? (ICmpInst::isSigned(Pred) ? C2->ssub_ov(*C1, Overflow)4443                                                 : C2->usub_ov(*C1, Overflow))4444                     : (ICmpInst::isSigned(Pred) ? C2->sadd_ov(*C1, Overflow)4445                                                 : C2->uadd_ov(*C1, Overflow));4446    if (!Overflow || ICmpInst::isEquality(Pred)) {4447      Value *NewCmp = Builder.CreateICmp(4448          Pred, WO->getLHS(), ConstantInt::get(WO->getLHS()->getType(), NewC));4449      return BinaryOperator::CreateOr(Ov, NewCmp);4450    }4451  }4452 4453  // Try to fold the pattern "Overflow | icmp pred Res, C2" into a single4454  // comparison instruction for umul.with.overflow.4455  if (Value *R = foldOrUnsignedUMulOverflowICmp(I, Builder, DL))4456    return replaceInstUsesWith(I, R);4457 4458  // (~x) | y  -->  ~(x & (~y))  iff that gets rid of inversions4459  if (sinkNotIntoOtherHandOfLogicalOp(I))4460    return &I;4461 4462  // Improve "get low bit mask up to and including bit X" pattern:4463  //   (1 << X) | ((1 << X) + -1)  -->  -1 l>> (bitwidth(x) - 1 - X)4464  if (match(&I, m_c_Or(m_Add(m_Shl(m_One(), m_Value(X)), m_AllOnes()),4465                       m_Shl(m_One(), m_Deferred(X)))) &&4466      match(&I, m_c_Or(m_OneUse(m_Value()), m_Value()))) {4467    Value *Sub = Builder.CreateSub(4468        ConstantInt::get(Ty, Ty->getScalarSizeInBits() - 1), X);4469    return BinaryOperator::CreateLShr(Constant::getAllOnesValue(Ty), Sub);4470  }4471 4472  // An or recurrence w/loop invariant step is equivelent to (or start, step)4473  PHINode *PN = nullptr;4474  Value *Start = nullptr, *Step = nullptr;4475  if (matchSimpleRecurrence(&I, PN, Start, Step) && DT.dominates(Step, PN))4476    return replaceInstUsesWith(I, Builder.CreateOr(Start, Step));4477 4478  // (A & B) | (C | D) or (C | D) | (A & B)4479  // Can be combined if C or D is of type (A/B & X)4480  if (match(&I, m_c_Or(m_OneUse(m_And(m_Value(A), m_Value(B))),4481                       m_OneUse(m_Or(m_Value(C), m_Value(D)))))) {4482    // (A & B) | (C | ?) -> C | (? | (A & B))4483    // (A & B) | (C | ?) -> C | (? | (A & B))4484    // (A & B) | (C | ?) -> C | (? | (A & B))4485    // (A & B) | (C | ?) -> C | (? | (A & B))4486    // (C | ?) | (A & B) -> C | (? | (A & B))4487    // (C | ?) | (A & B) -> C | (? | (A & B))4488    // (C | ?) | (A & B) -> C | (? | (A & B))4489    // (C | ?) | (A & B) -> C | (? | (A & B))4490    if (match(D, m_OneUse(m_c_And(m_Specific(A), m_Value()))) ||4491        match(D, m_OneUse(m_c_And(m_Specific(B), m_Value()))))4492      return BinaryOperator::CreateOr(4493          C, Builder.CreateOr(D, Builder.CreateAnd(A, B)));4494    // (A & B) | (? | D) -> (? | (A & B)) | D4495    // (A & B) | (? | D) -> (? | (A & B)) | D4496    // (A & B) | (? | D) -> (? | (A & B)) | D4497    // (A & B) | (? | D) -> (? | (A & B)) | D4498    // (? | D) | (A & B) -> (? | (A & B)) | D4499    // (? | D) | (A & B) -> (? | (A & B)) | D4500    // (? | D) | (A & B) -> (? | (A & B)) | D4501    // (? | D) | (A & B) -> (? | (A & B)) | D4502    if (match(C, m_OneUse(m_c_And(m_Specific(A), m_Value()))) ||4503        match(C, m_OneUse(m_c_And(m_Specific(B), m_Value()))))4504      return BinaryOperator::CreateOr(4505          Builder.CreateOr(C, Builder.CreateAnd(A, B)), D);4506  }4507 4508  if (Instruction *R = reassociateForUses(I, Builder))4509    return R;4510 4511  if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))4512    return Canonicalized;4513 4514  if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1))4515    return Folded;4516 4517  if (Instruction *Res = foldBinOpOfDisplacedShifts(I))4518    return Res;4519 4520  // If we are setting the sign bit of a floating-point value, convert4521  // this to fneg(fabs), then cast back to integer.4522  //4523  // If the result isn't immediately cast back to a float, this will increase4524  // the number of instructions. This is still probably a better canonical form4525  // as it enables FP value tracking.4526  //4527  // Assumes any IEEE-represented type has the sign bit in the high bit.4528  //4529  // This is generous interpretation of noimplicitfloat, this is not a true4530  // floating-point operation.4531  Value *CastOp;4532  if (match(Op0, m_ElementWiseBitCast(m_Value(CastOp))) &&4533      match(Op1, m_SignMask()) &&4534      !Builder.GetInsertBlock()->getParent()->hasFnAttribute(4535          Attribute::NoImplicitFloat)) {4536    Type *EltTy = CastOp->getType()->getScalarType();4537    if (EltTy->isFloatingPointTy() &&4538        APFloat::hasSignBitInMSB(EltTy->getFltSemantics())) {4539      Value *FAbs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, CastOp);4540      Value *FNegFAbs = Builder.CreateFNeg(FAbs);4541      return new BitCastInst(FNegFAbs, I.getType());4542    }4543  }4544 4545  // (X & C1) | C2 -> X & (C1 | C2) iff (X & C2) == C24546  if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C1)))) &&4547      match(Op1, m_APInt(C2))) {4548    KnownBits KnownX = computeKnownBits(X, &I);4549    if ((KnownX.One & *C2) == *C2)4550      return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *C1 | *C2));4551  }4552 4553  if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))4554    return Res;4555 4556  if (Value *V =4557          simplifyAndOrWithOpReplaced(Op0, Op1, Constant::getNullValue(Ty),4558                                      /*SimplifyOnly*/ false, *this))4559    return BinaryOperator::CreateOr(V, Op1);4560  if (Value *V =4561          simplifyAndOrWithOpReplaced(Op1, Op0, Constant::getNullValue(Ty),4562                                      /*SimplifyOnly*/ false, *this))4563    return BinaryOperator::CreateOr(Op0, V);4564 4565  if (cast<PossiblyDisjointInst>(I).isDisjoint())4566    if (Value *V = SimplifyAddWithRemainder(I))4567      return replaceInstUsesWith(I, V);4568 4569  if (Value *Res = FoldOrOfSelectSmaxToAbs(I, Builder))4570    return replaceInstUsesWith(I, Res);4571 4572  return nullptr;4573}4574 4575/// A ^ B can be specified using other logic ops in a variety of patterns. We4576/// can fold these early and efficiently by morphing an existing instruction.4577static Instruction *foldXorToXor(BinaryOperator &I,4578                                 InstCombiner::BuilderTy &Builder) {4579  assert(I.getOpcode() == Instruction::Xor);4580  Value *Op0 = I.getOperand(0);4581  Value *Op1 = I.getOperand(1);4582  Value *A, *B;4583 4584  // There are 4 commuted variants for each of the basic patterns.4585 4586  // (A & B) ^ (A | B) -> A ^ B4587  // (A & B) ^ (B | A) -> A ^ B4588  // (A | B) ^ (A & B) -> A ^ B4589  // (A | B) ^ (B & A) -> A ^ B4590  if (match(&I, m_c_Xor(m_And(m_Value(A), m_Value(B)),4591                        m_c_Or(m_Deferred(A), m_Deferred(B)))))4592    return BinaryOperator::CreateXor(A, B);4593 4594  // (A | ~B) ^ (~A | B) -> A ^ B4595  // (~B | A) ^ (~A | B) -> A ^ B4596  // (~A | B) ^ (A | ~B) -> A ^ B4597  // (B | ~A) ^ (A | ~B) -> A ^ B4598  if (match(&I, m_Xor(m_c_Or(m_Value(A), m_Not(m_Value(B))),4599                      m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))4600    return BinaryOperator::CreateXor(A, B);4601 4602  // (A & ~B) ^ (~A & B) -> A ^ B4603  // (~B & A) ^ (~A & B) -> A ^ B4604  // (~A & B) ^ (A & ~B) -> A ^ B4605  // (B & ~A) ^ (A & ~B) -> A ^ B4606  if (match(&I, m_Xor(m_c_And(m_Value(A), m_Not(m_Value(B))),4607                      m_c_And(m_Not(m_Deferred(A)), m_Deferred(B)))))4608    return BinaryOperator::CreateXor(A, B);4609 4610  // For the remaining cases we need to get rid of one of the operands.4611  if (!Op0->hasOneUse() && !Op1->hasOneUse())4612    return nullptr;4613 4614  // (A | B) ^ ~(A & B) -> ~(A ^ B)4615  // (A | B) ^ ~(B & A) -> ~(A ^ B)4616  // (A & B) ^ ~(A | B) -> ~(A ^ B)4617  // (A & B) ^ ~(B | A) -> ~(A ^ B)4618  // Complexity sorting ensures the not will be on the right side.4619  if ((match(Op0, m_Or(m_Value(A), m_Value(B))) &&4620       match(Op1, m_Not(m_c_And(m_Specific(A), m_Specific(B))))) ||4621      (match(Op0, m_And(m_Value(A), m_Value(B))) &&4622       match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))))4623    return BinaryOperator::CreateNot(Builder.CreateXor(A, B));4624 4625  return nullptr;4626}4627 4628Value *InstCombinerImpl::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS,4629                                        BinaryOperator &I) {4630  assert(I.getOpcode() == Instruction::Xor && I.getOperand(0) == LHS &&4631         I.getOperand(1) == RHS && "Should be 'xor' with these operands");4632 4633  ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();4634  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);4635  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);4636 4637  if (predicatesFoldable(PredL, PredR)) {4638    if (LHS0 == RHS1 && LHS1 == RHS0) {4639      std::swap(LHS0, LHS1);4640      PredL = ICmpInst::getSwappedPredicate(PredL);4641    }4642    if (LHS0 == RHS0 && LHS1 == RHS1) {4643      // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)4644      unsigned Code = getICmpCode(PredL) ^ getICmpCode(PredR);4645      bool IsSigned = LHS->isSigned() || RHS->isSigned();4646      return getNewICmpValue(Code, IsSigned, LHS0, LHS1, Builder);4647    }4648  }4649 4650  const APInt *LC, *RC;4651  if (match(LHS1, m_APInt(LC)) && match(RHS1, m_APInt(RC)) &&4652      LHS0->getType() == RHS0->getType() &&4653      LHS0->getType()->isIntOrIntVectorTy()) {4654    // Convert xor of signbit tests to signbit test of xor'd values:4655    // (X > -1) ^ (Y > -1) --> (X ^ Y) < 04656    // (X <  0) ^ (Y <  0) --> (X ^ Y) < 04657    // (X > -1) ^ (Y <  0) --> (X ^ Y) > -14658    // (X <  0) ^ (Y > -1) --> (X ^ Y) > -14659    bool TrueIfSignedL, TrueIfSignedR;4660    if ((LHS->hasOneUse() || RHS->hasOneUse()) &&4661        isSignBitCheck(PredL, *LC, TrueIfSignedL) &&4662        isSignBitCheck(PredR, *RC, TrueIfSignedR)) {4663      Value *XorLR = Builder.CreateXor(LHS0, RHS0);4664      return TrueIfSignedL == TrueIfSignedR ? Builder.CreateIsNeg(XorLR) :4665                                              Builder.CreateIsNotNeg(XorLR);4666    }4667 4668    // Fold (icmp pred1 X, C1) ^ (icmp pred2 X, C2)4669    // into a single comparison using range-based reasoning.4670    if (LHS0 == RHS0) {4671      ConstantRange CR1 = ConstantRange::makeExactICmpRegion(PredL, *LC);4672      ConstantRange CR2 = ConstantRange::makeExactICmpRegion(PredR, *RC);4673      auto CRUnion = CR1.exactUnionWith(CR2);4674      auto CRIntersect = CR1.exactIntersectWith(CR2);4675      if (CRUnion && CRIntersect)4676        if (auto CR = CRUnion->exactIntersectWith(CRIntersect->inverse())) {4677          if (CR->isFullSet())4678            return ConstantInt::getTrue(I.getType());4679          if (CR->isEmptySet())4680            return ConstantInt::getFalse(I.getType());4681 4682          CmpInst::Predicate NewPred;4683          APInt NewC, Offset;4684          CR->getEquivalentICmp(NewPred, NewC, Offset);4685 4686          if ((Offset.isZero() && (LHS->hasOneUse() || RHS->hasOneUse())) ||4687              (LHS->hasOneUse() && RHS->hasOneUse())) {4688            Value *NewV = LHS0;4689            Type *Ty = LHS0->getType();4690            if (!Offset.isZero())4691              NewV = Builder.CreateAdd(NewV, ConstantInt::get(Ty, Offset));4692            return Builder.CreateICmp(NewPred, NewV,4693                                      ConstantInt::get(Ty, NewC));4694          }4695        }4696    }4697 4698    // Fold (icmp eq/ne (X & Pow2), 0) ^ (icmp eq/ne (Y & Pow2), 0) into4699    // (icmp eq/ne ((X ^ Y) & Pow2), 0)4700    Value *X, *Y, *Pow2;4701    if (ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&4702        LC->isZero() && RC->isZero() && LHS->hasOneUse() && RHS->hasOneUse() &&4703        match(LHS0, m_And(m_Value(X), m_Value(Pow2))) &&4704        match(RHS0, m_And(m_Value(Y), m_Specific(Pow2))) &&4705        isKnownToBeAPowerOfTwo(Pow2, /*OrZero=*/true, &I)) {4706      Value *Xor = Builder.CreateXor(X, Y);4707      Value *And = Builder.CreateAnd(Xor, Pow2);4708      return Builder.CreateICmp(PredL == PredR ? ICmpInst::ICMP_NE4709                                               : ICmpInst::ICMP_EQ,4710                                And, ConstantInt::getNullValue(Xor->getType()));4711    }4712  }4713 4714  // Instead of trying to imitate the folds for and/or, decompose this 'xor'4715  // into those logic ops. That is, try to turn this into an and-of-icmps4716  // because we have many folds for that pattern.4717  //4718  // This is based on a truth table definition of xor:4719  // X ^ Y --> (X | Y) & !(X & Y)4720  if (Value *OrICmp = simplifyBinOp(Instruction::Or, LHS, RHS, SQ)) {4721    // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y).4722    // TODO: If OrICmp is false, the whole thing is false (InstSimplify?).4723    if (Value *AndICmp = simplifyBinOp(Instruction::And, LHS, RHS, SQ)) {4724      // TODO: Independently handle cases where the 'and' side is a constant.4725      ICmpInst *X = nullptr, *Y = nullptr;4726      if (OrICmp == LHS && AndICmp == RHS) {4727        // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS  --> X & !Y4728        X = LHS;4729        Y = RHS;4730      }4731      if (OrICmp == RHS && AndICmp == LHS) {4732        // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS  --> !Y & X4733        X = RHS;4734        Y = LHS;4735      }4736      if (X && Y && (Y->hasOneUse() || canFreelyInvertAllUsersOf(Y, &I))) {4737        // Invert the predicate of 'Y', thus inverting its output.4738        Y->setPredicate(Y->getInversePredicate());4739        // So, are there other uses of Y?4740        if (!Y->hasOneUse()) {4741          // We need to adapt other uses of Y though. Get a value that matches4742          // the original value of Y before inversion. While this increases4743          // immediate instruction count, we have just ensured that all the4744          // users are freely-invertible, so that 'not' *will* get folded away.4745          BuilderTy::InsertPointGuard Guard(Builder);4746          // Set insertion point to right after the Y.4747          Builder.SetInsertPoint(Y->getParent(), ++(Y->getIterator()));4748          Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");4749          // Replace all uses of Y (excluding the one in NotY!) with NotY.4750          Worklist.pushUsersToWorkList(*Y);4751          Y->replaceUsesWithIf(NotY,4752                               [NotY](Use &U) { return U.getUser() != NotY; });4753        }4754        // All done.4755        return Builder.CreateAnd(LHS, RHS);4756      }4757    }4758  }4759 4760  return nullptr;4761}4762 4763/// If we have a masked merge, in the canonical form of:4764/// (assuming that A only has one use.)4765///   |        A  |  |B|4766///   ((x ^ y) & M) ^ y4767///    |  D  |4768/// * If M is inverted:4769///      |  D  |4770///     ((x ^ y) & ~M) ^ y4771///   We can canonicalize by swapping the final xor operand4772///   to eliminate the 'not' of the mask.4773///     ((x ^ y) & M) ^ x4774/// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops4775///   because that shortens the dependency chain and improves analysis:4776///     (x & M) | (y & ~M)4777static Instruction *visitMaskedMerge(BinaryOperator &I,4778                                     InstCombiner::BuilderTy &Builder) {4779  Value *B, *X, *D;4780  Value *M;4781  if (!match(&I, m_c_Xor(m_Value(B),4782                         m_OneUse(m_c_And(4783                             m_Value(D, m_c_Xor(m_Deferred(B), m_Value(X))),4784                             m_Value(M))))))4785    return nullptr;4786 4787  Value *NotM;4788  if (match(M, m_Not(m_Value(NotM)))) {4789    // De-invert the mask and swap the value in B part.4790    Value *NewA = Builder.CreateAnd(D, NotM);4791    return BinaryOperator::CreateXor(NewA, X);4792  }4793 4794  Constant *C;4795  if (D->hasOneUse() && match(M, m_Constant(C))) {4796    // Propagating undef is unsafe. Clamp undef elements to -1.4797    Type *EltTy = C->getType()->getScalarType();4798    C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy));4799    // Unfold.4800    Value *LHS = Builder.CreateAnd(X, C);4801    Value *NotC = Builder.CreateNot(C);4802    Value *RHS = Builder.CreateAnd(B, NotC);4803    return BinaryOperator::CreateOr(LHS, RHS);4804  }4805 4806  return nullptr;4807}4808 4809static Instruction *foldNotXor(BinaryOperator &I,4810                               InstCombiner::BuilderTy &Builder) {4811  Value *X, *Y;4812  // FIXME: one-use check is not needed in general, but currently we are unable4813  // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182)4814  if (!match(&I, m_Not(m_OneUse(m_Xor(m_Value(X), m_Value(Y))))))4815    return nullptr;4816 4817  auto hasCommonOperand = [](Value *A, Value *B, Value *C, Value *D) {4818    return A == C || A == D || B == C || B == D;4819  };4820 4821  Value *A, *B, *C, *D;4822  // Canonicalize ~((A & B) ^ (A | ?)) -> (A & B) | ~(A | ?)4823  // 4 commuted variants4824  if (match(X, m_And(m_Value(A), m_Value(B))) &&4825      match(Y, m_Or(m_Value(C), m_Value(D))) && hasCommonOperand(A, B, C, D)) {4826    Value *NotY = Builder.CreateNot(Y);4827    return BinaryOperator::CreateOr(X, NotY);4828  };4829 4830  // Canonicalize ~((A | ?) ^ (A & B)) -> (A & B) | ~(A | ?)4831  // 4 commuted variants4832  if (match(Y, m_And(m_Value(A), m_Value(B))) &&4833      match(X, m_Or(m_Value(C), m_Value(D))) && hasCommonOperand(A, B, C, D)) {4834    Value *NotX = Builder.CreateNot(X);4835    return BinaryOperator::CreateOr(Y, NotX);4836  };4837 4838  return nullptr;4839}4840 4841/// Canonicalize a shifty way to code absolute value to the more common pattern4842/// that uses negation and select.4843static Instruction *canonicalizeAbs(BinaryOperator &Xor,4844                                    InstCombiner::BuilderTy &Builder) {4845  assert(Xor.getOpcode() == Instruction::Xor && "Expected an xor instruction.");4846 4847  // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1.4848  // We're relying on the fact that we only do this transform when the shift has4849  // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase4850  // instructions).4851  Value *Op0 = Xor.getOperand(0), *Op1 = Xor.getOperand(1);4852  if (Op0->hasNUses(2))4853    std::swap(Op0, Op1);4854 4855  Type *Ty = Xor.getType();4856  Value *A;4857  const APInt *ShAmt;4858  if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&4859      Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&4860      match(Op0, m_OneUse(m_c_Add(m_Specific(A), m_Specific(Op1))))) {4861    // Op1 = ashr i32 A, 31   ; smear the sign bit4862    // xor (add A, Op1), Op1  ; add -1 and flip bits if negative4863    // --> (A < 0) ? -A : A4864    Value *IsNeg = Builder.CreateIsNeg(A);4865    // Copy the nsw flags from the add to the negate.4866    auto *Add = cast<BinaryOperator>(Op0);4867    Value *NegA = Add->hasNoUnsignedWrap()4868                      ? Constant::getNullValue(A->getType())4869                      : Builder.CreateNeg(A, "", Add->hasNoSignedWrap());4870    return SelectInst::Create(IsNeg, NegA, A);4871  }4872  return nullptr;4873}4874 4875static bool canFreelyInvert(InstCombiner &IC, Value *Op,4876                            Instruction *IgnoredUser) {4877  auto *I = dyn_cast<Instruction>(Op);4878  return I && IC.isFreeToInvert(I, /*WillInvertAllUses=*/true) &&4879         IC.canFreelyInvertAllUsersOf(I, IgnoredUser);4880}4881 4882static Value *freelyInvert(InstCombinerImpl &IC, Value *Op,4883                           Instruction *IgnoredUser) {4884  auto *I = cast<Instruction>(Op);4885  IC.Builder.SetInsertPoint(*I->getInsertionPointAfterDef());4886  Value *NotOp = IC.Builder.CreateNot(Op, Op->getName() + ".not");4887  Op->replaceUsesWithIf(NotOp,4888                        [NotOp](Use &U) { return U.getUser() != NotOp; });4889  IC.freelyInvertAllUsersOf(NotOp, IgnoredUser);4890  return NotOp;4891}4892 4893// Transform4894//   z = ~(x &/| y)4895// into:4896//   z = ((~x) |/& (~y))4897// iff both x and y are free to invert and all uses of z can be freely updated.4898bool InstCombinerImpl::sinkNotIntoLogicalOp(Instruction &I) {4899  Value *Op0, *Op1;4900  if (!match(&I, m_LogicalOp(m_Value(Op0), m_Value(Op1))))4901    return false;4902 4903  // If this logic op has not been simplified yet, just bail out and let that4904  // happen first. Otherwise, the code below may wrongly invert.4905  if (Op0 == Op1)4906    return false;4907 4908  // If one of the operands is a user of the other,4909  // freelyInvert->freelyInvertAllUsersOf will change the operands of I, which4910  // may cause miscompilation.4911  if (match(Op0, m_Not(m_Specific(Op1))) || match(Op1, m_Not(m_Specific(Op0))))4912    return false;4913 4914  Instruction::BinaryOps NewOpc =4915      match(&I, m_LogicalAnd()) ? Instruction::Or : Instruction::And;4916  bool IsBinaryOp = isa<BinaryOperator>(I);4917 4918  // Can our users be adapted?4919  if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))4920    return false;4921 4922  // And can the operands be adapted?4923  if (!canFreelyInvert(*this, Op0, &I) || !canFreelyInvert(*this, Op1, &I))4924    return false;4925 4926  Op0 = freelyInvert(*this, Op0, &I);4927  Op1 = freelyInvert(*this, Op1, &I);4928 4929  Builder.SetInsertPoint(*I.getInsertionPointAfterDef());4930  Value *NewLogicOp;4931  if (IsBinaryOp)4932    NewLogicOp = Builder.CreateBinOp(NewOpc, Op0, Op1, I.getName() + ".not");4933  else4934    NewLogicOp =4935        Builder.CreateLogicalOp(NewOpc, Op0, Op1, I.getName() + ".not");4936 4937  replaceInstUsesWith(I, NewLogicOp);4938  // We can not just create an outer `not`, it will most likely be immediately4939  // folded back, reconstructing our initial pattern, and causing an4940  // infinite combine loop, so immediately manually fold it away.4941  freelyInvertAllUsersOf(NewLogicOp);4942  return true;4943}4944 4945// Transform4946//   z = (~x) &/| y4947// into:4948//   z = ~(x |/& (~y))4949// iff y is free to invert and all uses of z can be freely updated.4950bool InstCombinerImpl::sinkNotIntoOtherHandOfLogicalOp(Instruction &I) {4951  Value *Op0, *Op1;4952  if (!match(&I, m_LogicalOp(m_Value(Op0), m_Value(Op1))))4953    return false;4954  Instruction::BinaryOps NewOpc =4955      match(&I, m_LogicalAnd()) ? Instruction::Or : Instruction::And;4956  bool IsBinaryOp = isa<BinaryOperator>(I);4957 4958  Value *NotOp0 = nullptr;4959  Value *NotOp1 = nullptr;4960  Value **OpToInvert = nullptr;4961  if (match(Op0, m_Not(m_Value(NotOp0))) && canFreelyInvert(*this, Op1, &I)) {4962    Op0 = NotOp0;4963    OpToInvert = &Op1;4964  } else if (match(Op1, m_Not(m_Value(NotOp1))) &&4965             canFreelyInvert(*this, Op0, &I)) {4966    Op1 = NotOp1;4967    OpToInvert = &Op0;4968  } else4969    return false;4970 4971  // And can our users be adapted?4972  if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))4973    return false;4974 4975  *OpToInvert = freelyInvert(*this, *OpToInvert, &I);4976 4977  Builder.SetInsertPoint(*I.getInsertionPointAfterDef());4978  Value *NewBinOp;4979  if (IsBinaryOp)4980    NewBinOp = Builder.CreateBinOp(NewOpc, Op0, Op1, I.getName() + ".not");4981  else4982    NewBinOp = Builder.CreateLogicalOp(NewOpc, Op0, Op1, I.getName() + ".not");4983  replaceInstUsesWith(I, NewBinOp);4984  // We can not just create an outer `not`, it will most likely be immediately4985  // folded back, reconstructing our initial pattern, and causing an4986  // infinite combine loop, so immediately manually fold it away.4987  freelyInvertAllUsersOf(NewBinOp);4988  return true;4989}4990 4991Instruction *InstCombinerImpl::foldNot(BinaryOperator &I) {4992  Value *NotOp;4993  if (!match(&I, m_Not(m_Value(NotOp))))4994    return nullptr;4995 4996  // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand.4997  // We must eliminate the and/or (one-use) for these transforms to not increase4998  // the instruction count.4999  //5000  // ~(~X & Y) --> (X | ~Y)5001  // ~(Y & ~X) --> (X | ~Y)5002  //5003  // Note: The logical matches do not check for the commuted patterns because5004  //       those are handled via SimplifySelectsFeedingBinaryOp().5005  Type *Ty = I.getType();5006  Value *X, *Y;5007  if (match(NotOp, m_OneUse(m_c_And(m_Not(m_Value(X)), m_Value(Y))))) {5008    Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");5009    return BinaryOperator::CreateOr(X, NotY);5010  }5011  if (match(NotOp, m_OneUse(m_LogicalAnd(m_Not(m_Value(X)), m_Value(Y))))) {5012    Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");5013    return SelectInst::Create(X, ConstantInt::getTrue(Ty), NotY);5014  }5015 5016  // ~(~X | Y) --> (X & ~Y)5017  // ~(Y | ~X) --> (X & ~Y)5018  if (match(NotOp, m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y))))) {5019    Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");5020    return BinaryOperator::CreateAnd(X, NotY);5021  }5022  if (match(NotOp, m_OneUse(m_LogicalOr(m_Not(m_Value(X)), m_Value(Y))))) {5023    Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");5024    return SelectInst::Create(X, NotY, ConstantInt::getFalse(Ty));5025  }5026 5027  // Is this a 'not' (~) fed by a binary operator?5028  BinaryOperator *NotVal;5029  if (match(NotOp, m_BinOp(NotVal))) {5030    // ~((-X) | Y) --> (X - 1) & (~Y)5031    if (match(NotVal,5032              m_OneUse(m_c_Or(m_OneUse(m_Neg(m_Value(X))), m_Value(Y))))) {5033      Value *DecX = Builder.CreateAdd(X, ConstantInt::getAllOnesValue(Ty));5034      Value *NotY = Builder.CreateNot(Y);5035      return BinaryOperator::CreateAnd(DecX, NotY);5036    }5037 5038    // ~(~X >>s Y) --> (X >>s Y)5039    if (match(NotVal, m_AShr(m_Not(m_Value(X)), m_Value(Y))))5040      return BinaryOperator::CreateAShr(X, Y);5041 5042    // Treat lshr with non-negative operand as ashr.5043    // ~(~X >>u Y) --> (X >>s Y) iff X is known negative5044    if (match(NotVal, m_LShr(m_Not(m_Value(X)), m_Value(Y))) &&5045        isKnownNegative(X, SQ.getWithInstruction(NotVal)))5046      return BinaryOperator::CreateAShr(X, Y);5047 5048    // Bit-hack form of a signbit test for iN type:5049    // ~(X >>s (N - 1)) --> sext i1 (X > -1) to iN5050    unsigned FullShift = Ty->getScalarSizeInBits() - 1;5051    if (match(NotVal, m_OneUse(m_AShr(m_Value(X), m_SpecificInt(FullShift))))) {5052      Value *IsNotNeg = Builder.CreateIsNotNeg(X, "isnotneg");5053      return new SExtInst(IsNotNeg, Ty);5054    }5055 5056    // If we are inverting a right-shifted constant, we may be able to eliminate5057    // the 'not' by inverting the constant and using the opposite shift type.5058    // Canonicalization rules ensure that only a negative constant uses 'ashr',5059    // but we must check that in case that transform has not fired yet.5060 5061    // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits)5062    Constant *C;5063    if (match(NotVal, m_AShr(m_Constant(C), m_Value(Y))) &&5064        match(C, m_Negative()))5065      return BinaryOperator::CreateLShr(ConstantExpr::getNot(C), Y);5066 5067    // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits)5068    if (match(NotVal, m_LShr(m_Constant(C), m_Value(Y))) &&5069        match(C, m_NonNegative()))5070      return BinaryOperator::CreateAShr(ConstantExpr::getNot(C), Y);5071 5072    // ~(X + C) --> ~C - X5073    if (match(NotVal, m_Add(m_Value(X), m_ImmConstant(C))))5074      return BinaryOperator::CreateSub(ConstantExpr::getNot(C), X);5075 5076    // ~(X - Y) --> ~X + Y5077    // FIXME: is it really beneficial to sink the `not` here?5078    if (match(NotVal, m_Sub(m_Value(X), m_Value(Y))))5079      if (isa<Constant>(X) || NotVal->hasOneUse())5080        return BinaryOperator::CreateAdd(Builder.CreateNot(X), Y);5081 5082    // ~(~X + Y) --> X - Y5083    if (match(NotVal, m_c_Add(m_Not(m_Value(X)), m_Value(Y))))5084      return BinaryOperator::CreateWithCopiedFlags(Instruction::Sub, X, Y,5085                                                   NotVal);5086  }5087 5088  // not (cmp A, B) = !cmp A, B5089  CmpPredicate Pred;5090  if (match(NotOp, m_Cmp(Pred, m_Value(), m_Value())) &&5091      (NotOp->hasOneUse() ||5092       InstCombiner::canFreelyInvertAllUsersOf(cast<Instruction>(NotOp),5093                                               /*IgnoredUser=*/nullptr))) {5094    cast<CmpInst>(NotOp)->setPredicate(CmpInst::getInversePredicate(Pred));5095    freelyInvertAllUsersOf(NotOp);5096    return &I;5097  }5098 5099  // not (bitcast (cmp A, B) --> bitcast (!cmp A, B)5100  if (match(NotOp, m_OneUse(m_BitCast(m_Value(X)))) &&5101      match(X, m_OneUse(m_Cmp(Pred, m_Value(), m_Value())))) {5102    cast<CmpInst>(X)->setPredicate(CmpInst::getInversePredicate(Pred));5103    return new BitCastInst(X, Ty);5104  }5105 5106  // Move a 'not' ahead of casts of a bool to enable logic reduction:5107  // not (bitcast (sext i1 X)) --> bitcast (sext (not i1 X))5108  if (match(NotOp, m_OneUse(m_BitCast(m_OneUse(m_SExt(m_Value(X)))))) &&5109      X->getType()->isIntOrIntVectorTy(1)) {5110    Type *SextTy = cast<BitCastOperator>(NotOp)->getSrcTy();5111    Value *NotX = Builder.CreateNot(X);5112    Value *Sext = Builder.CreateSExt(NotX, SextTy);5113    return new BitCastInst(Sext, Ty);5114  }5115 5116  if (auto *NotOpI = dyn_cast<Instruction>(NotOp))5117    if (sinkNotIntoLogicalOp(*NotOpI))5118      return &I;5119 5120  // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max:5121  // ~min(~X, ~Y) --> max(X, Y)5122  // ~max(~X, Y) --> min(X, ~Y)5123  auto *II = dyn_cast<IntrinsicInst>(NotOp);5124  if (II && II->hasOneUse()) {5125    if (match(NotOp, m_c_MaxOrMin(m_Not(m_Value(X)), m_Value(Y)))) {5126      Intrinsic::ID InvID = getInverseMinMaxIntrinsic(II->getIntrinsicID());5127      Value *NotY = Builder.CreateNot(Y);5128      Value *InvMaxMin = Builder.CreateBinaryIntrinsic(InvID, X, NotY);5129      return replaceInstUsesWith(I, InvMaxMin);5130    }5131 5132    if (II->getIntrinsicID() == Intrinsic::is_fpclass) {5133      ConstantInt *ClassMask = cast<ConstantInt>(II->getArgOperand(1));5134      II->setArgOperand(5135          1, ConstantInt::get(ClassMask->getType(),5136                              ~ClassMask->getZExtValue() & fcAllFlags));5137      return replaceInstUsesWith(I, II);5138    }5139  }5140 5141  if (NotOp->hasOneUse()) {5142    // Pull 'not' into operands of select if both operands are one-use compares5143    // or one is one-use compare and the other one is a constant.5144    // Inverting the predicates eliminates the 'not' operation.5145    // Example:5146    //   not (select ?, (cmp TPred, ?, ?), (cmp FPred, ?, ?) -->5147    //     select ?, (cmp InvTPred, ?, ?), (cmp InvFPred, ?, ?)5148    //   not (select ?, (cmp TPred, ?, ?), true -->5149    //     select ?, (cmp InvTPred, ?, ?), false5150    if (auto *Sel = dyn_cast<SelectInst>(NotOp)) {5151      Value *TV = Sel->getTrueValue();5152      Value *FV = Sel->getFalseValue();5153      auto *CmpT = dyn_cast<CmpInst>(TV);5154      auto *CmpF = dyn_cast<CmpInst>(FV);5155      bool InvertibleT = (CmpT && CmpT->hasOneUse()) || isa<Constant>(TV);5156      bool InvertibleF = (CmpF && CmpF->hasOneUse()) || isa<Constant>(FV);5157      if (InvertibleT && InvertibleF) {5158        if (CmpT)5159          CmpT->setPredicate(CmpT->getInversePredicate());5160        else5161          Sel->setTrueValue(ConstantExpr::getNot(cast<Constant>(TV)));5162        if (CmpF)5163          CmpF->setPredicate(CmpF->getInversePredicate());5164        else5165          Sel->setFalseValue(ConstantExpr::getNot(cast<Constant>(FV)));5166        return replaceInstUsesWith(I, Sel);5167      }5168    }5169  }5170 5171  if (Instruction *NewXor = foldNotXor(I, Builder))5172    return NewXor;5173 5174  // TODO: Could handle multi-use better by checking if all uses of NotOp (other5175  // than I) can be inverted.5176  if (Value *R = getFreelyInverted(NotOp, NotOp->hasOneUse(), &Builder))5177    return replaceInstUsesWith(I, R);5178 5179  return nullptr;5180}5181 5182// FIXME: We use commutative matchers (m_c_*) for some, but not all, matches5183// here. We should standardize that construct where it is needed or choose some5184// other way to ensure that commutated variants of patterns are not missed.5185Instruction *InstCombinerImpl::visitXor(BinaryOperator &I) {5186  if (Value *V = simplifyXorInst(I.getOperand(0), I.getOperand(1),5187                                 SQ.getWithInstruction(&I)))5188    return replaceInstUsesWith(I, V);5189 5190  if (SimplifyAssociativeOrCommutative(I))5191    return &I;5192 5193  if (Instruction *X = foldVectorBinop(I))5194    return X;5195 5196  if (Instruction *Phi = foldBinopWithPhiOperands(I))5197    return Phi;5198 5199  if (Instruction *NewXor = foldXorToXor(I, Builder))5200    return NewXor;5201 5202  // (A&B)^(A&C) -> A&(B^C) etc5203  if (Value *V = foldUsingDistributiveLaws(I))5204    return replaceInstUsesWith(I, V);5205 5206  // See if we can simplify any instructions used by the instruction whose sole5207  // purpose is to compute bits we don't care about.5208  if (SimplifyDemandedInstructionBits(I))5209    return &I;5210 5211  if (Instruction *R = foldNot(I))5212    return R;5213 5214  if (Instruction *R = foldBinOpShiftWithShift(I))5215    return R;5216 5217  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);5218  Value *X, *Y, *M;5219 5220  // (X | Y) ^ M -> (X ^ M) ^ Y5221  // (X | Y) ^ M -> (Y ^ M) ^ X5222  if (match(&I, m_c_Xor(m_OneUse(m_DisjointOr(m_Value(X), m_Value(Y))),5223                        m_Value(M)))) {5224    if (Value *XorAC = simplifyXorInst(X, M, SQ.getWithInstruction(&I)))5225      return BinaryOperator::CreateXor(XorAC, Y);5226 5227    if (Value *XorBC = simplifyXorInst(Y, M, SQ.getWithInstruction(&I)))5228      return BinaryOperator::CreateXor(XorBC, X);5229  }5230 5231  // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M)5232  // This it a special case in haveNoCommonBitsSet, but the computeKnownBits5233  // calls in there are unnecessary as SimplifyDemandedInstructionBits should5234  // have already taken care of those cases.5235  if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(M)), m_Value()),5236                        m_c_And(m_Deferred(M), m_Value())))) {5237    if (isGuaranteedNotToBeUndef(M))5238      return BinaryOperator::CreateDisjointOr(Op0, Op1);5239    else5240      return BinaryOperator::CreateOr(Op0, Op1);5241  }5242 5243  if (Instruction *Xor = visitMaskedMerge(I, Builder))5244    return Xor;5245 5246  Constant *C1;5247  if (match(Op1, m_Constant(C1))) {5248    Constant *C2;5249 5250    if (match(Op0, m_OneUse(m_Or(m_Value(X), m_ImmConstant(C2)))) &&5251        match(C1, m_ImmConstant())) {5252      // (X | C2) ^ C1 --> (X & ~C2) ^ (C1^C2)5253      C2 = Constant::replaceUndefsWith(5254          C2, Constant::getAllOnesValue(C2->getType()->getScalarType()));5255      Value *And = Builder.CreateAnd(5256          X, Constant::mergeUndefsWith(ConstantExpr::getNot(C2), C1));5257      return BinaryOperator::CreateXor(5258          And, Constant::mergeUndefsWith(ConstantExpr::getXor(C1, C2), C1));5259    }5260 5261    // Use DeMorgan and reassociation to eliminate a 'not' op.5262    if (match(Op0, m_OneUse(m_Or(m_Not(m_Value(X)), m_Constant(C2))))) {5263      // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C15264      Value *And = Builder.CreateAnd(X, ConstantExpr::getNot(C2));5265      return BinaryOperator::CreateXor(And, ConstantExpr::getNot(C1));5266    }5267    if (match(Op0, m_OneUse(m_And(m_Not(m_Value(X)), m_Constant(C2))))) {5268      // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C15269      Value *Or = Builder.CreateOr(X, ConstantExpr::getNot(C2));5270      return BinaryOperator::CreateXor(Or, ConstantExpr::getNot(C1));5271    }5272 5273    // Convert xor ([trunc] (ashr X, BW-1)), C =>5274    //   select(X >s -1, C, ~C)5275    // The ashr creates "AllZeroOrAllOne's", which then optionally inverses the5276    // constant depending on whether this input is less than 0.5277    const APInt *CA;5278    if (match(Op0, m_OneUse(m_TruncOrSelf(5279                       m_AShr(m_Value(X), m_APIntAllowPoison(CA))))) &&5280        *CA == X->getType()->getScalarSizeInBits() - 1 &&5281        !match(C1, m_AllOnes())) {5282      assert(!C1->isZeroValue() && "Unexpected xor with 0");5283      Value *IsNotNeg = Builder.CreateIsNotNeg(X);5284      return SelectInst::Create(IsNotNeg, Op1, Builder.CreateNot(Op1));5285    }5286  }5287 5288  Type *Ty = I.getType();5289  {5290    const APInt *RHSC;5291    if (match(Op1, m_APInt(RHSC))) {5292      Value *X;5293      const APInt *C;5294      // (C - X) ^ signmaskC --> (C + signmaskC) - X5295      if (RHSC->isSignMask() && match(Op0, m_Sub(m_APInt(C), m_Value(X))))5296        return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C + *RHSC), X);5297 5298      // (X + C) ^ signmaskC --> X + (C + signmaskC)5299      if (RHSC->isSignMask() && match(Op0, m_Add(m_Value(X), m_APInt(C))))5300        return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C + *RHSC));5301 5302      // (X | C) ^ RHSC --> X ^ (C ^ RHSC) iff X & C == 05303      if (match(Op0, m_Or(m_Value(X), m_APInt(C))) &&5304          MaskedValueIsZero(X, *C, &I))5305        return BinaryOperator::CreateXor(X, ConstantInt::get(Ty, *C ^ *RHSC));5306 5307      // When X is a power-of-two or zero and zero input is poison:5308      // ctlz(i32 X) ^ 31 --> cttz(X)5309      // cttz(i32 X) ^ 31 --> ctlz(X)5310      auto *II = dyn_cast<IntrinsicInst>(Op0);5311      if (II && II->hasOneUse() && *RHSC == Ty->getScalarSizeInBits() - 1) {5312        Intrinsic::ID IID = II->getIntrinsicID();5313        if ((IID == Intrinsic::ctlz || IID == Intrinsic::cttz) &&5314            match(II->getArgOperand(1), m_One()) &&5315            isKnownToBeAPowerOfTwo(II->getArgOperand(0), /*OrZero */ true)) {5316          IID = (IID == Intrinsic::ctlz) ? Intrinsic::cttz : Intrinsic::ctlz;5317          Function *F =5318              Intrinsic::getOrInsertDeclaration(II->getModule(), IID, Ty);5319          return CallInst::Create(F, {II->getArgOperand(0), Builder.getTrue()});5320        }5321      }5322 5323      // If RHSC is inverting the remaining bits of shifted X,5324      // canonicalize to a 'not' before the shift to help SCEV and codegen:5325      // (X << C) ^ RHSC --> ~X << C5326      if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_APInt(C)))) &&5327          *RHSC == APInt::getAllOnes(Ty->getScalarSizeInBits()).shl(*C)) {5328        Value *NotX = Builder.CreateNot(X);5329        return BinaryOperator::CreateShl(NotX, ConstantInt::get(Ty, *C));5330      }5331      // (X >>u C) ^ RHSC --> ~X >>u C5332      if (match(Op0, m_OneUse(m_LShr(m_Value(X), m_APInt(C)))) &&5333          *RHSC == APInt::getAllOnes(Ty->getScalarSizeInBits()).lshr(*C)) {5334        Value *NotX = Builder.CreateNot(X);5335        return BinaryOperator::CreateLShr(NotX, ConstantInt::get(Ty, *C));5336      }5337      // TODO: We could handle 'ashr' here as well. That would be matching5338      //       a 'not' op and moving it before the shift. Doing that requires5339      //       preventing the inverse fold in canShiftBinOpWithConstantRHS().5340    }5341 5342    // If we are XORing the sign bit of a floating-point value, convert5343    // this to fneg, then cast back to integer.5344    //5345    // This is generous interpretation of noimplicitfloat, this is not a true5346    // floating-point operation.5347    //5348    // Assumes any IEEE-represented type has the sign bit in the high bit.5349    // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt5350    Value *CastOp;5351    if (match(Op0, m_ElementWiseBitCast(m_Value(CastOp))) &&5352        match(Op1, m_SignMask()) &&5353        !Builder.GetInsertBlock()->getParent()->hasFnAttribute(5354            Attribute::NoImplicitFloat)) {5355      Type *EltTy = CastOp->getType()->getScalarType();5356      if (EltTy->isFloatingPointTy() &&5357          APFloat::hasSignBitInMSB(EltTy->getFltSemantics())) {5358        Value *FNeg = Builder.CreateFNeg(CastOp);5359        return new BitCastInst(FNeg, I.getType());5360      }5361    }5362  }5363 5364  // FIXME: This should not be limited to scalar (pull into APInt match above).5365  {5366    Value *X;5367    ConstantInt *C1, *C2, *C3;5368    // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)5369    if (match(Op1, m_ConstantInt(C3)) &&5370        match(Op0, m_LShr(m_Xor(m_Value(X), m_ConstantInt(C1)),5371                          m_ConstantInt(C2))) &&5372        Op0->hasOneUse()) {5373      // fold (C1 >> C2) ^ C35374      APInt FoldConst = C1->getValue().lshr(C2->getValue());5375      FoldConst ^= C3->getValue();5376      // Prepare the two operands.5377      auto *Opnd0 = Builder.CreateLShr(X, C2);5378      Opnd0->takeName(Op0);5379      return BinaryOperator::CreateXor(Opnd0, ConstantInt::get(Ty, FoldConst));5380    }5381  }5382 5383  if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))5384    return FoldedLogic;5385 5386  // Y ^ (X | Y) --> X & ~Y5387  // Y ^ (Y | X) --> X & ~Y5388  if (match(Op1, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op0)))))5389    return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op0));5390  // (X | Y) ^ Y --> X & ~Y5391  // (Y | X) ^ Y --> X & ~Y5392  if (match(Op0, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op1)))))5393    return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op1));5394 5395  // Y ^ (X & Y) --> ~X & Y5396  // Y ^ (Y & X) --> ~X & Y5397  if (match(Op1, m_OneUse(m_c_And(m_Value(X), m_Specific(Op0)))))5398    return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(X));5399  // (X & Y) ^ Y --> ~X & Y5400  // (Y & X) ^ Y --> ~X & Y5401  // Canonical form is (X & C) ^ C; don't touch that.5402  // TODO: A 'not' op is better for analysis and codegen, but demanded bits must5403  //       be fixed to prefer that (otherwise we get infinite looping).5404  if (!match(Op1, m_Constant()) &&5405      match(Op0, m_OneUse(m_c_And(m_Value(X), m_Specific(Op1)))))5406    return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(X));5407 5408  Value *A, *B, *C;5409  // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants.5410  if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),5411                        m_OneUse(m_c_Or(m_Deferred(A), m_Value(C))))))5412      return BinaryOperator::CreateXor(5413          Builder.CreateAnd(Builder.CreateNot(A), C), B);5414 5415  // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants.5416  if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),5417                        m_OneUse(m_c_Or(m_Deferred(B), m_Value(C))))))5418      return BinaryOperator::CreateXor(5419          Builder.CreateAnd(Builder.CreateNot(B), C), A);5420 5421  // (A & B) ^ (A ^ B) -> (A | B)5422  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&5423      match(Op1, m_c_Xor(m_Specific(A), m_Specific(B))))5424    return BinaryOperator::CreateOr(A, B);5425  // (A ^ B) ^ (A & B) -> (A | B)5426  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&5427      match(Op1, m_c_And(m_Specific(A), m_Specific(B))))5428    return BinaryOperator::CreateOr(A, B);5429 5430  // (A & ~B) ^ ~A -> ~(A & B)5431  // (~B & A) ^ ~A -> ~(A & B)5432  if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&5433      match(Op1, m_Not(m_Specific(A))))5434    return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));5435 5436  // (~A & B) ^ A --> A | B -- There are 4 commuted variants.5437  if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(A)), m_Value(B)), m_Deferred(A))))5438    return BinaryOperator::CreateOr(A, B);5439 5440  // (~A | B) ^ A --> ~(A & B)5441  if (match(Op0, m_OneUse(m_c_Or(m_Not(m_Specific(Op1)), m_Value(B)))))5442    return BinaryOperator::CreateNot(Builder.CreateAnd(Op1, B));5443 5444  // A ^ (~A | B) --> ~(A & B)5445  if (match(Op1, m_OneUse(m_c_Or(m_Not(m_Specific(Op0)), m_Value(B)))))5446    return BinaryOperator::CreateNot(Builder.CreateAnd(Op0, B));5447 5448  // (A | B) ^ (A | C) --> (B ^ C) & ~A -- There are 4 commuted variants.5449  // TODO: Loosen one-use restriction if common operand is a constant.5450  Value *D;5451  if (match(Op0, m_OneUse(m_Or(m_Value(A), m_Value(B)))) &&5452      match(Op1, m_OneUse(m_Or(m_Value(C), m_Value(D))))) {5453    if (B == C || B == D)5454      std::swap(A, B);5455    if (A == C)5456      std::swap(C, D);5457    if (A == D) {5458      Value *NotA = Builder.CreateNot(A);5459      return BinaryOperator::CreateAnd(Builder.CreateXor(B, C), NotA);5460    }5461  }5462 5463  // (A & B) ^ (A | C) --> A ? ~B : C -- There are 4 commuted variants.5464  if (I.getType()->isIntOrIntVectorTy(1) &&5465      match(&I, m_c_Xor(m_OneUse(m_LogicalAnd(m_Value(A), m_Value(B))),5466                        m_OneUse(m_LogicalOr(m_Value(C), m_Value(D)))))) {5467    bool NeedFreeze = isa<SelectInst>(Op0) && isa<SelectInst>(Op1) && B == D;5468    if (B == C || B == D)5469      std::swap(A, B);5470    if (A == C)5471      std::swap(C, D);5472    if (A == D) {5473      if (NeedFreeze)5474        A = Builder.CreateFreeze(A);5475      Value *NotB = Builder.CreateNot(B);5476      return SelectInst::Create(A, NotB, C);5477    }5478  }5479 5480  if (auto *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))5481    if (auto *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))5482      if (Value *V = foldXorOfICmps(LHS, RHS, I))5483        return replaceInstUsesWith(I, V);5484 5485  if (Instruction *CastedXor = foldCastedBitwiseLogic(I))5486    return CastedXor;5487 5488  if (Instruction *Abs = canonicalizeAbs(I, Builder))5489    return Abs;5490 5491  // Otherwise, if all else failed, try to hoist the xor-by-constant:5492  //   (X ^ C) ^ Y --> (X ^ Y) ^ C5493  // Just like we do in other places, we completely avoid the fold5494  // for constantexprs, at least to avoid endless combine loop.5495  if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(X, m_Unless(m_ConstantExpr())),5496                                       m_ImmConstant(C1))),5497                        m_Value(Y))))5498    return BinaryOperator::CreateXor(Builder.CreateXor(X, Y), C1);5499 5500  if (Instruction *R = reassociateForUses(I, Builder))5501    return R;5502 5503  if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))5504    return Canonicalized;5505 5506  if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1))5507    return Folded;5508 5509  if (Instruction *Folded = canonicalizeConditionalNegationViaMathToSelect(I))5510    return Folded;5511 5512  if (Instruction *Res = foldBinOpOfDisplacedShifts(I))5513    return Res;5514 5515  if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))5516    return Res;5517 5518  return nullptr;5519}5520