LLVM API Documentation

InstCombineAndOrXor.cpp
Go to the documentation of this file.
00001 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This file implements the visitAnd, visitOr, and visitXor functions.
00011 //
00012 //===----------------------------------------------------------------------===//
00013 
00014 #include "InstCombine.h"
00015 #include "llvm/Analysis/InstructionSimplify.h"
00016 #include "llvm/IR/ConstantRange.h"
00017 #include "llvm/IR/Intrinsics.h"
00018 #include "llvm/IR/PatternMatch.h"
00019 #include "llvm/Transforms/Utils/CmpInstAnalysis.h"
00020 using namespace llvm;
00021 using namespace PatternMatch;
00022 
00023 #define DEBUG_TYPE "instcombine"
00024 
00025 /// isFreeToInvert - Return true if the specified value is free to invert (apply
00026 /// ~ to).  This happens in cases where the ~ can be eliminated.
00027 static inline bool isFreeToInvert(Value *V) {
00028   // ~(~(X)) -> X.
00029   if (BinaryOperator::isNot(V))
00030     return true;
00031 
00032   // Constants can be considered to be not'ed values.
00033   if (isa<ConstantInt>(V))
00034     return true;
00035 
00036   // Compares can be inverted if they have a single use.
00037   if (CmpInst *CI = dyn_cast<CmpInst>(V))
00038     return CI->hasOneUse();
00039 
00040   return false;
00041 }
00042 
00043 static inline Value *dyn_castNotVal(Value *V) {
00044   // If this is not(not(x)) don't return that this is a not: we want the two
00045   // not's to be folded first.
00046   if (BinaryOperator::isNot(V)) {
00047     Value *Operand = BinaryOperator::getNotArgument(V);
00048     if (!isFreeToInvert(Operand))
00049       return Operand;
00050   }
00051 
00052   // Constants can be considered to be not'ed values...
00053   if (ConstantInt *C = dyn_cast<ConstantInt>(V))
00054     return ConstantInt::get(C->getType(), ~C->getValue());
00055   return nullptr;
00056 }
00057 
00058 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
00059 /// predicate into a three bit mask. It also returns whether it is an ordered
00060 /// predicate by reference.
00061 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
00062   isOrdered = false;
00063   switch (CC) {
00064   case FCmpInst::FCMP_ORD: isOrdered = true; return 0;  // 000
00065   case FCmpInst::FCMP_UNO:                   return 0;  // 000
00066   case FCmpInst::FCMP_OGT: isOrdered = true; return 1;  // 001
00067   case FCmpInst::FCMP_UGT:                   return 1;  // 001
00068   case FCmpInst::FCMP_OEQ: isOrdered = true; return 2;  // 010
00069   case FCmpInst::FCMP_UEQ:                   return 2;  // 010
00070   case FCmpInst::FCMP_OGE: isOrdered = true; return 3;  // 011
00071   case FCmpInst::FCMP_UGE:                   return 3;  // 011
00072   case FCmpInst::FCMP_OLT: isOrdered = true; return 4;  // 100
00073   case FCmpInst::FCMP_ULT:                   return 4;  // 100
00074   case FCmpInst::FCMP_ONE: isOrdered = true; return 5;  // 101
00075   case FCmpInst::FCMP_UNE:                   return 5;  // 101
00076   case FCmpInst::FCMP_OLE: isOrdered = true; return 6;  // 110
00077   case FCmpInst::FCMP_ULE:                   return 6;  // 110
00078     // True -> 7
00079   default:
00080     // Not expecting FCMP_FALSE and FCMP_TRUE;
00081     llvm_unreachable("Unexpected FCmp predicate!");
00082   }
00083 }
00084 
00085 /// getNewICmpValue - This is the complement of getICmpCode, which turns an
00086 /// opcode and two operands into either a constant true or false, or a brand
00087 /// new ICmp instruction. The sign is passed in to determine which kind
00088 /// of predicate to use in the new icmp instruction.
00089 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
00090                               InstCombiner::BuilderTy *Builder) {
00091   ICmpInst::Predicate NewPred;
00092   if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
00093     return NewConstant;
00094   return Builder->CreateICmp(NewPred, LHS, RHS);
00095 }
00096 
00097 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
00098 /// opcode and two operands into either a FCmp instruction. isordered is passed
00099 /// in to determine which kind of predicate to use in the new fcmp instruction.
00100 static Value *getFCmpValue(bool isordered, unsigned code,
00101                            Value *LHS, Value *RHS,
00102                            InstCombiner::BuilderTy *Builder) {
00103   CmpInst::Predicate Pred;
00104   switch (code) {
00105   default: llvm_unreachable("Illegal FCmp code!");
00106   case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
00107   case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
00108   case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
00109   case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
00110   case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
00111   case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
00112   case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
00113   case 7:
00114     if (!isordered) return ConstantInt::getTrue(LHS->getContext());
00115     Pred = FCmpInst::FCMP_ORD; break;
00116   }
00117   return Builder->CreateFCmp(Pred, LHS, RHS);
00118 }
00119 
00120 // OptAndOp - This handles expressions of the form ((val OP C1) & C2).  Where
00121 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'.  Op is
00122 // guaranteed to be a binary operator.
00123 Instruction *InstCombiner::OptAndOp(Instruction *Op,
00124                                     ConstantInt *OpRHS,
00125                                     ConstantInt *AndRHS,
00126                                     BinaryOperator &TheAnd) {
00127   Value *X = Op->getOperand(0);
00128   Constant *Together = nullptr;
00129   if (!Op->isShift())
00130     Together = ConstantExpr::getAnd(AndRHS, OpRHS);
00131 
00132   switch (Op->getOpcode()) {
00133   case Instruction::Xor:
00134     if (Op->hasOneUse()) {
00135       // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
00136       Value *And = Builder->CreateAnd(X, AndRHS);
00137       And->takeName(Op);
00138       return BinaryOperator::CreateXor(And, Together);
00139     }
00140     break;
00141   case Instruction::Or:
00142     if (Op->hasOneUse()){
00143       if (Together != OpRHS) {
00144         // (X | C1) & C2 --> (X | (C1&C2)) & C2
00145         Value *Or = Builder->CreateOr(X, Together);
00146         Or->takeName(Op);
00147         return BinaryOperator::CreateAnd(Or, AndRHS);
00148       }
00149 
00150       ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
00151       if (TogetherCI && !TogetherCI->isZero()){
00152         // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
00153         // NOTE: This reduces the number of bits set in the & mask, which
00154         // can expose opportunities for store narrowing.
00155         Together = ConstantExpr::getXor(AndRHS, Together);
00156         Value *And = Builder->CreateAnd(X, Together);
00157         And->takeName(Op);
00158         return BinaryOperator::CreateOr(And, OpRHS);
00159       }
00160     }
00161 
00162     break;
00163   case Instruction::Add:
00164     if (Op->hasOneUse()) {
00165       // Adding a one to a single bit bit-field should be turned into an XOR
00166       // of the bit.  First thing to check is to see if this AND is with a
00167       // single bit constant.
00168       const APInt &AndRHSV = AndRHS->getValue();
00169 
00170       // If there is only one bit set.
00171       if (AndRHSV.isPowerOf2()) {
00172         // Ok, at this point, we know that we are masking the result of the
00173         // ADD down to exactly one bit.  If the constant we are adding has
00174         // no bits set below this bit, then we can eliminate the ADD.
00175         const APInt& AddRHS = OpRHS->getValue();
00176 
00177         // Check to see if any bits below the one bit set in AndRHSV are set.
00178         if ((AddRHS & (AndRHSV-1)) == 0) {
00179           // If not, the only thing that can effect the output of the AND is
00180           // the bit specified by AndRHSV.  If that bit is set, the effect of
00181           // the XOR is to toggle the bit.  If it is clear, then the ADD has
00182           // no effect.
00183           if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
00184             TheAnd.setOperand(0, X);
00185             return &TheAnd;
00186           } else {
00187             // Pull the XOR out of the AND.
00188             Value *NewAnd = Builder->CreateAnd(X, AndRHS);
00189             NewAnd->takeName(Op);
00190             return BinaryOperator::CreateXor(NewAnd, AndRHS);
00191           }
00192         }
00193       }
00194     }
00195     break;
00196 
00197   case Instruction::Shl: {
00198     // We know that the AND will not produce any of the bits shifted in, so if
00199     // the anded constant includes them, clear them now!
00200     //
00201     uint32_t BitWidth = AndRHS->getType()->getBitWidth();
00202     uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
00203     APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
00204     ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
00205 
00206     if (CI->getValue() == ShlMask)
00207       // Masking out bits that the shift already masks.
00208       return ReplaceInstUsesWith(TheAnd, Op);   // No need for the and.
00209 
00210     if (CI != AndRHS) {                  // Reducing bits set in and.
00211       TheAnd.setOperand(1, CI);
00212       return &TheAnd;
00213     }
00214     break;
00215   }
00216   case Instruction::LShr: {
00217     // We know that the AND will not produce any of the bits shifted in, so if
00218     // the anded constant includes them, clear them now!  This only applies to
00219     // unsigned shifts, because a signed shr may bring in set bits!
00220     //
00221     uint32_t BitWidth = AndRHS->getType()->getBitWidth();
00222     uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
00223     APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
00224     ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
00225 
00226     if (CI->getValue() == ShrMask)
00227       // Masking out bits that the shift already masks.
00228       return ReplaceInstUsesWith(TheAnd, Op);
00229 
00230     if (CI != AndRHS) {
00231       TheAnd.setOperand(1, CI);  // Reduce bits set in and cst.
00232       return &TheAnd;
00233     }
00234     break;
00235   }
00236   case Instruction::AShr:
00237     // Signed shr.
00238     // See if this is shifting in some sign extension, then masking it out
00239     // with an and.
00240     if (Op->hasOneUse()) {
00241       uint32_t BitWidth = AndRHS->getType()->getBitWidth();
00242       uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
00243       APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
00244       Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
00245       if (C == AndRHS) {          // Masking out bits shifted in.
00246         // (Val ashr C1) & C2 -> (Val lshr C1) & C2
00247         // Make the argument unsigned.
00248         Value *ShVal = Op->getOperand(0);
00249         ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
00250         return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
00251       }
00252     }
00253     break;
00254   }
00255   return nullptr;
00256 }
00257 
00258 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
00259 /// (V < Lo || V >= Hi).  In practice, we emit the more efficient
00260 /// (V-Lo) <u Hi-Lo.  This method expects that Lo <= Hi. isSigned indicates
00261 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
00262 /// insert new instructions.
00263 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
00264                                      bool isSigned, bool Inside) {
00265   assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
00266             ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
00267          "Lo is not <= Hi in range emission code!");
00268 
00269   if (Inside) {
00270     if (Lo == Hi)  // Trivially false.
00271       return Builder->getFalse();
00272 
00273     // V >= Min && V < Hi --> V < Hi
00274     if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
00275       ICmpInst::Predicate pred = (isSigned ?
00276         ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
00277       return Builder->CreateICmp(pred, V, Hi);
00278     }
00279 
00280     // Emit V-Lo <u Hi-Lo
00281     Constant *NegLo = ConstantExpr::getNeg(Lo);
00282     Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
00283     Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
00284     return Builder->CreateICmpULT(Add, UpperBound);
00285   }
00286 
00287   if (Lo == Hi)  // Trivially true.
00288     return Builder->getTrue();
00289 
00290   // V < Min || V >= Hi -> V > Hi-1
00291   Hi = SubOne(cast<ConstantInt>(Hi));
00292   if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
00293     ICmpInst::Predicate pred = (isSigned ?
00294         ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
00295     return Builder->CreateICmp(pred, V, Hi);
00296   }
00297 
00298   // Emit V-Lo >u Hi-1-Lo
00299   // Note that Hi has already had one subtracted from it, above.
00300   ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
00301   Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
00302   Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
00303   return Builder->CreateICmpUGT(Add, LowerBound);
00304 }
00305 
00306 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
00307 // any number of 0s on either side.  The 1s are allowed to wrap from LSB to
00308 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.  0x0F0F0000 is
00309 // not, since all 1s are not contiguous.
00310 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
00311   const APInt& V = Val->getValue();
00312   uint32_t BitWidth = Val->getType()->getBitWidth();
00313   if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
00314 
00315   // look for the first zero bit after the run of ones
00316   MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
00317   // look for the first non-zero bit
00318   ME = V.getActiveBits();
00319   return true;
00320 }
00321 
00322 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
00323 /// where isSub determines whether the operator is a sub.  If we can fold one of
00324 /// the following xforms:
00325 ///
00326 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
00327 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
00328 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
00329 ///
00330 /// return (A +/- B).
00331 ///
00332 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
00333                                         ConstantInt *Mask, bool isSub,
00334                                         Instruction &I) {
00335   Instruction *LHSI = dyn_cast<Instruction>(LHS);
00336   if (!LHSI || LHSI->getNumOperands() != 2 ||
00337       !isa<ConstantInt>(LHSI->getOperand(1))) return nullptr;
00338 
00339   ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
00340 
00341   switch (LHSI->getOpcode()) {
00342   default: return nullptr;
00343   case Instruction::And:
00344     if (ConstantExpr::getAnd(N, Mask) == Mask) {
00345       // If the AndRHS is a power of two minus one (0+1+), this is simple.
00346       if ((Mask->getValue().countLeadingZeros() +
00347            Mask->getValue().countPopulation()) ==
00348           Mask->getValue().getBitWidth())
00349         break;
00350 
00351       // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
00352       // part, we don't need any explicit masks to take them out of A.  If that
00353       // is all N is, ignore it.
00354       uint32_t MB = 0, ME = 0;
00355       if (isRunOfOnes(Mask, MB, ME)) {  // begin/end bit of run, inclusive
00356         uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
00357         APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
00358         if (MaskedValueIsZero(RHS, Mask))
00359           break;
00360       }
00361     }
00362     return nullptr;
00363   case Instruction::Or:
00364   case Instruction::Xor:
00365     // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
00366     if ((Mask->getValue().countLeadingZeros() +
00367          Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
00368         && ConstantExpr::getAnd(N, Mask)->isNullValue())
00369       break;
00370     return nullptr;
00371   }
00372 
00373   if (isSub)
00374     return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
00375   return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
00376 }
00377 
00378 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
00379 /// One of A and B is considered the mask, the other the value. This is
00380 /// described as the "AMask" or "BMask" part of the enum. If the enum
00381 /// contains only "Mask", then both A and B can be considered masks.
00382 /// If A is the mask, then it was proven, that (A & C) == C. This
00383 /// is trivial if C == A, or C == 0. If both A and C are constants, this
00384 /// proof is also easy.
00385 /// For the following explanations we assume that A is the mask.
00386 /// The part "AllOnes" declares, that the comparison is true only
00387 /// if (A & B) == A, or all bits of A are set in B.
00388 ///   Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
00389 /// The part "AllZeroes" declares, that the comparison is true only
00390 /// if (A & B) == 0, or all bits of A are cleared in B.
00391 ///   Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
00392 /// The part "Mixed" declares, that (A & B) == C and C might or might not
00393 /// contain any number of one bits and zero bits.
00394 ///   Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
00395 /// The Part "Not" means, that in above descriptions "==" should be replaced
00396 /// by "!=".
00397 ///   Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
00398 /// If the mask A contains a single bit, then the following is equivalent:
00399 ///    (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
00400 ///    (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
00401 enum MaskedICmpType {
00402   FoldMskICmp_AMask_AllOnes           =     1,
00403   FoldMskICmp_AMask_NotAllOnes        =     2,
00404   FoldMskICmp_BMask_AllOnes           =     4,
00405   FoldMskICmp_BMask_NotAllOnes        =     8,
00406   FoldMskICmp_Mask_AllZeroes          =    16,
00407   FoldMskICmp_Mask_NotAllZeroes       =    32,
00408   FoldMskICmp_AMask_Mixed             =    64,
00409   FoldMskICmp_AMask_NotMixed          =   128,
00410   FoldMskICmp_BMask_Mixed             =   256,
00411   FoldMskICmp_BMask_NotMixed          =   512
00412 };
00413 
00414 /// return the set of pattern classes (from MaskedICmpType)
00415 /// that (icmp SCC (A & B), C) satisfies
00416 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
00417                                     ICmpInst::Predicate SCC)
00418 {
00419   ConstantInt *ACst = dyn_cast<ConstantInt>(A);
00420   ConstantInt *BCst = dyn_cast<ConstantInt>(B);
00421   ConstantInt *CCst = dyn_cast<ConstantInt>(C);
00422   bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
00423   bool icmp_abit = (ACst && !ACst->isZero() &&
00424                     ACst->getValue().isPowerOf2());
00425   bool icmp_bbit = (BCst && !BCst->isZero() &&
00426                     BCst->getValue().isPowerOf2());
00427   unsigned result = 0;
00428   if (CCst && CCst->isZero()) {
00429     // if C is zero, then both A and B qualify as mask
00430     result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
00431                           FoldMskICmp_Mask_AllZeroes |
00432                           FoldMskICmp_AMask_Mixed |
00433                           FoldMskICmp_BMask_Mixed)
00434                        : (FoldMskICmp_Mask_NotAllZeroes |
00435                           FoldMskICmp_Mask_NotAllZeroes |
00436                           FoldMskICmp_AMask_NotMixed |
00437                           FoldMskICmp_BMask_NotMixed));
00438     if (icmp_abit)
00439       result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
00440                             FoldMskICmp_AMask_NotMixed)
00441                          : (FoldMskICmp_AMask_AllOnes |
00442                             FoldMskICmp_AMask_Mixed));
00443     if (icmp_bbit)
00444       result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
00445                             FoldMskICmp_BMask_NotMixed)
00446                          : (FoldMskICmp_BMask_AllOnes |
00447                             FoldMskICmp_BMask_Mixed));
00448     return result;
00449   }
00450   if (A == C) {
00451     result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
00452                           FoldMskICmp_AMask_Mixed)
00453                        : (FoldMskICmp_AMask_NotAllOnes |
00454                           FoldMskICmp_AMask_NotMixed));
00455     if (icmp_abit)
00456       result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
00457                             FoldMskICmp_AMask_NotMixed)
00458                          : (FoldMskICmp_Mask_AllZeroes |
00459                             FoldMskICmp_AMask_Mixed));
00460   } else if (ACst && CCst &&
00461              ConstantExpr::getAnd(ACst, CCst) == CCst) {
00462     result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
00463                        : FoldMskICmp_AMask_NotMixed);
00464   }
00465   if (B == C) {
00466     result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
00467                           FoldMskICmp_BMask_Mixed)
00468                        : (FoldMskICmp_BMask_NotAllOnes |
00469                           FoldMskICmp_BMask_NotMixed));
00470     if (icmp_bbit)
00471       result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
00472                             FoldMskICmp_BMask_NotMixed)
00473                          : (FoldMskICmp_Mask_AllZeroes |
00474                             FoldMskICmp_BMask_Mixed));
00475   } else if (BCst && CCst &&
00476              ConstantExpr::getAnd(BCst, CCst) == CCst) {
00477     result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
00478                        : FoldMskICmp_BMask_NotMixed);
00479   }
00480   return result;
00481 }
00482 
00483 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
00484 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
00485 /// is adjacent to the corresponding normal flag (recording ==), this just
00486 /// involves swapping those bits over.
00487 static unsigned conjugateICmpMask(unsigned Mask) {
00488   unsigned NewMask;
00489   NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes |
00490                      FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed |
00491                      FoldMskICmp_BMask_Mixed))
00492             << 1;
00493 
00494   NewMask |=
00495       (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes |
00496                FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed |
00497                FoldMskICmp_BMask_NotMixed))
00498       >> 1;
00499 
00500   return NewMask;
00501 }
00502 
00503 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
00504 /// if possible. The returned predicate is either == or !=. Returns false if
00505 /// decomposition fails.
00506 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
00507                                  Value *&X, Value *&Y, Value *&Z) {
00508   ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1));
00509   if (!C)
00510     return false;
00511 
00512   switch (I->getPredicate()) {
00513   default:
00514     return false;
00515   case ICmpInst::ICMP_SLT:
00516     // X < 0 is equivalent to (X & SignBit) != 0.
00517     if (!C->isZero())
00518       return false;
00519     Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
00520     Pred = ICmpInst::ICMP_NE;
00521     break;
00522   case ICmpInst::ICMP_SGT:
00523     // X > -1 is equivalent to (X & SignBit) == 0.
00524     if (!C->isAllOnesValue())
00525       return false;
00526     Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
00527     Pred = ICmpInst::ICMP_EQ;
00528     break;
00529   case ICmpInst::ICMP_ULT:
00530     // X <u 2^n is equivalent to (X & ~(2^n-1)) == 0.
00531     if (!C->getValue().isPowerOf2())
00532       return false;
00533     Y = ConstantInt::get(I->getContext(), -C->getValue());
00534     Pred = ICmpInst::ICMP_EQ;
00535     break;
00536   case ICmpInst::ICMP_UGT:
00537     // X >u 2^n-1 is equivalent to (X & ~(2^n-1)) != 0.
00538     if (!(C->getValue() + 1).isPowerOf2())
00539       return false;
00540     Y = ConstantInt::get(I->getContext(), ~C->getValue());
00541     Pred = ICmpInst::ICMP_NE;
00542     break;
00543   }
00544 
00545   X = I->getOperand(0);
00546   Z = ConstantInt::getNullValue(C->getType());
00547   return true;
00548 }
00549 
00550 /// foldLogOpOfMaskedICmpsHelper:
00551 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
00552 /// return the set of pattern classes (from MaskedICmpType)
00553 /// that both LHS and RHS satisfy
00554 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
00555                                              Value*& B, Value*& C,
00556                                              Value*& D, Value*& E,
00557                                              ICmpInst *LHS, ICmpInst *RHS,
00558                                              ICmpInst::Predicate &LHSCC,
00559                                              ICmpInst::Predicate &RHSCC) {
00560   if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
00561   // vectors are not (yet?) supported
00562   if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
00563 
00564   // Here comes the tricky part:
00565   // LHS might be of the form L11 & L12 == X, X == L21 & L22,
00566   // and L11 & L12 == L21 & L22. The same goes for RHS.
00567   // Now we must find those components L** and R**, that are equal, so
00568   // that we can extract the parameters A, B, C, D, and E for the canonical
00569   // above.
00570   Value *L1 = LHS->getOperand(0);
00571   Value *L2 = LHS->getOperand(1);
00572   Value *L11,*L12,*L21,*L22;
00573   // Check whether the icmp can be decomposed into a bit test.
00574   if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
00575     L21 = L22 = L1 = nullptr;
00576   } else {
00577     // Look for ANDs in the LHS icmp.
00578     if (!L1->getType()->isIntegerTy()) {
00579       // You can icmp pointers, for example. They really aren't masks.
00580       L11 = L12 = nullptr;
00581     } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
00582       // Any icmp can be viewed as being trivially masked; if it allows us to
00583       // remove one, it's worth it.
00584       L11 = L1;
00585       L12 = Constant::getAllOnesValue(L1->getType());
00586     }
00587 
00588     if (!L2->getType()->isIntegerTy()) {
00589       // You can icmp pointers, for example. They really aren't masks.
00590       L21 = L22 = nullptr;
00591     } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
00592       L21 = L2;
00593       L22 = Constant::getAllOnesValue(L2->getType());
00594     }
00595   }
00596 
00597   // Bail if LHS was a icmp that can't be decomposed into an equality.
00598   if (!ICmpInst::isEquality(LHSCC))
00599     return 0;
00600 
00601   Value *R1 = RHS->getOperand(0);
00602   Value *R2 = RHS->getOperand(1);
00603   Value *R11,*R12;
00604   bool ok = false;
00605   if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
00606     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
00607       A = R11; D = R12;
00608     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
00609       A = R12; D = R11;
00610     } else {
00611       return 0;
00612     }
00613     E = R2; R1 = nullptr; ok = true;
00614   } else if (R1->getType()->isIntegerTy()) {
00615     if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
00616       // As before, model no mask as a trivial mask if it'll let us do an
00617       // optimisation.
00618       R11 = R1;
00619       R12 = Constant::getAllOnesValue(R1->getType());
00620     }
00621 
00622     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
00623       A = R11; D = R12; E = R2; ok = true;
00624     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
00625       A = R12; D = R11; E = R2; ok = true;
00626     }
00627   }
00628 
00629   // Bail if RHS was a icmp that can't be decomposed into an equality.
00630   if (!ICmpInst::isEquality(RHSCC))
00631     return 0;
00632 
00633   // Look for ANDs in on the right side of the RHS icmp.
00634   if (!ok && R2->getType()->isIntegerTy()) {
00635     if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
00636       R11 = R2;
00637       R12 = Constant::getAllOnesValue(R2->getType());
00638     }
00639 
00640     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
00641       A = R11; D = R12; E = R1; ok = true;
00642     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
00643       A = R12; D = R11; E = R1; ok = true;
00644     } else {
00645       return 0;
00646     }
00647   }
00648   if (!ok)
00649     return 0;
00650 
00651   if (L11 == A) {
00652     B = L12; C = L2;
00653   } else if (L12 == A) {
00654     B = L11; C = L2;
00655   } else if (L21 == A) {
00656     B = L22; C = L1;
00657   } else if (L22 == A) {
00658     B = L21; C = L1;
00659   }
00660 
00661   unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
00662   unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
00663   return left_type & right_type;
00664 }
00665 /// foldLogOpOfMaskedICmps:
00666 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
00667 /// into a single (icmp(A & X) ==/!= Y)
00668 static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
00669                                      llvm::InstCombiner::BuilderTy* Builder) {
00670   Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
00671   ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
00672   unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
00673                                                LHSCC, RHSCC);
00674   if (mask == 0) return nullptr;
00675   assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
00676          "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
00677 
00678   // In full generality:
00679   //     (icmp (A & B) Op C) | (icmp (A & D) Op E)
00680   // ==  ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
00681   //
00682   // If the latter can be converted into (icmp (A & X) Op Y) then the former is
00683   // equivalent to (icmp (A & X) !Op Y).
00684   //
00685   // Therefore, we can pretend for the rest of this function that we're dealing
00686   // with the conjunction, provided we flip the sense of any comparisons (both
00687   // input and output).
00688 
00689   // In most cases we're going to produce an EQ for the "&&" case.
00690   ICmpInst::Predicate NEWCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
00691   if (!IsAnd) {
00692     // Convert the masking analysis into its equivalent with negated
00693     // comparisons.
00694     mask = conjugateICmpMask(mask);
00695   }
00696 
00697   if (mask & FoldMskICmp_Mask_AllZeroes) {
00698     // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
00699     // -> (icmp eq (A & (B|D)), 0)
00700     Value* newOr = Builder->CreateOr(B, D);
00701     Value* newAnd = Builder->CreateAnd(A, newOr);
00702     // we can't use C as zero, because we might actually handle
00703     //   (icmp ne (A & B), B) & (icmp ne (A & D), D)
00704     // with B and D, having a single bit set
00705     Value* zero = Constant::getNullValue(A->getType());
00706     return Builder->CreateICmp(NEWCC, newAnd, zero);
00707   }
00708   if (mask & FoldMskICmp_BMask_AllOnes) {
00709     // (icmp eq (A & B), B) & (icmp eq (A & D), D)
00710     // -> (icmp eq (A & (B|D)), (B|D))
00711     Value* newOr = Builder->CreateOr(B, D);
00712     Value* newAnd = Builder->CreateAnd(A, newOr);
00713     return Builder->CreateICmp(NEWCC, newAnd, newOr);
00714   }
00715   if (mask & FoldMskICmp_AMask_AllOnes) {
00716     // (icmp eq (A & B), A) & (icmp eq (A & D), A)
00717     // -> (icmp eq (A & (B&D)), A)
00718     Value* newAnd1 = Builder->CreateAnd(B, D);
00719     Value* newAnd = Builder->CreateAnd(A, newAnd1);
00720     return Builder->CreateICmp(NEWCC, newAnd, A);
00721   }
00722 
00723   // Remaining cases assume at least that B and D are constant, and depend on
00724   // their actual values. This isn't strictly, necessary, just a "handle the
00725   // easy cases for now" decision.
00726   ConstantInt *BCst = dyn_cast<ConstantInt>(B);
00727   if (!BCst) return nullptr;
00728   ConstantInt *DCst = dyn_cast<ConstantInt>(D);
00729   if (!DCst) return nullptr;
00730 
00731   if (mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) {
00732     // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
00733     // (icmp ne (A & B), B) & (icmp ne (A & D), D)
00734     //     -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
00735     // Only valid if one of the masks is a superset of the other (check "B&D" is
00736     // the same as either B or D).
00737     APInt NewMask = BCst->getValue() & DCst->getValue();
00738 
00739     if (NewMask == BCst->getValue())
00740       return LHS;
00741     else if (NewMask == DCst->getValue())
00742       return RHS;
00743   }
00744   if (mask & FoldMskICmp_AMask_NotAllOnes) {
00745     // (icmp ne (A & B), B) & (icmp ne (A & D), D)
00746     //     -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
00747     // Only valid if one of the masks is a superset of the other (check "B|D" is
00748     // the same as either B or D).
00749     APInt NewMask = BCst->getValue() | DCst->getValue();
00750 
00751     if (NewMask == BCst->getValue())
00752       return LHS;
00753     else if (NewMask == DCst->getValue())
00754       return RHS;
00755   }
00756   if (mask & FoldMskICmp_BMask_Mixed) {
00757     // (icmp eq (A & B), C) & (icmp eq (A & D), E)
00758     // We already know that B & C == C && D & E == E.
00759     // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
00760     // C and E, which are shared by both the mask B and the mask D, don't
00761     // contradict, then we can transform to
00762     // -> (icmp eq (A & (B|D)), (C|E))
00763     // Currently, we only handle the case of B, C, D, and E being constant.
00764     // we can't simply use C and E, because we might actually handle
00765     //   (icmp ne (A & B), B) & (icmp eq (A & D), D)
00766     // with B and D, having a single bit set
00767     ConstantInt *CCst = dyn_cast<ConstantInt>(C);
00768     if (!CCst) return nullptr;
00769     if (LHSCC != NEWCC)
00770       CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) );
00771     ConstantInt *ECst = dyn_cast<ConstantInt>(E);
00772     if (!ECst) return nullptr;
00773     if (RHSCC != NEWCC)
00774       ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) );
00775     ConstantInt* MCst = dyn_cast<ConstantInt>(
00776       ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst),
00777                            ConstantExpr::getXor(CCst, ECst)) );
00778     // if there is a conflict we should actually return a false for the
00779     // whole construct
00780     if (!MCst->isZero())
00781       return nullptr;
00782     Value *newOr1 = Builder->CreateOr(B, D);
00783     Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
00784     Value *newAnd = Builder->CreateAnd(A, newOr1);
00785     return Builder->CreateICmp(NEWCC, newAnd, newOr2);
00786   }
00787   return nullptr;
00788 }
00789 
00790 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
00791 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
00792   ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
00793 
00794   // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
00795   if (PredicatesFoldable(LHSCC, RHSCC)) {
00796     if (LHS->getOperand(0) == RHS->getOperand(1) &&
00797         LHS->getOperand(1) == RHS->getOperand(0))
00798       LHS->swapOperands();
00799     if (LHS->getOperand(0) == RHS->getOperand(0) &&
00800         LHS->getOperand(1) == RHS->getOperand(1)) {
00801       Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
00802       unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
00803       bool isSigned = LHS->isSigned() || RHS->isSigned();
00804       return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
00805     }
00806   }
00807 
00808   // handle (roughly):  (icmp eq (A & B), C) & (icmp eq (A & D), E)
00809   if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
00810     return V;
00811 
00812   // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
00813   Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
00814   ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
00815   ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
00816   if (!LHSCst || !RHSCst) return nullptr;
00817 
00818   if (LHSCst == RHSCst && LHSCC == RHSCC) {
00819     // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
00820     // where C is a power of 2
00821     if (LHSCC == ICmpInst::ICMP_ULT &&
00822         LHSCst->getValue().isPowerOf2()) {
00823       Value *NewOr = Builder->CreateOr(Val, Val2);
00824       return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
00825     }
00826 
00827     // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
00828     if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
00829       Value *NewOr = Builder->CreateOr(Val, Val2);
00830       return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
00831     }
00832   }
00833 
00834   // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
00835   // where CMAX is the all ones value for the truncated type,
00836   // iff the lower bits of C2 and CA are zero.
00837   if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
00838       LHS->hasOneUse() && RHS->hasOneUse()) {
00839     Value *V;
00840     ConstantInt *AndCst, *SmallCst = nullptr, *BigCst = nullptr;
00841 
00842     // (trunc x) == C1 & (and x, CA) == C2
00843     // (and x, CA) == C2 & (trunc x) == C1
00844     if (match(Val2, m_Trunc(m_Value(V))) &&
00845         match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
00846       SmallCst = RHSCst;
00847       BigCst = LHSCst;
00848     } else if (match(Val, m_Trunc(m_Value(V))) &&
00849                match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
00850       SmallCst = LHSCst;
00851       BigCst = RHSCst;
00852     }
00853 
00854     if (SmallCst && BigCst) {
00855       unsigned BigBitSize = BigCst->getType()->getBitWidth();
00856       unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
00857 
00858       // Check that the low bits are zero.
00859       APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
00860       if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
00861         Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
00862         APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
00863         Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
00864         return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
00865       }
00866     }
00867   }
00868 
00869   // From here on, we only handle:
00870   //    (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
00871   if (Val != Val2) return nullptr;
00872 
00873   // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
00874   if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
00875       RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
00876       LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
00877       RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
00878     return nullptr;
00879 
00880   // Make a constant range that's the intersection of the two icmp ranges.
00881   // If the intersection is empty, we know that the result is false.
00882   ConstantRange LHSRange =
00883     ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
00884   ConstantRange RHSRange =
00885     ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
00886 
00887   if (LHSRange.intersectWith(RHSRange).isEmptySet())
00888     return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
00889 
00890   // We can't fold (ugt x, C) & (sgt x, C2).
00891   if (!PredicatesFoldable(LHSCC, RHSCC))
00892     return nullptr;
00893 
00894   // Ensure that the larger constant is on the RHS.
00895   bool ShouldSwap;
00896   if (CmpInst::isSigned(LHSCC) ||
00897       (ICmpInst::isEquality(LHSCC) &&
00898        CmpInst::isSigned(RHSCC)))
00899     ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
00900   else
00901     ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
00902 
00903   if (ShouldSwap) {
00904     std::swap(LHS, RHS);
00905     std::swap(LHSCst, RHSCst);
00906     std::swap(LHSCC, RHSCC);
00907   }
00908 
00909   // At this point, we know we have two icmp instructions
00910   // comparing a value against two constants and and'ing the result
00911   // together.  Because of the above check, we know that we only have
00912   // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
00913   // (from the icmp folding check above), that the two constants
00914   // are not equal and that the larger constant is on the RHS
00915   assert(LHSCst != RHSCst && "Compares not folded above?");
00916 
00917   switch (LHSCC) {
00918   default: llvm_unreachable("Unknown integer condition code!");
00919   case ICmpInst::ICMP_EQ:
00920     switch (RHSCC) {
00921     default: llvm_unreachable("Unknown integer condition code!");
00922     case ICmpInst::ICMP_NE:         // (X == 13 & X != 15) -> X == 13
00923     case ICmpInst::ICMP_ULT:        // (X == 13 & X <  15) -> X == 13
00924     case ICmpInst::ICMP_SLT:        // (X == 13 & X <  15) -> X == 13
00925       return LHS;
00926     }
00927   case ICmpInst::ICMP_NE:
00928     switch (RHSCC) {
00929     default: llvm_unreachable("Unknown integer condition code!");
00930     case ICmpInst::ICMP_ULT:
00931       if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
00932         return Builder->CreateICmpULT(Val, LHSCst);
00933       break;                        // (X != 13 & X u< 15) -> no change
00934     case ICmpInst::ICMP_SLT:
00935       if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
00936         return Builder->CreateICmpSLT(Val, LHSCst);
00937       break;                        // (X != 13 & X s< 15) -> no change
00938     case ICmpInst::ICMP_EQ:         // (X != 13 & X == 15) -> X == 15
00939     case ICmpInst::ICMP_UGT:        // (X != 13 & X u> 15) -> X u> 15
00940     case ICmpInst::ICMP_SGT:        // (X != 13 & X s> 15) -> X s> 15
00941       return RHS;
00942     case ICmpInst::ICMP_NE:
00943       // Special case to get the ordering right when the values wrap around
00944       // zero.
00945       if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue())
00946         std::swap(LHSCst, RHSCst);
00947       if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
00948         Constant *AddCST = ConstantExpr::getNeg(LHSCst);
00949         Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
00950         return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1),
00951                                       Val->getName()+".cmp");
00952       }
00953       break;                        // (X != 13 & X != 15) -> no change
00954     }
00955     break;
00956   case ICmpInst::ICMP_ULT:
00957     switch (RHSCC) {
00958     default: llvm_unreachable("Unknown integer condition code!");
00959     case ICmpInst::ICMP_EQ:         // (X u< 13 & X == 15) -> false
00960     case ICmpInst::ICMP_UGT:        // (X u< 13 & X u> 15) -> false
00961       return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
00962     case ICmpInst::ICMP_SGT:        // (X u< 13 & X s> 15) -> no change
00963       break;
00964     case ICmpInst::ICMP_NE:         // (X u< 13 & X != 15) -> X u< 13
00965     case ICmpInst::ICMP_ULT:        // (X u< 13 & X u< 15) -> X u< 13
00966       return LHS;
00967     case ICmpInst::ICMP_SLT:        // (X u< 13 & X s< 15) -> no change
00968       break;
00969     }
00970     break;
00971   case ICmpInst::ICMP_SLT:
00972     switch (RHSCC) {
00973     default: llvm_unreachable("Unknown integer condition code!");
00974     case ICmpInst::ICMP_UGT:        // (X s< 13 & X u> 15) -> no change
00975       break;
00976     case ICmpInst::ICMP_NE:         // (X s< 13 & X != 15) -> X < 13
00977     case ICmpInst::ICMP_SLT:        // (X s< 13 & X s< 15) -> X < 13
00978       return LHS;
00979     case ICmpInst::ICMP_ULT:        // (X s< 13 & X u< 15) -> no change
00980       break;
00981     }
00982     break;
00983   case ICmpInst::ICMP_UGT:
00984     switch (RHSCC) {
00985     default: llvm_unreachable("Unknown integer condition code!");
00986     case ICmpInst::ICMP_EQ:         // (X u> 13 & X == 15) -> X == 15
00987     case ICmpInst::ICMP_UGT:        // (X u> 13 & X u> 15) -> X u> 15
00988       return RHS;
00989     case ICmpInst::ICMP_SGT:        // (X u> 13 & X s> 15) -> no change
00990       break;
00991     case ICmpInst::ICMP_NE:
00992       if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
00993         return Builder->CreateICmp(LHSCC, Val, RHSCst);
00994       break;                        // (X u> 13 & X != 15) -> no change
00995     case ICmpInst::ICMP_ULT:        // (X u> 13 & X u< 15) -> (X-14) <u 1
00996       return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
00997     case ICmpInst::ICMP_SLT:        // (X u> 13 & X s< 15) -> no change
00998       break;
00999     }
01000     break;
01001   case ICmpInst::ICMP_SGT:
01002     switch (RHSCC) {
01003     default: llvm_unreachable("Unknown integer condition code!");
01004     case ICmpInst::ICMP_EQ:         // (X s> 13 & X == 15) -> X == 15
01005     case ICmpInst::ICMP_SGT:        // (X s> 13 & X s> 15) -> X s> 15
01006       return RHS;
01007     case ICmpInst::ICMP_UGT:        // (X s> 13 & X u> 15) -> no change
01008       break;
01009     case ICmpInst::ICMP_NE:
01010       if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
01011         return Builder->CreateICmp(LHSCC, Val, RHSCst);
01012       break;                        // (X s> 13 & X != 15) -> no change
01013     case ICmpInst::ICMP_SLT:        // (X s> 13 & X s< 15) -> (X-14) s< 1
01014       return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
01015     case ICmpInst::ICMP_ULT:        // (X s> 13 & X u< 15) -> no change
01016       break;
01017     }
01018     break;
01019   }
01020 
01021   return nullptr;
01022 }
01023 
01024 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp).  NOTE: Unlike the rest of
01025 /// instcombine, this returns a Value which should already be inserted into the
01026 /// function.
01027 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
01028   if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
01029       RHS->getPredicate() == FCmpInst::FCMP_ORD) {
01030     if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
01031       return nullptr;
01032 
01033     // (fcmp ord x, c) & (fcmp ord y, c)  -> (fcmp ord x, y)
01034     if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
01035       if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
01036         // If either of the constants are nans, then the whole thing returns
01037         // false.
01038         if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
01039           return Builder->getFalse();
01040         return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
01041       }
01042 
01043     // Handle vector zeros.  This occurs because the canonical form of
01044     // "fcmp ord x,x" is "fcmp ord x, 0".
01045     if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
01046         isa<ConstantAggregateZero>(RHS->getOperand(1)))
01047       return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
01048     return nullptr;
01049   }
01050 
01051   Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
01052   Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
01053   FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
01054 
01055 
01056   if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
01057     // Swap RHS operands to match LHS.
01058     Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
01059     std::swap(Op1LHS, Op1RHS);
01060   }
01061 
01062   if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
01063     // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
01064     if (Op0CC == Op1CC)
01065       return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
01066     if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
01067       return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
01068     if (Op0CC == FCmpInst::FCMP_TRUE)
01069       return RHS;
01070     if (Op1CC == FCmpInst::FCMP_TRUE)
01071       return LHS;
01072 
01073     bool Op0Ordered;
01074     bool Op1Ordered;
01075     unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
01076     unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
01077     // uno && ord -> false
01078     if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered)
01079         return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
01080     if (Op1Pred == 0) {
01081       std::swap(LHS, RHS);
01082       std::swap(Op0Pred, Op1Pred);
01083       std::swap(Op0Ordered, Op1Ordered);
01084     }
01085     if (Op0Pred == 0) {
01086       // uno && ueq -> uno && (uno || eq) -> uno
01087       // ord && olt -> ord && (ord && lt) -> olt
01088       if (!Op0Ordered && (Op0Ordered == Op1Ordered))
01089         return LHS;
01090       if (Op0Ordered && (Op0Ordered == Op1Ordered))
01091         return RHS;
01092 
01093       // uno && oeq -> uno && (ord && eq) -> false
01094       if (!Op0Ordered)
01095         return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
01096       // ord && ueq -> ord && (uno || eq) -> oeq
01097       return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
01098     }
01099   }
01100 
01101   return nullptr;
01102 }
01103 
01104 
01105 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
01106   bool Changed = SimplifyAssociativeOrCommutative(I);
01107   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
01108 
01109   if (Value *V = SimplifyVectorOp(I))
01110     return ReplaceInstUsesWith(I, V);
01111 
01112   if (Value *V = SimplifyAndInst(Op0, Op1, DL))
01113     return ReplaceInstUsesWith(I, V);
01114 
01115   // (A|B)&(A|C) -> A|(B&C) etc
01116   if (Value *V = SimplifyUsingDistributiveLaws(I))
01117     return ReplaceInstUsesWith(I, V);
01118 
01119   // See if we can simplify any instructions used by the instruction whose sole
01120   // purpose is to compute bits we don't care about.
01121   if (SimplifyDemandedInstructionBits(I))
01122     return &I;
01123 
01124   if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
01125     const APInt &AndRHSMask = AndRHS->getValue();
01126 
01127     // Optimize a variety of ((val OP C1) & C2) combinations...
01128     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
01129       Value *Op0LHS = Op0I->getOperand(0);
01130       Value *Op0RHS = Op0I->getOperand(1);
01131       switch (Op0I->getOpcode()) {
01132       default: break;
01133       case Instruction::Xor:
01134       case Instruction::Or: {
01135         // If the mask is only needed on one incoming arm, push it up.
01136         if (!Op0I->hasOneUse()) break;
01137 
01138         APInt NotAndRHS(~AndRHSMask);
01139         if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
01140           // Not masking anything out for the LHS, move to RHS.
01141           Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
01142                                              Op0RHS->getName()+".masked");
01143           return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
01144         }
01145         if (!isa<Constant>(Op0RHS) &&
01146             MaskedValueIsZero(Op0RHS, NotAndRHS)) {
01147           // Not masking anything out for the RHS, move to LHS.
01148           Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
01149                                              Op0LHS->getName()+".masked");
01150           return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
01151         }
01152 
01153         break;
01154       }
01155       case Instruction::Add:
01156         // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
01157         // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
01158         // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
01159         if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
01160           return BinaryOperator::CreateAnd(V, AndRHS);
01161         if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
01162           return BinaryOperator::CreateAnd(V, AndRHS);  // Add commutes
01163         break;
01164 
01165       case Instruction::Sub:
01166         // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
01167         // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
01168         // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
01169         if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
01170           return BinaryOperator::CreateAnd(V, AndRHS);
01171 
01172         // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
01173         // has 1's for all bits that the subtraction with A might affect.
01174         if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
01175           uint32_t BitWidth = AndRHSMask.getBitWidth();
01176           uint32_t Zeros = AndRHSMask.countLeadingZeros();
01177           APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
01178 
01179           if (MaskedValueIsZero(Op0LHS, Mask)) {
01180             Value *NewNeg = Builder->CreateNeg(Op0RHS);
01181             return BinaryOperator::CreateAnd(NewNeg, AndRHS);
01182           }
01183         }
01184         break;
01185 
01186       case Instruction::Shl:
01187       case Instruction::LShr:
01188         // (1 << x) & 1 --> zext(x == 0)
01189         // (1 >> x) & 1 --> zext(x == 0)
01190         if (AndRHSMask == 1 && Op0LHS == AndRHS) {
01191           Value *NewICmp =
01192             Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
01193           return new ZExtInst(NewICmp, I.getType());
01194         }
01195         break;
01196       }
01197 
01198       if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
01199         if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
01200           return Res;
01201     }
01202 
01203     // If this is an integer truncation, and if the source is an 'and' with
01204     // immediate, transform it.  This frequently occurs for bitfield accesses.
01205     {
01206       Value *X = nullptr; ConstantInt *YC = nullptr;
01207       if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
01208         // Change: and (trunc (and X, YC) to T), C2
01209         // into  : and (trunc X to T), trunc(YC) & C2
01210         // This will fold the two constants together, which may allow
01211         // other simplifications.
01212         Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
01213         Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
01214         C3 = ConstantExpr::getAnd(C3, AndRHS);
01215         return BinaryOperator::CreateAnd(NewCast, C3);
01216       }
01217     }
01218 
01219     // Try to fold constant and into select arguments.
01220     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
01221       if (Instruction *R = FoldOpIntoSelect(I, SI))
01222         return R;
01223     if (isa<PHINode>(Op0))
01224       if (Instruction *NV = FoldOpIntoPhi(I))
01225         return NV;
01226   }
01227 
01228 
01229   // (~A & ~B) == (~(A | B)) - De Morgan's Law
01230   if (Value *Op0NotVal = dyn_castNotVal(Op0))
01231     if (Value *Op1NotVal = dyn_castNotVal(Op1))
01232       if (Op0->hasOneUse() && Op1->hasOneUse()) {
01233         Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
01234                                       I.getName()+".demorgan");
01235         return BinaryOperator::CreateNot(Or);
01236       }
01237 
01238   {
01239     Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
01240     // (A|B) & ~(A&B) -> A^B
01241     if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
01242         match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
01243         ((A == C && B == D) || (A == D && B == C)))
01244       return BinaryOperator::CreateXor(A, B);
01245 
01246     // ~(A&B) & (A|B) -> A^B
01247     if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
01248         match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
01249         ((A == C && B == D) || (A == D && B == C)))
01250       return BinaryOperator::CreateXor(A, B);
01251 
01252     // A&(A^B) => A & ~B
01253     {
01254       Value *tmpOp0 = Op0;
01255       Value *tmpOp1 = Op1;
01256       if (Op0->hasOneUse() &&
01257           match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
01258         if (A == Op1 || B == Op1 ) {
01259           tmpOp1 = Op0;
01260           tmpOp0 = Op1;
01261           // Simplify below
01262         }
01263       }
01264 
01265       if (tmpOp1->hasOneUse() &&
01266           match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
01267         if (B == tmpOp0) {
01268           std::swap(A, B);
01269         }
01270         // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
01271         // A is originally -1 (or a vector of -1 and undefs), then we enter
01272         // an endless loop. By checking that A is non-constant we ensure that
01273         // we will never get to the loop.
01274         if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
01275           return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
01276       }
01277     }
01278 
01279     // (A&((~A)|B)) -> A&B
01280     if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
01281         match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
01282       return BinaryOperator::CreateAnd(A, Op1);
01283     if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
01284         match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
01285       return BinaryOperator::CreateAnd(A, Op0);
01286 
01287     // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
01288     if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
01289       if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
01290         if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
01291           return BinaryOperator::CreateAnd(Op0, Builder->CreateNot(C));
01292 
01293     // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
01294     if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
01295       if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
01296         if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
01297           return BinaryOperator::CreateAnd(Op1, Builder->CreateNot(C));
01298   }
01299 
01300   if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
01301     if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
01302       if (Value *Res = FoldAndOfICmps(LHS, RHS))
01303         return ReplaceInstUsesWith(I, Res);
01304 
01305   // If and'ing two fcmp, try combine them into one.
01306   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
01307     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
01308       if (Value *Res = FoldAndOfFCmps(LHS, RHS))
01309         return ReplaceInstUsesWith(I, Res);
01310 
01311 
01312   // fold (and (cast A), (cast B)) -> (cast (and A, B))
01313   if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
01314     if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
01315       Type *SrcTy = Op0C->getOperand(0)->getType();
01316       if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
01317           SrcTy == Op1C->getOperand(0)->getType() &&
01318           SrcTy->isIntOrIntVectorTy()) {
01319         Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
01320 
01321         // Only do this if the casts both really cause code to be generated.
01322         if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
01323             ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
01324           Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
01325           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
01326         }
01327 
01328         // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
01329         // cast is otherwise not optimizable.  This happens for vector sexts.
01330         if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
01331           if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
01332             if (Value *Res = FoldAndOfICmps(LHS, RHS))
01333               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
01334 
01335         // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
01336         // cast is otherwise not optimizable.  This happens for vector sexts.
01337         if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
01338           if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
01339             if (Value *Res = FoldAndOfFCmps(LHS, RHS))
01340               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
01341       }
01342     }
01343 
01344   // (X >> Z) & (Y >> Z)  -> (X&Y) >> Z  for all shifts.
01345   if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
01346     if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
01347       if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
01348           SI0->getOperand(1) == SI1->getOperand(1) &&
01349           (SI0->hasOneUse() || SI1->hasOneUse())) {
01350         Value *NewOp =
01351           Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
01352                              SI0->getName());
01353         return BinaryOperator::Create(SI1->getOpcode(), NewOp,
01354                                       SI1->getOperand(1));
01355       }
01356   }
01357 
01358   {
01359     Value *X = nullptr;
01360     bool OpsSwapped = false;
01361     // Canonicalize SExt or Not to the LHS
01362     if (match(Op1, m_SExt(m_Value())) ||
01363         match(Op1, m_Not(m_Value()))) {
01364       std::swap(Op0, Op1);
01365       OpsSwapped = true;
01366     }
01367 
01368     // Fold (and (sext bool to A), B) --> (select bool, B, 0)
01369     if (match(Op0, m_SExt(m_Value(X))) &&
01370         X->getType()->getScalarType()->isIntegerTy(1)) {
01371       Value *Zero = Constant::getNullValue(Op1->getType());
01372       return SelectInst::Create(X, Op1, Zero);
01373     }
01374 
01375     // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
01376     if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
01377         X->getType()->getScalarType()->isIntegerTy(1)) {
01378       Value *Zero = Constant::getNullValue(Op0->getType());
01379       return SelectInst::Create(X, Zero, Op1);
01380     }
01381 
01382     if (OpsSwapped)
01383       std::swap(Op0, Op1);
01384   }
01385 
01386   return Changed ? &I : nullptr;
01387 }
01388 
01389 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
01390 /// capable of providing pieces of a bswap.  The subexpression provides pieces
01391 /// of a bswap if it is proven that each of the non-zero bytes in the output of
01392 /// the expression came from the corresponding "byte swapped" byte in some other
01393 /// value.  For example, if the current subexpression is "(shl i32 %X, 24)" then
01394 /// we know that the expression deposits the low byte of %X into the high byte
01395 /// of the bswap result and that all other bytes are zero.  This expression is
01396 /// accepted, the high byte of ByteValues is set to X to indicate a correct
01397 /// match.
01398 ///
01399 /// This function returns true if the match was unsuccessful and false if so.
01400 /// On entry to the function the "OverallLeftShift" is a signed integer value
01401 /// indicating the number of bytes that the subexpression is later shifted.  For
01402 /// example, if the expression is later right shifted by 16 bits, the
01403 /// OverallLeftShift value would be -2 on entry.  This is used to specify which
01404 /// byte of ByteValues is actually being set.
01405 ///
01406 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
01407 /// byte is masked to zero by a user.  For example, in (X & 255), X will be
01408 /// processed with a bytemask of 1.  Because bytemask is 32-bits, this limits
01409 /// this function to working on up to 32-byte (256 bit) values.  ByteMask is
01410 /// always in the local (OverallLeftShift) coordinate space.
01411 ///
01412 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
01413                               SmallVectorImpl<Value *> &ByteValues) {
01414   if (Instruction *I = dyn_cast<Instruction>(V)) {
01415     // If this is an or instruction, it may be an inner node of the bswap.
01416     if (I->getOpcode() == Instruction::Or) {
01417       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
01418                                ByteValues) ||
01419              CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
01420                                ByteValues);
01421     }
01422 
01423     // If this is a logical shift by a constant multiple of 8, recurse with
01424     // OverallLeftShift and ByteMask adjusted.
01425     if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
01426       unsigned ShAmt =
01427         cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
01428       // Ensure the shift amount is defined and of a byte value.
01429       if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
01430         return true;
01431 
01432       unsigned ByteShift = ShAmt >> 3;
01433       if (I->getOpcode() == Instruction::Shl) {
01434         // X << 2 -> collect(X, +2)
01435         OverallLeftShift += ByteShift;
01436         ByteMask >>= ByteShift;
01437       } else {
01438         // X >>u 2 -> collect(X, -2)
01439         OverallLeftShift -= ByteShift;
01440         ByteMask <<= ByteShift;
01441         ByteMask &= (~0U >> (32-ByteValues.size()));
01442       }
01443 
01444       if (OverallLeftShift >= (int)ByteValues.size()) return true;
01445       if (OverallLeftShift <= -(int)ByteValues.size()) return true;
01446 
01447       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
01448                                ByteValues);
01449     }
01450 
01451     // If this is a logical 'and' with a mask that clears bytes, clear the
01452     // corresponding bytes in ByteMask.
01453     if (I->getOpcode() == Instruction::And &&
01454         isa<ConstantInt>(I->getOperand(1))) {
01455       // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
01456       unsigned NumBytes = ByteValues.size();
01457       APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
01458       const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
01459 
01460       for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
01461         // If this byte is masked out by a later operation, we don't care what
01462         // the and mask is.
01463         if ((ByteMask & (1 << i)) == 0)
01464           continue;
01465 
01466         // If the AndMask is all zeros for this byte, clear the bit.
01467         APInt MaskB = AndMask & Byte;
01468         if (MaskB == 0) {
01469           ByteMask &= ~(1U << i);
01470           continue;
01471         }
01472 
01473         // If the AndMask is not all ones for this byte, it's not a bytezap.
01474         if (MaskB != Byte)
01475           return true;
01476 
01477         // Otherwise, this byte is kept.
01478       }
01479 
01480       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
01481                                ByteValues);
01482     }
01483   }
01484 
01485   // Okay, we got to something that isn't a shift, 'or' or 'and'.  This must be
01486   // the input value to the bswap.  Some observations: 1) if more than one byte
01487   // is demanded from this input, then it could not be successfully assembled
01488   // into a byteswap.  At least one of the two bytes would not be aligned with
01489   // their ultimate destination.
01490   if (!isPowerOf2_32(ByteMask)) return true;
01491   unsigned InputByteNo = countTrailingZeros(ByteMask);
01492 
01493   // 2) The input and ultimate destinations must line up: if byte 3 of an i32
01494   // is demanded, it needs to go into byte 0 of the result.  This means that the
01495   // byte needs to be shifted until it lands in the right byte bucket.  The
01496   // shift amount depends on the position: if the byte is coming from the high
01497   // part of the value (e.g. byte 3) then it must be shifted right.  If from the
01498   // low part, it must be shifted left.
01499   unsigned DestByteNo = InputByteNo + OverallLeftShift;
01500   if (ByteValues.size()-1-DestByteNo != InputByteNo)
01501     return true;
01502 
01503   // If the destination byte value is already defined, the values are or'd
01504   // together, which isn't a bswap (unless it's an or of the same bits).
01505   if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
01506     return true;
01507   ByteValues[DestByteNo] = V;
01508   return false;
01509 }
01510 
01511 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
01512 /// If so, insert the new bswap intrinsic and return it.
01513 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
01514   IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
01515   if (!ITy || ITy->getBitWidth() % 16 ||
01516       // ByteMask only allows up to 32-byte values.
01517       ITy->getBitWidth() > 32*8)
01518     return nullptr;   // Can only bswap pairs of bytes.  Can't do vectors.
01519 
01520   /// ByteValues - For each byte of the result, we keep track of which value
01521   /// defines each byte.
01522   SmallVector<Value*, 8> ByteValues;
01523   ByteValues.resize(ITy->getBitWidth()/8);
01524 
01525   // Try to find all the pieces corresponding to the bswap.
01526   uint32_t ByteMask = ~0U >> (32-ByteValues.size());
01527   if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
01528     return nullptr;
01529 
01530   // Check to see if all of the bytes come from the same value.
01531   Value *V = ByteValues[0];
01532   if (!V) return nullptr;  // Didn't find a byte?  Must be zero.
01533 
01534   // Check to make sure that all of the bytes come from the same value.
01535   for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
01536     if (ByteValues[i] != V)
01537       return nullptr;
01538   Module *M = I.getParent()->getParent()->getParent();
01539   Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
01540   return CallInst::Create(F, V);
01541 }
01542 
01543 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D).  Check
01544 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
01545 /// we can simplify this expression to "cond ? C : D or B".
01546 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
01547                                          Value *C, Value *D) {
01548   // If A is not a select of -1/0, this cannot match.
01549   Value *Cond = nullptr;
01550   if (!match(A, m_SExt(m_Value(Cond))) ||
01551       !Cond->getType()->isIntegerTy(1))
01552     return nullptr;
01553 
01554   // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
01555   if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
01556     return SelectInst::Create(Cond, C, B);
01557   if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
01558     return SelectInst::Create(Cond, C, B);
01559 
01560   // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
01561   if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
01562     return SelectInst::Create(Cond, C, D);
01563   if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
01564     return SelectInst::Create(Cond, C, D);
01565   return nullptr;
01566 }
01567 
01568 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
01569 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
01570   ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
01571 
01572   // Fold (iszero(A & K1) | iszero(A & K2)) ->  (A & (K1 | K2)) != (K1 | K2)
01573   // if K1 and K2 are a one-bit mask.
01574   ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
01575   ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
01576 
01577   if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() &&
01578       RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
01579 
01580     BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0));
01581     BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0));
01582     if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() &&
01583         LAnd->getOpcode() == Instruction::And &&
01584         RAnd->getOpcode() == Instruction::And) {
01585 
01586       Value *Mask = nullptr;
01587       Value *Masked = nullptr;
01588       if (LAnd->getOperand(0) == RAnd->getOperand(0) &&
01589           isKnownToBeAPowerOfTwo(LAnd->getOperand(1)) &&
01590           isKnownToBeAPowerOfTwo(RAnd->getOperand(1))) {
01591         Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1));
01592         Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask);
01593       } else if (LAnd->getOperand(1) == RAnd->getOperand(1) &&
01594                  isKnownToBeAPowerOfTwo(LAnd->getOperand(0)) &&
01595                  isKnownToBeAPowerOfTwo(RAnd->getOperand(0))) {
01596         Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0));
01597         Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask);
01598       }
01599 
01600       if (Masked)
01601         return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask);
01602     }
01603   }
01604 
01605   // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
01606   if (PredicatesFoldable(LHSCC, RHSCC)) {
01607     if (LHS->getOperand(0) == RHS->getOperand(1) &&
01608         LHS->getOperand(1) == RHS->getOperand(0))
01609       LHS->swapOperands();
01610     if (LHS->getOperand(0) == RHS->getOperand(0) &&
01611         LHS->getOperand(1) == RHS->getOperand(1)) {
01612       Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
01613       unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
01614       bool isSigned = LHS->isSigned() || RHS->isSigned();
01615       return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
01616     }
01617   }
01618 
01619   // handle (roughly):
01620   // (icmp ne (A & B), C) | (icmp ne (A & D), E)
01621   if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
01622     return V;
01623 
01624   Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
01625   if (LHS->hasOneUse() || RHS->hasOneUse()) {
01626     // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
01627     // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
01628     Value *A = nullptr, *B = nullptr;
01629     if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
01630       B = Val;
01631       if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
01632         A = Val2;
01633       else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
01634         A = RHS->getOperand(1);
01635     }
01636     // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
01637     // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
01638     else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
01639       B = Val2;
01640       if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
01641         A = Val;
01642       else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
01643         A = LHS->getOperand(1);
01644     }
01645     if (A && B)
01646       return Builder->CreateICmp(
01647           ICmpInst::ICMP_UGE,
01648           Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
01649   }
01650 
01651   // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
01652   if (!LHSCst || !RHSCst) return nullptr;
01653 
01654   if (LHSCst == RHSCst && LHSCC == RHSCC) {
01655     // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
01656     if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
01657       Value *NewOr = Builder->CreateOr(Val, Val2);
01658       return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
01659     }
01660   }
01661 
01662   // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
01663   //   iff C2 + CA == C1.
01664   if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
01665     ConstantInt *AddCst;
01666     if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
01667       if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
01668         return Builder->CreateICmpULE(Val, LHSCst);
01669   }
01670 
01671   // From here on, we only handle:
01672   //    (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
01673   if (Val != Val2) return nullptr;
01674 
01675   // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
01676   if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
01677       RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
01678       LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
01679       RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
01680     return nullptr;
01681 
01682   // We can't fold (ugt x, C) | (sgt x, C2).
01683   if (!PredicatesFoldable(LHSCC, RHSCC))
01684     return nullptr;
01685 
01686   // Ensure that the larger constant is on the RHS.
01687   bool ShouldSwap;
01688   if (CmpInst::isSigned(LHSCC) ||
01689       (ICmpInst::isEquality(LHSCC) &&
01690        CmpInst::isSigned(RHSCC)))
01691     ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
01692   else
01693     ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
01694 
01695   if (ShouldSwap) {
01696     std::swap(LHS, RHS);
01697     std::swap(LHSCst, RHSCst);
01698     std::swap(LHSCC, RHSCC);
01699   }
01700 
01701   // At this point, we know we have two icmp instructions
01702   // comparing a value against two constants and or'ing the result
01703   // together.  Because of the above check, we know that we only have
01704   // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
01705   // icmp folding check above), that the two constants are not
01706   // equal.
01707   assert(LHSCst != RHSCst && "Compares not folded above?");
01708 
01709   switch (LHSCC) {
01710   default: llvm_unreachable("Unknown integer condition code!");
01711   case ICmpInst::ICMP_EQ:
01712     switch (RHSCC) {
01713     default: llvm_unreachable("Unknown integer condition code!");
01714     case ICmpInst::ICMP_EQ:
01715       if (LHS->getOperand(0) == RHS->getOperand(0)) {
01716         // if LHSCst and RHSCst differ only by one bit:
01717         // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
01718         assert(LHSCst->getValue().ule(LHSCst->getValue()));
01719 
01720         APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
01721         if (Xor.isPowerOf2()) {
01722           Value *NegCst = Builder->getInt(~Xor);
01723           Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
01724           return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
01725         }
01726       }
01727 
01728       if (LHSCst == SubOne(RHSCst)) {
01729         // (X == 13 | X == 14) -> X-13 <u 2
01730         Constant *AddCST = ConstantExpr::getNeg(LHSCst);
01731         Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
01732         AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
01733         return Builder->CreateICmpULT(Add, AddCST);
01734       }
01735 
01736       break;                         // (X == 13 | X == 15) -> no change
01737     case ICmpInst::ICMP_UGT:         // (X == 13 | X u> 14) -> no change
01738     case ICmpInst::ICMP_SGT:         // (X == 13 | X s> 14) -> no change
01739       break;
01740     case ICmpInst::ICMP_NE:          // (X == 13 | X != 15) -> X != 15
01741     case ICmpInst::ICMP_ULT:         // (X == 13 | X u< 15) -> X u< 15
01742     case ICmpInst::ICMP_SLT:         // (X == 13 | X s< 15) -> X s< 15
01743       return RHS;
01744     }
01745     break;
01746   case ICmpInst::ICMP_NE:
01747     switch (RHSCC) {
01748     default: llvm_unreachable("Unknown integer condition code!");
01749     case ICmpInst::ICMP_EQ:          // (X != 13 | X == 15) -> X != 13
01750     case ICmpInst::ICMP_UGT:         // (X != 13 | X u> 15) -> X != 13
01751     case ICmpInst::ICMP_SGT:         // (X != 13 | X s> 15) -> X != 13
01752       return LHS;
01753     case ICmpInst::ICMP_NE:          // (X != 13 | X != 15) -> true
01754     case ICmpInst::ICMP_ULT:         // (X != 13 | X u< 15) -> true
01755     case ICmpInst::ICMP_SLT:         // (X != 13 | X s< 15) -> true
01756       return Builder->getTrue();
01757     }
01758   case ICmpInst::ICMP_ULT:
01759     switch (RHSCC) {
01760     default: llvm_unreachable("Unknown integer condition code!");
01761     case ICmpInst::ICMP_EQ:         // (X u< 13 | X == 14) -> no change
01762       break;
01763     case ICmpInst::ICMP_UGT:        // (X u< 13 | X u> 15) -> (X-13) u> 2
01764       // If RHSCst is [us]MAXINT, it is always false.  Not handling
01765       // this can cause overflow.
01766       if (RHSCst->isMaxValue(false))
01767         return LHS;
01768       return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
01769     case ICmpInst::ICMP_SGT:        // (X u< 13 | X s> 15) -> no change
01770       break;
01771     case ICmpInst::ICMP_NE:         // (X u< 13 | X != 15) -> X != 15
01772     case ICmpInst::ICMP_ULT:        // (X u< 13 | X u< 15) -> X u< 15
01773       return RHS;
01774     case ICmpInst::ICMP_SLT:        // (X u< 13 | X s< 15) -> no change
01775       break;
01776     }
01777     break;
01778   case ICmpInst::ICMP_SLT:
01779     switch (RHSCC) {
01780     default: llvm_unreachable("Unknown integer condition code!");
01781     case ICmpInst::ICMP_EQ:         // (X s< 13 | X == 14) -> no change
01782       break;
01783     case ICmpInst::ICMP_SGT:        // (X s< 13 | X s> 15) -> (X-13) s> 2
01784       // If RHSCst is [us]MAXINT, it is always false.  Not handling
01785       // this can cause overflow.
01786       if (RHSCst->isMaxValue(true))
01787         return LHS;
01788       return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
01789     case ICmpInst::ICMP_UGT:        // (X s< 13 | X u> 15) -> no change
01790       break;
01791     case ICmpInst::ICMP_NE:         // (X s< 13 | X != 15) -> X != 15
01792     case ICmpInst::ICMP_SLT:        // (X s< 13 | X s< 15) -> X s< 15
01793       return RHS;
01794     case ICmpInst::ICMP_ULT:        // (X s< 13 | X u< 15) -> no change
01795       break;
01796     }
01797     break;
01798   case ICmpInst::ICMP_UGT:
01799     switch (RHSCC) {
01800     default: llvm_unreachable("Unknown integer condition code!");
01801     case ICmpInst::ICMP_EQ:         // (X u> 13 | X == 15) -> X u> 13
01802     case ICmpInst::ICMP_UGT:        // (X u> 13 | X u> 15) -> X u> 13
01803       return LHS;
01804     case ICmpInst::ICMP_SGT:        // (X u> 13 | X s> 15) -> no change
01805       break;
01806     case ICmpInst::ICMP_NE:         // (X u> 13 | X != 15) -> true
01807     case ICmpInst::ICMP_ULT:        // (X u> 13 | X u< 15) -> true
01808       return Builder->getTrue();
01809     case ICmpInst::ICMP_SLT:        // (X u> 13 | X s< 15) -> no change
01810       break;
01811     }
01812     break;
01813   case ICmpInst::ICMP_SGT:
01814     switch (RHSCC) {
01815     default: llvm_unreachable("Unknown integer condition code!");
01816     case ICmpInst::ICMP_EQ:         // (X s> 13 | X == 15) -> X > 13
01817     case ICmpInst::ICMP_SGT:        // (X s> 13 | X s> 15) -> X > 13
01818       return LHS;
01819     case ICmpInst::ICMP_UGT:        // (X s> 13 | X u> 15) -> no change
01820       break;
01821     case ICmpInst::ICMP_NE:         // (X s> 13 | X != 15) -> true
01822     case ICmpInst::ICMP_SLT:        // (X s> 13 | X s< 15) -> true
01823       return Builder->getTrue();
01824     case ICmpInst::ICMP_ULT:        // (X s> 13 | X u< 15) -> no change
01825       break;
01826     }
01827     break;
01828   }
01829   return nullptr;
01830 }
01831 
01832 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp).  NOTE: Unlike the rest of
01833 /// instcombine, this returns a Value which should already be inserted into the
01834 /// function.
01835 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
01836   if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
01837       RHS->getPredicate() == FCmpInst::FCMP_UNO &&
01838       LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
01839     if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
01840       if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
01841         // If either of the constants are nans, then the whole thing returns
01842         // true.
01843         if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
01844           return Builder->getTrue();
01845 
01846         // Otherwise, no need to compare the two constants, compare the
01847         // rest.
01848         return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
01849       }
01850 
01851     // Handle vector zeros.  This occurs because the canonical form of
01852     // "fcmp uno x,x" is "fcmp uno x, 0".
01853     if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
01854         isa<ConstantAggregateZero>(RHS->getOperand(1)))
01855       return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
01856 
01857     return nullptr;
01858   }
01859 
01860   Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
01861   Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
01862   FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
01863 
01864   if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
01865     // Swap RHS operands to match LHS.
01866     Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
01867     std::swap(Op1LHS, Op1RHS);
01868   }
01869   if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
01870     // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
01871     if (Op0CC == Op1CC)
01872       return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
01873     if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
01874       return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
01875     if (Op0CC == FCmpInst::FCMP_FALSE)
01876       return RHS;
01877     if (Op1CC == FCmpInst::FCMP_FALSE)
01878       return LHS;
01879     bool Op0Ordered;
01880     bool Op1Ordered;
01881     unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
01882     unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
01883     if (Op0Ordered == Op1Ordered) {
01884       // If both are ordered or unordered, return a new fcmp with
01885       // or'ed predicates.
01886       return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
01887     }
01888   }
01889   return nullptr;
01890 }
01891 
01892 /// FoldOrWithConstants - This helper function folds:
01893 ///
01894 ///     ((A | B) & C1) | (B & C2)
01895 ///
01896 /// into:
01897 ///
01898 ///     (A & C1) | B
01899 ///
01900 /// when the XOR of the two constants is "all ones" (-1).
01901 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
01902                                                Value *A, Value *B, Value *C) {
01903   ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
01904   if (!CI1) return nullptr;
01905 
01906   Value *V1 = nullptr;
01907   ConstantInt *CI2 = nullptr;
01908   if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr;
01909 
01910   APInt Xor = CI1->getValue() ^ CI2->getValue();
01911   if (!Xor.isAllOnesValue()) return nullptr;
01912 
01913   if (V1 == A || V1 == B) {
01914     Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
01915     return BinaryOperator::CreateOr(NewOp, V1);
01916   }
01917 
01918   return nullptr;
01919 }
01920 
01921 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
01922   bool Changed = SimplifyAssociativeOrCommutative(I);
01923   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
01924 
01925   if (Value *V = SimplifyVectorOp(I))
01926     return ReplaceInstUsesWith(I, V);
01927 
01928   if (Value *V = SimplifyOrInst(Op0, Op1, DL))
01929     return ReplaceInstUsesWith(I, V);
01930 
01931   // (A&B)|(A&C) -> A&(B|C) etc
01932   if (Value *V = SimplifyUsingDistributiveLaws(I))
01933     return ReplaceInstUsesWith(I, V);
01934 
01935   // See if we can simplify any instructions used by the instruction whose sole
01936   // purpose is to compute bits we don't care about.
01937   if (SimplifyDemandedInstructionBits(I))
01938     return &I;
01939 
01940   if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
01941     ConstantInt *C1 = nullptr; Value *X = nullptr;
01942     // (X & C1) | C2 --> (X | C2) & (C1|C2)
01943     // iff (C1 & C2) == 0.
01944     if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
01945         (RHS->getValue() & C1->getValue()) != 0 &&
01946         Op0->hasOneUse()) {
01947       Value *Or = Builder->CreateOr(X, RHS);
01948       Or->takeName(Op0);
01949       return BinaryOperator::CreateAnd(Or,
01950                              Builder->getInt(RHS->getValue() | C1->getValue()));
01951     }
01952 
01953     // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
01954     if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
01955         Op0->hasOneUse()) {
01956       Value *Or = Builder->CreateOr(X, RHS);
01957       Or->takeName(Op0);
01958       return BinaryOperator::CreateXor(Or,
01959                             Builder->getInt(C1->getValue() & ~RHS->getValue()));
01960     }
01961 
01962     // Try to fold constant and into select arguments.
01963     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
01964       if (Instruction *R = FoldOpIntoSelect(I, SI))
01965         return R;
01966 
01967     if (isa<PHINode>(Op0))
01968       if (Instruction *NV = FoldOpIntoPhi(I))
01969         return NV;
01970   }
01971 
01972   Value *A = nullptr, *B = nullptr;
01973   ConstantInt *C1 = nullptr, *C2 = nullptr;
01974 
01975   // (A | B) | C  and  A | (B | C)                  -> bswap if possible.
01976   // (A >> B) | (C << D)  and  (A << B) | (B >> C)  -> bswap if possible.
01977   if (match(Op0, m_Or(m_Value(), m_Value())) ||
01978       match(Op1, m_Or(m_Value(), m_Value())) ||
01979       (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
01980        match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
01981     if (Instruction *BSwap = MatchBSwap(I))
01982       return BSwap;
01983   }
01984 
01985   // (X^C)|Y -> (X|Y)^C iff Y&C == 0
01986   if (Op0->hasOneUse() &&
01987       match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
01988       MaskedValueIsZero(Op1, C1->getValue())) {
01989     Value *NOr = Builder->CreateOr(A, Op1);
01990     NOr->takeName(Op0);
01991     return BinaryOperator::CreateXor(NOr, C1);
01992   }
01993 
01994   // Y|(X^C) -> (X|Y)^C iff Y&C == 0
01995   if (Op1->hasOneUse() &&
01996       match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
01997       MaskedValueIsZero(Op0, C1->getValue())) {
01998     Value *NOr = Builder->CreateOr(A, Op0);
01999     NOr->takeName(Op0);
02000     return BinaryOperator::CreateXor(NOr, C1);
02001   }
02002 
02003   // ((~A & B) | A) -> (A | B)
02004   if (match(Op0, m_And(m_Not(m_Value(A)), m_Value(B))) &&
02005       match(Op1, m_Specific(A)))
02006     return BinaryOperator::CreateOr(A, B);
02007 
02008   // ((A & B) | ~A) -> (~A | B)
02009   if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
02010       match(Op1, m_Not(m_Specific(A))))
02011     return BinaryOperator::CreateOr(Builder->CreateNot(A), B);
02012 
02013   // (A & C)|(B & D)
02014   Value *C = nullptr, *D = nullptr;
02015   if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
02016       match(Op1, m_And(m_Value(B), m_Value(D)))) {
02017     Value *V1 = nullptr, *V2 = nullptr;
02018     C1 = dyn_cast<ConstantInt>(C);
02019     C2 = dyn_cast<ConstantInt>(D);
02020     if (C1 && C2) {  // (A & C1)|(B & C2)
02021       if ((C1->getValue() & C2->getValue()) == 0) {
02022         // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
02023         // iff (C1&C2) == 0 and (N&~C1) == 0
02024         if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
02025             ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) ||  // (V|N)
02026              (V2 == B && MaskedValueIsZero(V1, ~C1->getValue()))))   // (N|V)
02027           return BinaryOperator::CreateAnd(A,
02028                                 Builder->getInt(C1->getValue()|C2->getValue()));
02029         // Or commutes, try both ways.
02030         if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
02031             ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) ||  // (V|N)
02032              (V2 == A && MaskedValueIsZero(V1, ~C2->getValue()))))   // (N|V)
02033           return BinaryOperator::CreateAnd(B,
02034                                 Builder->getInt(C1->getValue()|C2->getValue()));
02035 
02036         // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
02037         // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
02038         ConstantInt *C3 = nullptr, *C4 = nullptr;
02039         if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
02040             (C3->getValue() & ~C1->getValue()) == 0 &&
02041             match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
02042             (C4->getValue() & ~C2->getValue()) == 0) {
02043           V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
02044           return BinaryOperator::CreateAnd(V2,
02045                                 Builder->getInt(C1->getValue()|C2->getValue()));
02046         }
02047       }
02048     }
02049 
02050     // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) ->  C0 ? A : B, and commuted variants.
02051     // Don't do this for vector select idioms, the code generator doesn't handle
02052     // them well yet.
02053     if (!I.getType()->isVectorTy()) {
02054       if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
02055         return Match;
02056       if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
02057         return Match;
02058       if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
02059         return Match;
02060       if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
02061         return Match;
02062     }
02063 
02064     // ((A&~B)|(~A&B)) -> A^B
02065     if ((match(C, m_Not(m_Specific(D))) &&
02066          match(B, m_Not(m_Specific(A)))))
02067       return BinaryOperator::CreateXor(A, D);
02068     // ((~B&A)|(~A&B)) -> A^B
02069     if ((match(A, m_Not(m_Specific(D))) &&
02070          match(B, m_Not(m_Specific(C)))))
02071       return BinaryOperator::CreateXor(C, D);
02072     // ((A&~B)|(B&~A)) -> A^B
02073     if ((match(C, m_Not(m_Specific(B))) &&
02074          match(D, m_Not(m_Specific(A)))))
02075       return BinaryOperator::CreateXor(A, B);
02076     // ((~B&A)|(B&~A)) -> A^B
02077     if ((match(A, m_Not(m_Specific(B))) &&
02078          match(D, m_Not(m_Specific(C)))))
02079       return BinaryOperator::CreateXor(C, B);
02080 
02081     // ((A|B)&1)|(B&-2) -> (A&1) | B
02082     if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
02083         match(A, m_Or(m_Specific(B), m_Value(V1)))) {
02084       Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
02085       if (Ret) return Ret;
02086     }
02087     // (B&-2)|((A|B)&1) -> (A&1) | B
02088     if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
02089         match(B, m_Or(m_Value(V1), m_Specific(A)))) {
02090       Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
02091       if (Ret) return Ret;
02092     }
02093   }
02094 
02095   // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
02096   if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
02097     if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
02098       if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
02099         return BinaryOperator::CreateOr(Op0, C);
02100 
02101   // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
02102   if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
02103     if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
02104       if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
02105         return BinaryOperator::CreateOr(Op1, C);
02106 
02107   // (X >> Z) | (Y >> Z)  -> (X|Y) >> Z  for all shifts.
02108   if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
02109     if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
02110       if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
02111           SI0->getOperand(1) == SI1->getOperand(1) &&
02112           (SI0->hasOneUse() || SI1->hasOneUse())) {
02113         Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
02114                                          SI0->getName());
02115         return BinaryOperator::Create(SI1->getOpcode(), NewOp,
02116                                       SI1->getOperand(1));
02117       }
02118   }
02119 
02120   // (~A | ~B) == (~(A & B)) - De Morgan's Law
02121   if (Value *Op0NotVal = dyn_castNotVal(Op0))
02122     if (Value *Op1NotVal = dyn_castNotVal(Op1))
02123       if (Op0->hasOneUse() && Op1->hasOneUse()) {
02124         Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
02125                                         I.getName()+".demorgan");
02126         return BinaryOperator::CreateNot(And);
02127       }
02128 
02129   // Canonicalize xor to the RHS.
02130   bool SwappedForXor = false;
02131   if (match(Op0, m_Xor(m_Value(), m_Value()))) {
02132     std::swap(Op0, Op1);
02133     SwappedForXor = true;
02134   }
02135 
02136   // A | ( A ^ B) -> A |  B
02137   // A | (~A ^ B) -> A | ~B
02138   // (A & B) | (A ^ B)
02139   if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
02140     if (Op0 == A || Op0 == B)
02141       return BinaryOperator::CreateOr(A, B);
02142 
02143     if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
02144         match(Op0, m_And(m_Specific(B), m_Specific(A))))
02145       return BinaryOperator::CreateOr(A, B);
02146 
02147     if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
02148       Value *Not = Builder->CreateNot(B, B->getName()+".not");
02149       return BinaryOperator::CreateOr(Not, Op0);
02150     }
02151     if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
02152       Value *Not = Builder->CreateNot(A, A->getName()+".not");
02153       return BinaryOperator::CreateOr(Not, Op0);
02154     }
02155   }
02156 
02157   // A | ~(A | B) -> A | ~B
02158   // A | ~(A ^ B) -> A | ~B
02159   if (match(Op1, m_Not(m_Value(A))))
02160     if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
02161       if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
02162           Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
02163                                B->getOpcode() == Instruction::Xor)) {
02164         Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
02165                                                  B->getOperand(0);
02166         Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
02167         return BinaryOperator::CreateOr(Not, Op0);
02168       }
02169 
02170   if (SwappedForXor)
02171     std::swap(Op0, Op1);
02172 
02173   if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
02174     if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
02175       if (Value *Res = FoldOrOfICmps(LHS, RHS))
02176         return ReplaceInstUsesWith(I, Res);
02177 
02178   // (fcmp uno x, c) | (fcmp uno y, c)  -> (fcmp uno x, y)
02179   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
02180     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
02181       if (Value *Res = FoldOrOfFCmps(LHS, RHS))
02182         return ReplaceInstUsesWith(I, Res);
02183 
02184   // fold (or (cast A), (cast B)) -> (cast (or A, B))
02185   if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
02186     CastInst *Op1C = dyn_cast<CastInst>(Op1);
02187     if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
02188       Type *SrcTy = Op0C->getOperand(0)->getType();
02189       if (SrcTy == Op1C->getOperand(0)->getType() &&
02190           SrcTy->isIntOrIntVectorTy()) {
02191         Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
02192 
02193         if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
02194             // Only do this if the casts both really cause code to be
02195             // generated.
02196             ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
02197             ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
02198           Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
02199           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
02200         }
02201 
02202         // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
02203         // cast is otherwise not optimizable.  This happens for vector sexts.
02204         if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
02205           if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
02206             if (Value *Res = FoldOrOfICmps(LHS, RHS))
02207               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
02208 
02209         // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
02210         // cast is otherwise not optimizable.  This happens for vector sexts.
02211         if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
02212           if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
02213             if (Value *Res = FoldOrOfFCmps(LHS, RHS))
02214               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
02215       }
02216     }
02217   }
02218 
02219   // or(sext(A), B) -> A ? -1 : B where A is an i1
02220   // or(A, sext(B)) -> B ? -1 : A where B is an i1
02221   if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
02222     return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
02223   if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
02224     return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
02225 
02226   // Note: If we've gotten to the point of visiting the outer OR, then the
02227   // inner one couldn't be simplified.  If it was a constant, then it won't
02228   // be simplified by a later pass either, so we try swapping the inner/outer
02229   // ORs in the hopes that we'll be able to simplify it this way.
02230   // (X|C) | V --> (X|V) | C
02231   if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
02232       match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
02233     Value *Inner = Builder->CreateOr(A, Op1);
02234     Inner->takeName(Op0);
02235     return BinaryOperator::CreateOr(Inner, C1);
02236   }
02237 
02238   // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
02239   // Since this OR statement hasn't been optimized further yet, we hope
02240   // that this transformation will allow the new ORs to be optimized.
02241   {
02242     Value *X = nullptr, *Y = nullptr;
02243     if (Op0->hasOneUse() && Op1->hasOneUse() &&
02244         match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
02245         match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
02246       Value *orTrue = Builder->CreateOr(A, C);
02247       Value *orFalse = Builder->CreateOr(B, D);
02248       return SelectInst::Create(X, orTrue, orFalse);
02249     }
02250   }
02251 
02252   return Changed ? &I : nullptr;
02253 }
02254 
02255 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
02256   bool Changed = SimplifyAssociativeOrCommutative(I);
02257   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
02258 
02259   if (Value *V = SimplifyVectorOp(I))
02260     return ReplaceInstUsesWith(I, V);
02261 
02262   if (Value *V = SimplifyXorInst(Op0, Op1, DL))
02263     return ReplaceInstUsesWith(I, V);
02264 
02265   // (A&B)^(A&C) -> A&(B^C) etc
02266   if (Value *V = SimplifyUsingDistributiveLaws(I))
02267     return ReplaceInstUsesWith(I, V);
02268 
02269   // See if we can simplify any instructions used by the instruction whose sole
02270   // purpose is to compute bits we don't care about.
02271   if (SimplifyDemandedInstructionBits(I))
02272     return &I;
02273 
02274   // Is this a ~ operation?
02275   if (Value *NotOp = dyn_castNotVal(&I)) {
02276     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
02277       if (Op0I->getOpcode() == Instruction::And ||
02278           Op0I->getOpcode() == Instruction::Or) {
02279         // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
02280         // ~(~X | Y) === (X & ~Y) - De Morgan's Law
02281         if (dyn_castNotVal(Op0I->getOperand(1)))
02282           Op0I->swapOperands();
02283         if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
02284           Value *NotY =
02285             Builder->CreateNot(Op0I->getOperand(1),
02286                                Op0I->getOperand(1)->getName()+".not");
02287           if (Op0I->getOpcode() == Instruction::And)
02288             return BinaryOperator::CreateOr(Op0NotVal, NotY);
02289           return BinaryOperator::CreateAnd(Op0NotVal, NotY);
02290         }
02291 
02292         // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
02293         // ~(X | Y) === (~X & ~Y) - De Morgan's Law
02294         if (isFreeToInvert(Op0I->getOperand(0)) &&
02295             isFreeToInvert(Op0I->getOperand(1))) {
02296           Value *NotX =
02297             Builder->CreateNot(Op0I->getOperand(0), "notlhs");
02298           Value *NotY =
02299             Builder->CreateNot(Op0I->getOperand(1), "notrhs");
02300           if (Op0I->getOpcode() == Instruction::And)
02301             return BinaryOperator::CreateOr(NotX, NotY);
02302           return BinaryOperator::CreateAnd(NotX, NotY);
02303         }
02304 
02305       } else if (Op0I->getOpcode() == Instruction::AShr) {
02306         // ~(~X >>s Y) --> (X >>s Y)
02307         if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
02308           return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
02309       }
02310     }
02311   }
02312 
02313 
02314   if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
02315     if (RHS->isOne() && Op0->hasOneUse())
02316       // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
02317       if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
02318         return CmpInst::Create(CI->getOpcode(),
02319                                CI->getInversePredicate(),
02320                                CI->getOperand(0), CI->getOperand(1));
02321 
02322     // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
02323     if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
02324       if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
02325         if (CI->hasOneUse() && Op0C->hasOneUse()) {
02326           Instruction::CastOps Opcode = Op0C->getOpcode();
02327           if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
02328               (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
02329                                             Op0C->getDestTy()))) {
02330             CI->setPredicate(CI->getInversePredicate());
02331             return CastInst::Create(Opcode, CI, Op0C->getType());
02332           }
02333         }
02334       }
02335     }
02336 
02337     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
02338       // ~(c-X) == X-c-1 == X+(-c-1)
02339       if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
02340         if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
02341           Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
02342           Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
02343                                       ConstantInt::get(I.getType(), 1));
02344           return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
02345         }
02346 
02347       if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
02348         if (Op0I->getOpcode() == Instruction::Add) {
02349           // ~(X-c) --> (-c-1)-X
02350           if (RHS->isAllOnesValue()) {
02351             Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
02352             return BinaryOperator::CreateSub(
02353                            ConstantExpr::getSub(NegOp0CI,
02354                                       ConstantInt::get(I.getType(), 1)),
02355                                       Op0I->getOperand(0));
02356           } else if (RHS->getValue().isSignBit()) {
02357             // (X + C) ^ signbit -> (X + C + signbit)
02358             Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
02359             return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
02360 
02361           }
02362         } else if (Op0I->getOpcode() == Instruction::Or) {
02363           // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
02364           if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
02365             Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
02366             // Anything in both C1 and C2 is known to be zero, remove it from
02367             // NewRHS.
02368             Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
02369             NewRHS = ConstantExpr::getAnd(NewRHS,
02370                                        ConstantExpr::getNot(CommonBits));
02371             Worklist.Add(Op0I);
02372             I.setOperand(0, Op0I->getOperand(0));
02373             I.setOperand(1, NewRHS);
02374             return &I;
02375           }
02376         } else if (Op0I->getOpcode() == Instruction::LShr) {
02377           // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
02378           // E1 = "X ^ C1"
02379           BinaryOperator *E1;
02380           ConstantInt *C1;
02381           if (Op0I->hasOneUse() &&
02382               (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
02383               E1->getOpcode() == Instruction::Xor &&
02384               (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
02385             // fold (C1 >> C2) ^ C3
02386             ConstantInt *C2 = Op0CI, *C3 = RHS;
02387             APInt FoldConst = C1->getValue().lshr(C2->getValue());
02388             FoldConst ^= C3->getValue();
02389             // Prepare the two operands.
02390             Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
02391             Opnd0->takeName(Op0I);
02392             cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
02393             Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
02394 
02395             return BinaryOperator::CreateXor(Opnd0, FoldVal);
02396           }
02397         }
02398       }
02399     }
02400 
02401     // Try to fold constant and into select arguments.
02402     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
02403       if (Instruction *R = FoldOpIntoSelect(I, SI))
02404         return R;
02405     if (isa<PHINode>(Op0))
02406       if (Instruction *NV = FoldOpIntoPhi(I))
02407         return NV;
02408   }
02409 
02410   BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
02411   if (Op1I) {
02412     Value *A, *B;
02413     if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
02414       if (A == Op0) {              // B^(B|A) == (A|B)^B
02415         Op1I->swapOperands();
02416         I.swapOperands();
02417         std::swap(Op0, Op1);
02418       } else if (B == Op0) {       // B^(A|B) == (A|B)^B
02419         I.swapOperands();     // Simplified below.
02420         std::swap(Op0, Op1);
02421       }
02422     } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
02423                Op1I->hasOneUse()){
02424       if (A == Op0) {                                      // A^(A&B) -> A^(B&A)
02425         Op1I->swapOperands();
02426         std::swap(A, B);
02427       }
02428       if (B == Op0) {                                      // A^(B&A) -> (B&A)^A
02429         I.swapOperands();     // Simplified below.
02430         std::swap(Op0, Op1);
02431       }
02432     }
02433   }
02434 
02435   BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
02436   if (Op0I) {
02437     Value *A, *B;
02438     if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
02439         Op0I->hasOneUse()) {
02440       if (A == Op1)                                  // (B|A)^B == (A|B)^B
02441         std::swap(A, B);
02442       if (B == Op1)                                  // (A|B)^B == A & ~B
02443         return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
02444     } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
02445                Op0I->hasOneUse()){
02446       if (A == Op1)                                        // (A&B)^A -> (B&A)^A
02447         std::swap(A, B);
02448       if (B == Op1 &&                                      // (B&A)^A == ~B & A
02449           !isa<ConstantInt>(Op1)) {  // Canonical form is (B&C)^C
02450         return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
02451       }
02452     }
02453   }
02454 
02455   // (X >> Z) ^ (Y >> Z)  -> (X^Y) >> Z  for all shifts.
02456   if (Op0I && Op1I && Op0I->isShift() &&
02457       Op0I->getOpcode() == Op1I->getOpcode() &&
02458       Op0I->getOperand(1) == Op1I->getOperand(1) &&
02459       (Op0I->hasOneUse() || Op1I->hasOneUse())) {
02460     Value *NewOp =
02461       Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
02462                          Op0I->getName());
02463     return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
02464                                   Op1I->getOperand(1));
02465   }
02466 
02467   if (Op0I && Op1I) {
02468     Value *A, *B, *C, *D;
02469     // (A & B)^(A | B) -> A ^ B
02470     if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
02471         match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
02472       if ((A == C && B == D) || (A == D && B == C))
02473         return BinaryOperator::CreateXor(A, B);
02474     }
02475     // (A | B)^(A & B) -> A ^ B
02476     if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
02477         match(Op1I, m_And(m_Value(C), m_Value(D)))) {
02478       if ((A == C && B == D) || (A == D && B == C))
02479         return BinaryOperator::CreateXor(A, B);
02480     }
02481     // (A & B) ^ (A ^ B) -> (A | B)
02482     if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
02483         match(Op1I, m_Xor(m_Specific(A), m_Specific(B))))
02484       return BinaryOperator::CreateOr(A, B);
02485     // (A ^ B) ^ (A & B) -> (A | B)
02486     if (match(Op0I, m_Xor(m_Value(A), m_Value(B))) &&
02487         match(Op1I, m_And(m_Specific(A), m_Specific(B))))
02488       return BinaryOperator::CreateOr(A, B);
02489   }
02490 
02491   // (A | B)^(~A) -> (A | ~B)
02492   Value *A = nullptr, *B = nullptr;
02493   if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
02494       match(Op1, m_Not(m_Specific(A))))
02495     return BinaryOperator::CreateOr(A, Builder->CreateNot(B));
02496 
02497   // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
02498   if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
02499     if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
02500       if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
02501         if (LHS->getOperand(0) == RHS->getOperand(1) &&
02502             LHS->getOperand(1) == RHS->getOperand(0))
02503           LHS->swapOperands();
02504         if (LHS->getOperand(0) == RHS->getOperand(0) &&
02505             LHS->getOperand(1) == RHS->getOperand(1)) {
02506           Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
02507           unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
02508           bool isSigned = LHS->isSigned() || RHS->isSigned();
02509           return ReplaceInstUsesWith(I,
02510                                getNewICmpValue(isSigned, Code, Op0, Op1,
02511                                                Builder));
02512         }
02513       }
02514 
02515   // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
02516   if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
02517     if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
02518       if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
02519         Type *SrcTy = Op0C->getOperand(0)->getType();
02520         if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
02521             // Only do this if the casts both really cause code to be generated.
02522             ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
02523                                I.getType()) &&
02524             ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
02525                                I.getType())) {
02526           Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
02527                                             Op1C->getOperand(0), I.getName());
02528           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
02529         }
02530       }
02531   }
02532 
02533   return Changed ? &I : nullptr;
02534 }