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 
01288   if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
01289     if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
01290       if (Value *Res = FoldAndOfICmps(LHS, RHS))
01291         return ReplaceInstUsesWith(I, Res);
01292 
01293   // If and'ing two fcmp, try combine them into one.
01294   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
01295     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
01296       if (Value *Res = FoldAndOfFCmps(LHS, RHS))
01297         return ReplaceInstUsesWith(I, Res);
01298 
01299 
01300   // fold (and (cast A), (cast B)) -> (cast (and A, B))
01301   if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
01302     if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
01303       Type *SrcTy = Op0C->getOperand(0)->getType();
01304       if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
01305           SrcTy == Op1C->getOperand(0)->getType() &&
01306           SrcTy->isIntOrIntVectorTy()) {
01307         Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
01308 
01309         // Only do this if the casts both really cause code to be generated.
01310         if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
01311             ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
01312           Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
01313           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
01314         }
01315 
01316         // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
01317         // cast is otherwise not optimizable.  This happens for vector sexts.
01318         if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
01319           if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
01320             if (Value *Res = FoldAndOfICmps(LHS, RHS))
01321               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
01322 
01323         // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
01324         // cast is otherwise not optimizable.  This happens for vector sexts.
01325         if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
01326           if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
01327             if (Value *Res = FoldAndOfFCmps(LHS, RHS))
01328               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
01329       }
01330     }
01331 
01332   // (X >> Z) & (Y >> Z)  -> (X&Y) >> Z  for all shifts.
01333   if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
01334     if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
01335       if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
01336           SI0->getOperand(1) == SI1->getOperand(1) &&
01337           (SI0->hasOneUse() || SI1->hasOneUse())) {
01338         Value *NewOp =
01339           Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
01340                              SI0->getName());
01341         return BinaryOperator::Create(SI1->getOpcode(), NewOp,
01342                                       SI1->getOperand(1));
01343       }
01344   }
01345 
01346   {
01347     Value *X = nullptr;
01348     bool OpsSwapped = false;
01349     // Canonicalize SExt or Not to the LHS
01350     if (match(Op1, m_SExt(m_Value())) ||
01351         match(Op1, m_Not(m_Value()))) {
01352       std::swap(Op0, Op1);
01353       OpsSwapped = true;
01354     }
01355 
01356     // Fold (and (sext bool to A), B) --> (select bool, B, 0)
01357     if (match(Op0, m_SExt(m_Value(X))) &&
01358         X->getType()->getScalarType()->isIntegerTy(1)) {
01359       Value *Zero = Constant::getNullValue(Op1->getType());
01360       return SelectInst::Create(X, Op1, Zero);
01361     }
01362 
01363     // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
01364     if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
01365         X->getType()->getScalarType()->isIntegerTy(1)) {
01366       Value *Zero = Constant::getNullValue(Op0->getType());
01367       return SelectInst::Create(X, Zero, Op1);
01368     }
01369 
01370     if (OpsSwapped)
01371       std::swap(Op0, Op1);
01372   }
01373 
01374   return Changed ? &I : nullptr;
01375 }
01376 
01377 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
01378 /// capable of providing pieces of a bswap.  The subexpression provides pieces
01379 /// of a bswap if it is proven that each of the non-zero bytes in the output of
01380 /// the expression came from the corresponding "byte swapped" byte in some other
01381 /// value.  For example, if the current subexpression is "(shl i32 %X, 24)" then
01382 /// we know that the expression deposits the low byte of %X into the high byte
01383 /// of the bswap result and that all other bytes are zero.  This expression is
01384 /// accepted, the high byte of ByteValues is set to X to indicate a correct
01385 /// match.
01386 ///
01387 /// This function returns true if the match was unsuccessful and false if so.
01388 /// On entry to the function the "OverallLeftShift" is a signed integer value
01389 /// indicating the number of bytes that the subexpression is later shifted.  For
01390 /// example, if the expression is later right shifted by 16 bits, the
01391 /// OverallLeftShift value would be -2 on entry.  This is used to specify which
01392 /// byte of ByteValues is actually being set.
01393 ///
01394 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
01395 /// byte is masked to zero by a user.  For example, in (X & 255), X will be
01396 /// processed with a bytemask of 1.  Because bytemask is 32-bits, this limits
01397 /// this function to working on up to 32-byte (256 bit) values.  ByteMask is
01398 /// always in the local (OverallLeftShift) coordinate space.
01399 ///
01400 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
01401                               SmallVectorImpl<Value *> &ByteValues) {
01402   if (Instruction *I = dyn_cast<Instruction>(V)) {
01403     // If this is an or instruction, it may be an inner node of the bswap.
01404     if (I->getOpcode() == Instruction::Or) {
01405       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
01406                                ByteValues) ||
01407              CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
01408                                ByteValues);
01409     }
01410 
01411     // If this is a logical shift by a constant multiple of 8, recurse with
01412     // OverallLeftShift and ByteMask adjusted.
01413     if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
01414       unsigned ShAmt =
01415         cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
01416       // Ensure the shift amount is defined and of a byte value.
01417       if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
01418         return true;
01419 
01420       unsigned ByteShift = ShAmt >> 3;
01421       if (I->getOpcode() == Instruction::Shl) {
01422         // X << 2 -> collect(X, +2)
01423         OverallLeftShift += ByteShift;
01424         ByteMask >>= ByteShift;
01425       } else {
01426         // X >>u 2 -> collect(X, -2)
01427         OverallLeftShift -= ByteShift;
01428         ByteMask <<= ByteShift;
01429         ByteMask &= (~0U >> (32-ByteValues.size()));
01430       }
01431 
01432       if (OverallLeftShift >= (int)ByteValues.size()) return true;
01433       if (OverallLeftShift <= -(int)ByteValues.size()) return true;
01434 
01435       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
01436                                ByteValues);
01437     }
01438 
01439     // If this is a logical 'and' with a mask that clears bytes, clear the
01440     // corresponding bytes in ByteMask.
01441     if (I->getOpcode() == Instruction::And &&
01442         isa<ConstantInt>(I->getOperand(1))) {
01443       // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
01444       unsigned NumBytes = ByteValues.size();
01445       APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
01446       const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
01447 
01448       for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
01449         // If this byte is masked out by a later operation, we don't care what
01450         // the and mask is.
01451         if ((ByteMask & (1 << i)) == 0)
01452           continue;
01453 
01454         // If the AndMask is all zeros for this byte, clear the bit.
01455         APInt MaskB = AndMask & Byte;
01456         if (MaskB == 0) {
01457           ByteMask &= ~(1U << i);
01458           continue;
01459         }
01460 
01461         // If the AndMask is not all ones for this byte, it's not a bytezap.
01462         if (MaskB != Byte)
01463           return true;
01464 
01465         // Otherwise, this byte is kept.
01466       }
01467 
01468       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
01469                                ByteValues);
01470     }
01471   }
01472 
01473   // Okay, we got to something that isn't a shift, 'or' or 'and'.  This must be
01474   // the input value to the bswap.  Some observations: 1) if more than one byte
01475   // is demanded from this input, then it could not be successfully assembled
01476   // into a byteswap.  At least one of the two bytes would not be aligned with
01477   // their ultimate destination.
01478   if (!isPowerOf2_32(ByteMask)) return true;
01479   unsigned InputByteNo = countTrailingZeros(ByteMask);
01480 
01481   // 2) The input and ultimate destinations must line up: if byte 3 of an i32
01482   // is demanded, it needs to go into byte 0 of the result.  This means that the
01483   // byte needs to be shifted until it lands in the right byte bucket.  The
01484   // shift amount depends on the position: if the byte is coming from the high
01485   // part of the value (e.g. byte 3) then it must be shifted right.  If from the
01486   // low part, it must be shifted left.
01487   unsigned DestByteNo = InputByteNo + OverallLeftShift;
01488   if (ByteValues.size()-1-DestByteNo != InputByteNo)
01489     return true;
01490 
01491   // If the destination byte value is already defined, the values are or'd
01492   // together, which isn't a bswap (unless it's an or of the same bits).
01493   if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
01494     return true;
01495   ByteValues[DestByteNo] = V;
01496   return false;
01497 }
01498 
01499 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
01500 /// If so, insert the new bswap intrinsic and return it.
01501 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
01502   IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
01503   if (!ITy || ITy->getBitWidth() % 16 ||
01504       // ByteMask only allows up to 32-byte values.
01505       ITy->getBitWidth() > 32*8)
01506     return nullptr;   // Can only bswap pairs of bytes.  Can't do vectors.
01507 
01508   /// ByteValues - For each byte of the result, we keep track of which value
01509   /// defines each byte.
01510   SmallVector<Value*, 8> ByteValues;
01511   ByteValues.resize(ITy->getBitWidth()/8);
01512 
01513   // Try to find all the pieces corresponding to the bswap.
01514   uint32_t ByteMask = ~0U >> (32-ByteValues.size());
01515   if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
01516     return nullptr;
01517 
01518   // Check to see if all of the bytes come from the same value.
01519   Value *V = ByteValues[0];
01520   if (!V) return nullptr;  // Didn't find a byte?  Must be zero.
01521 
01522   // Check to make sure that all of the bytes come from the same value.
01523   for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
01524     if (ByteValues[i] != V)
01525       return nullptr;
01526   Module *M = I.getParent()->getParent()->getParent();
01527   Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
01528   return CallInst::Create(F, V);
01529 }
01530 
01531 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D).  Check
01532 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
01533 /// we can simplify this expression to "cond ? C : D or B".
01534 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
01535                                          Value *C, Value *D) {
01536   // If A is not a select of -1/0, this cannot match.
01537   Value *Cond = nullptr;
01538   if (!match(A, m_SExt(m_Value(Cond))) ||
01539       !Cond->getType()->isIntegerTy(1))
01540     return nullptr;
01541 
01542   // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
01543   if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
01544     return SelectInst::Create(Cond, C, B);
01545   if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
01546     return SelectInst::Create(Cond, C, B);
01547 
01548   // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
01549   if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
01550     return SelectInst::Create(Cond, C, D);
01551   if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
01552     return SelectInst::Create(Cond, C, D);
01553   return nullptr;
01554 }
01555 
01556 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
01557 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
01558   ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
01559 
01560   // Fold (iszero(A & K1) | iszero(A & K2)) ->  (A & (K1 | K2)) != (K1 | K2)
01561   // if K1 and K2 are a one-bit mask.
01562   ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
01563   ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
01564 
01565   if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() &&
01566       RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
01567 
01568     BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0));
01569     BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0));
01570     if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() &&
01571         LAnd->getOpcode() == Instruction::And &&
01572         RAnd->getOpcode() == Instruction::And) {
01573 
01574       Value *Mask = nullptr;
01575       Value *Masked = nullptr;
01576       if (LAnd->getOperand(0) == RAnd->getOperand(0) &&
01577           isKnownToBeAPowerOfTwo(LAnd->getOperand(1)) &&
01578           isKnownToBeAPowerOfTwo(RAnd->getOperand(1))) {
01579         Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1));
01580         Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask);
01581       } else if (LAnd->getOperand(1) == RAnd->getOperand(1) &&
01582                  isKnownToBeAPowerOfTwo(LAnd->getOperand(0)) &&
01583                  isKnownToBeAPowerOfTwo(RAnd->getOperand(0))) {
01584         Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0));
01585         Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask);
01586       }
01587 
01588       if (Masked)
01589         return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask);
01590     }
01591   }
01592 
01593   // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
01594   if (PredicatesFoldable(LHSCC, RHSCC)) {
01595     if (LHS->getOperand(0) == RHS->getOperand(1) &&
01596         LHS->getOperand(1) == RHS->getOperand(0))
01597       LHS->swapOperands();
01598     if (LHS->getOperand(0) == RHS->getOperand(0) &&
01599         LHS->getOperand(1) == RHS->getOperand(1)) {
01600       Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
01601       unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
01602       bool isSigned = LHS->isSigned() || RHS->isSigned();
01603       return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
01604     }
01605   }
01606 
01607   // handle (roughly):
01608   // (icmp ne (A & B), C) | (icmp ne (A & D), E)
01609   if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
01610     return V;
01611 
01612   Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
01613   if (LHS->hasOneUse() || RHS->hasOneUse()) {
01614     // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
01615     // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
01616     Value *A = nullptr, *B = nullptr;
01617     if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
01618       B = Val;
01619       if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
01620         A = Val2;
01621       else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
01622         A = RHS->getOperand(1);
01623     }
01624     // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
01625     // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
01626     else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
01627       B = Val2;
01628       if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
01629         A = Val;
01630       else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
01631         A = LHS->getOperand(1);
01632     }
01633     if (A && B)
01634       return Builder->CreateICmp(
01635           ICmpInst::ICMP_UGE,
01636           Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
01637   }
01638 
01639   // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
01640   if (!LHSCst || !RHSCst) return nullptr;
01641 
01642   if (LHSCst == RHSCst && LHSCC == RHSCC) {
01643     // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
01644     if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
01645       Value *NewOr = Builder->CreateOr(Val, Val2);
01646       return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
01647     }
01648   }
01649 
01650   // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
01651   //   iff C2 + CA == C1.
01652   if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
01653     ConstantInt *AddCst;
01654     if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
01655       if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
01656         return Builder->CreateICmpULE(Val, LHSCst);
01657   }
01658 
01659   // From here on, we only handle:
01660   //    (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
01661   if (Val != Val2) return nullptr;
01662 
01663   // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
01664   if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
01665       RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
01666       LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
01667       RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
01668     return nullptr;
01669 
01670   // We can't fold (ugt x, C) | (sgt x, C2).
01671   if (!PredicatesFoldable(LHSCC, RHSCC))
01672     return nullptr;
01673 
01674   // Ensure that the larger constant is on the RHS.
01675   bool ShouldSwap;
01676   if (CmpInst::isSigned(LHSCC) ||
01677       (ICmpInst::isEquality(LHSCC) &&
01678        CmpInst::isSigned(RHSCC)))
01679     ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
01680   else
01681     ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
01682 
01683   if (ShouldSwap) {
01684     std::swap(LHS, RHS);
01685     std::swap(LHSCst, RHSCst);
01686     std::swap(LHSCC, RHSCC);
01687   }
01688 
01689   // At this point, we know we have two icmp instructions
01690   // comparing a value against two constants and or'ing the result
01691   // together.  Because of the above check, we know that we only have
01692   // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
01693   // icmp folding check above), that the two constants are not
01694   // equal.
01695   assert(LHSCst != RHSCst && "Compares not folded above?");
01696 
01697   switch (LHSCC) {
01698   default: llvm_unreachable("Unknown integer condition code!");
01699   case ICmpInst::ICMP_EQ:
01700     switch (RHSCC) {
01701     default: llvm_unreachable("Unknown integer condition code!");
01702     case ICmpInst::ICMP_EQ:
01703       if (LHS->getOperand(0) == RHS->getOperand(0)) {
01704         // if LHSCst and RHSCst differ only by one bit:
01705         // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
01706         assert(LHSCst->getValue().ule(LHSCst->getValue()));
01707 
01708         APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
01709         if (Xor.isPowerOf2()) {
01710           Value *NegCst = Builder->getInt(~Xor);
01711           Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
01712           return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
01713         }
01714       }
01715 
01716       if (LHSCst == SubOne(RHSCst)) {
01717         // (X == 13 | X == 14) -> X-13 <u 2
01718         Constant *AddCST = ConstantExpr::getNeg(LHSCst);
01719         Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
01720         AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
01721         return Builder->CreateICmpULT(Add, AddCST);
01722       }
01723 
01724       break;                         // (X == 13 | X == 15) -> no change
01725     case ICmpInst::ICMP_UGT:         // (X == 13 | X u> 14) -> no change
01726     case ICmpInst::ICMP_SGT:         // (X == 13 | X s> 14) -> no change
01727       break;
01728     case ICmpInst::ICMP_NE:          // (X == 13 | X != 15) -> X != 15
01729     case ICmpInst::ICMP_ULT:         // (X == 13 | X u< 15) -> X u< 15
01730     case ICmpInst::ICMP_SLT:         // (X == 13 | X s< 15) -> X s< 15
01731       return RHS;
01732     }
01733     break;
01734   case ICmpInst::ICMP_NE:
01735     switch (RHSCC) {
01736     default: llvm_unreachable("Unknown integer condition code!");
01737     case ICmpInst::ICMP_EQ:          // (X != 13 | X == 15) -> X != 13
01738     case ICmpInst::ICMP_UGT:         // (X != 13 | X u> 15) -> X != 13
01739     case ICmpInst::ICMP_SGT:         // (X != 13 | X s> 15) -> X != 13
01740       return LHS;
01741     case ICmpInst::ICMP_NE:          // (X != 13 | X != 15) -> true
01742     case ICmpInst::ICMP_ULT:         // (X != 13 | X u< 15) -> true
01743     case ICmpInst::ICMP_SLT:         // (X != 13 | X s< 15) -> true
01744       return Builder->getTrue();
01745     }
01746   case ICmpInst::ICMP_ULT:
01747     switch (RHSCC) {
01748     default: llvm_unreachable("Unknown integer condition code!");
01749     case ICmpInst::ICMP_EQ:         // (X u< 13 | X == 14) -> no change
01750       break;
01751     case ICmpInst::ICMP_UGT:        // (X u< 13 | X u> 15) -> (X-13) u> 2
01752       // If RHSCst is [us]MAXINT, it is always false.  Not handling
01753       // this can cause overflow.
01754       if (RHSCst->isMaxValue(false))
01755         return LHS;
01756       return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
01757     case ICmpInst::ICMP_SGT:        // (X u< 13 | X s> 15) -> no change
01758       break;
01759     case ICmpInst::ICMP_NE:         // (X u< 13 | X != 15) -> X != 15
01760     case ICmpInst::ICMP_ULT:        // (X u< 13 | X u< 15) -> X u< 15
01761       return RHS;
01762     case ICmpInst::ICMP_SLT:        // (X u< 13 | X s< 15) -> no change
01763       break;
01764     }
01765     break;
01766   case ICmpInst::ICMP_SLT:
01767     switch (RHSCC) {
01768     default: llvm_unreachable("Unknown integer condition code!");
01769     case ICmpInst::ICMP_EQ:         // (X s< 13 | X == 14) -> no change
01770       break;
01771     case ICmpInst::ICMP_SGT:        // (X s< 13 | X s> 15) -> (X-13) s> 2
01772       // If RHSCst is [us]MAXINT, it is always false.  Not handling
01773       // this can cause overflow.
01774       if (RHSCst->isMaxValue(true))
01775         return LHS;
01776       return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
01777     case ICmpInst::ICMP_UGT:        // (X s< 13 | X u> 15) -> no change
01778       break;
01779     case ICmpInst::ICMP_NE:         // (X s< 13 | X != 15) -> X != 15
01780     case ICmpInst::ICMP_SLT:        // (X s< 13 | X s< 15) -> X s< 15
01781       return RHS;
01782     case ICmpInst::ICMP_ULT:        // (X s< 13 | X u< 15) -> no change
01783       break;
01784     }
01785     break;
01786   case ICmpInst::ICMP_UGT:
01787     switch (RHSCC) {
01788     default: llvm_unreachable("Unknown integer condition code!");
01789     case ICmpInst::ICMP_EQ:         // (X u> 13 | X == 15) -> X u> 13
01790     case ICmpInst::ICMP_UGT:        // (X u> 13 | X u> 15) -> X u> 13
01791       return LHS;
01792     case ICmpInst::ICMP_SGT:        // (X u> 13 | X s> 15) -> no change
01793       break;
01794     case ICmpInst::ICMP_NE:         // (X u> 13 | X != 15) -> true
01795     case ICmpInst::ICMP_ULT:        // (X u> 13 | X u< 15) -> true
01796       return Builder->getTrue();
01797     case ICmpInst::ICMP_SLT:        // (X u> 13 | X s< 15) -> no change
01798       break;
01799     }
01800     break;
01801   case ICmpInst::ICMP_SGT:
01802     switch (RHSCC) {
01803     default: llvm_unreachable("Unknown integer condition code!");
01804     case ICmpInst::ICMP_EQ:         // (X s> 13 | X == 15) -> X > 13
01805     case ICmpInst::ICMP_SGT:        // (X s> 13 | X s> 15) -> X > 13
01806       return LHS;
01807     case ICmpInst::ICMP_UGT:        // (X s> 13 | X u> 15) -> no change
01808       break;
01809     case ICmpInst::ICMP_NE:         // (X s> 13 | X != 15) -> true
01810     case ICmpInst::ICMP_SLT:        // (X s> 13 | X s< 15) -> true
01811       return Builder->getTrue();
01812     case ICmpInst::ICMP_ULT:        // (X s> 13 | X u< 15) -> no change
01813       break;
01814     }
01815     break;
01816   }
01817   return nullptr;
01818 }
01819 
01820 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp).  NOTE: Unlike the rest of
01821 /// instcombine, this returns a Value which should already be inserted into the
01822 /// function.
01823 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
01824   if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
01825       RHS->getPredicate() == FCmpInst::FCMP_UNO &&
01826       LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
01827     if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
01828       if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
01829         // If either of the constants are nans, then the whole thing returns
01830         // true.
01831         if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
01832           return Builder->getTrue();
01833 
01834         // Otherwise, no need to compare the two constants, compare the
01835         // rest.
01836         return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
01837       }
01838 
01839     // Handle vector zeros.  This occurs because the canonical form of
01840     // "fcmp uno x,x" is "fcmp uno x, 0".
01841     if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
01842         isa<ConstantAggregateZero>(RHS->getOperand(1)))
01843       return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
01844 
01845     return nullptr;
01846   }
01847 
01848   Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
01849   Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
01850   FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
01851 
01852   if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
01853     // Swap RHS operands to match LHS.
01854     Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
01855     std::swap(Op1LHS, Op1RHS);
01856   }
01857   if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
01858     // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
01859     if (Op0CC == Op1CC)
01860       return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
01861     if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
01862       return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
01863     if (Op0CC == FCmpInst::FCMP_FALSE)
01864       return RHS;
01865     if (Op1CC == FCmpInst::FCMP_FALSE)
01866       return LHS;
01867     bool Op0Ordered;
01868     bool Op1Ordered;
01869     unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
01870     unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
01871     if (Op0Ordered == Op1Ordered) {
01872       // If both are ordered or unordered, return a new fcmp with
01873       // or'ed predicates.
01874       return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
01875     }
01876   }
01877   return nullptr;
01878 }
01879 
01880 /// FoldOrWithConstants - This helper function folds:
01881 ///
01882 ///     ((A | B) & C1) | (B & C2)
01883 ///
01884 /// into:
01885 ///
01886 ///     (A & C1) | B
01887 ///
01888 /// when the XOR of the two constants is "all ones" (-1).
01889 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
01890                                                Value *A, Value *B, Value *C) {
01891   ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
01892   if (!CI1) return nullptr;
01893 
01894   Value *V1 = nullptr;
01895   ConstantInt *CI2 = nullptr;
01896   if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr;
01897 
01898   APInt Xor = CI1->getValue() ^ CI2->getValue();
01899   if (!Xor.isAllOnesValue()) return nullptr;
01900 
01901   if (V1 == A || V1 == B) {
01902     Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
01903     return BinaryOperator::CreateOr(NewOp, V1);
01904   }
01905 
01906   return nullptr;
01907 }
01908 
01909 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
01910   bool Changed = SimplifyAssociativeOrCommutative(I);
01911   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
01912 
01913   if (Value *V = SimplifyVectorOp(I))
01914     return ReplaceInstUsesWith(I, V);
01915 
01916   if (Value *V = SimplifyOrInst(Op0, Op1, DL))
01917     return ReplaceInstUsesWith(I, V);
01918 
01919   // (A&B)|(A&C) -> A&(B|C) etc
01920   if (Value *V = SimplifyUsingDistributiveLaws(I))
01921     return ReplaceInstUsesWith(I, V);
01922 
01923   // See if we can simplify any instructions used by the instruction whose sole
01924   // purpose is to compute bits we don't care about.
01925   if (SimplifyDemandedInstructionBits(I))
01926     return &I;
01927 
01928   if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
01929     ConstantInt *C1 = nullptr; Value *X = nullptr;
01930     // (X & C1) | C2 --> (X | C2) & (C1|C2)
01931     // iff (C1 & C2) == 0.
01932     if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
01933         (RHS->getValue() & C1->getValue()) != 0 &&
01934         Op0->hasOneUse()) {
01935       Value *Or = Builder->CreateOr(X, RHS);
01936       Or->takeName(Op0);
01937       return BinaryOperator::CreateAnd(Or,
01938                              Builder->getInt(RHS->getValue() | C1->getValue()));
01939     }
01940 
01941     // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
01942     if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
01943         Op0->hasOneUse()) {
01944       Value *Or = Builder->CreateOr(X, RHS);
01945       Or->takeName(Op0);
01946       return BinaryOperator::CreateXor(Or,
01947                             Builder->getInt(C1->getValue() & ~RHS->getValue()));
01948     }
01949 
01950     // Try to fold constant and into select arguments.
01951     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
01952       if (Instruction *R = FoldOpIntoSelect(I, SI))
01953         return R;
01954 
01955     if (isa<PHINode>(Op0))
01956       if (Instruction *NV = FoldOpIntoPhi(I))
01957         return NV;
01958   }
01959 
01960   Value *A = nullptr, *B = nullptr;
01961   ConstantInt *C1 = nullptr, *C2 = nullptr;
01962 
01963   // (A | B) | C  and  A | (B | C)                  -> bswap if possible.
01964   // (A >> B) | (C << D)  and  (A << B) | (B >> C)  -> bswap if possible.
01965   if (match(Op0, m_Or(m_Value(), m_Value())) ||
01966       match(Op1, m_Or(m_Value(), m_Value())) ||
01967       (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
01968        match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
01969     if (Instruction *BSwap = MatchBSwap(I))
01970       return BSwap;
01971   }
01972 
01973   // (X^C)|Y -> (X|Y)^C iff Y&C == 0
01974   if (Op0->hasOneUse() &&
01975       match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
01976       MaskedValueIsZero(Op1, C1->getValue())) {
01977     Value *NOr = Builder->CreateOr(A, Op1);
01978     NOr->takeName(Op0);
01979     return BinaryOperator::CreateXor(NOr, C1);
01980   }
01981 
01982   // Y|(X^C) -> (X|Y)^C iff Y&C == 0
01983   if (Op1->hasOneUse() &&
01984       match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
01985       MaskedValueIsZero(Op0, C1->getValue())) {
01986     Value *NOr = Builder->CreateOr(A, Op0);
01987     NOr->takeName(Op0);
01988     return BinaryOperator::CreateXor(NOr, C1);
01989   }
01990 
01991   // (A & C)|(B & D)
01992   Value *C = nullptr, *D = nullptr;
01993   if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
01994       match(Op1, m_And(m_Value(B), m_Value(D)))) {
01995     Value *V1 = nullptr, *V2 = nullptr;
01996     C1 = dyn_cast<ConstantInt>(C);
01997     C2 = dyn_cast<ConstantInt>(D);
01998     if (C1 && C2) {  // (A & C1)|(B & C2)
01999       if ((C1->getValue() & C2->getValue()) == 0) {
02000         // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
02001         // iff (C1&C2) == 0 and (N&~C1) == 0
02002         if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
02003             ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) ||  // (V|N)
02004              (V2 == B && MaskedValueIsZero(V1, ~C1->getValue()))))   // (N|V)
02005           return BinaryOperator::CreateAnd(A,
02006                                 Builder->getInt(C1->getValue()|C2->getValue()));
02007         // Or commutes, try both ways.
02008         if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
02009             ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) ||  // (V|N)
02010              (V2 == A && MaskedValueIsZero(V1, ~C2->getValue()))))   // (N|V)
02011           return BinaryOperator::CreateAnd(B,
02012                                 Builder->getInt(C1->getValue()|C2->getValue()));
02013 
02014         // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
02015         // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
02016         ConstantInt *C3 = nullptr, *C4 = nullptr;
02017         if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
02018             (C3->getValue() & ~C1->getValue()) == 0 &&
02019             match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
02020             (C4->getValue() & ~C2->getValue()) == 0) {
02021           V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
02022           return BinaryOperator::CreateAnd(V2,
02023                                 Builder->getInt(C1->getValue()|C2->getValue()));
02024         }
02025       }
02026     }
02027 
02028     // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) ->  C0 ? A : B, and commuted variants.
02029     // Don't do this for vector select idioms, the code generator doesn't handle
02030     // them well yet.
02031     if (!I.getType()->isVectorTy()) {
02032       if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
02033         return Match;
02034       if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
02035         return Match;
02036       if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
02037         return Match;
02038       if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
02039         return Match;
02040     }
02041 
02042     // ((A&~B)|(~A&B)) -> A^B
02043     if ((match(C, m_Not(m_Specific(D))) &&
02044          match(B, m_Not(m_Specific(A)))))
02045       return BinaryOperator::CreateXor(A, D);
02046     // ((~B&A)|(~A&B)) -> A^B
02047     if ((match(A, m_Not(m_Specific(D))) &&
02048          match(B, m_Not(m_Specific(C)))))
02049       return BinaryOperator::CreateXor(C, D);
02050     // ((A&~B)|(B&~A)) -> A^B
02051     if ((match(C, m_Not(m_Specific(B))) &&
02052          match(D, m_Not(m_Specific(A)))))
02053       return BinaryOperator::CreateXor(A, B);
02054     // ((~B&A)|(B&~A)) -> A^B
02055     if ((match(A, m_Not(m_Specific(B))) &&
02056          match(D, m_Not(m_Specific(C)))))
02057       return BinaryOperator::CreateXor(C, B);
02058 
02059     // ((A|B)&1)|(B&-2) -> (A&1) | B
02060     if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
02061         match(A, m_Or(m_Specific(B), m_Value(V1)))) {
02062       Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
02063       if (Ret) return Ret;
02064     }
02065     // (B&-2)|((A|B)&1) -> (A&1) | B
02066     if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
02067         match(B, m_Or(m_Value(V1), m_Specific(A)))) {
02068       Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
02069       if (Ret) return Ret;
02070     }
02071   }
02072 
02073   // (X >> Z) | (Y >> Z)  -> (X|Y) >> Z  for all shifts.
02074   if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
02075     if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
02076       if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
02077           SI0->getOperand(1) == SI1->getOperand(1) &&
02078           (SI0->hasOneUse() || SI1->hasOneUse())) {
02079         Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
02080                                          SI0->getName());
02081         return BinaryOperator::Create(SI1->getOpcode(), NewOp,
02082                                       SI1->getOperand(1));
02083       }
02084   }
02085 
02086   // (~A | ~B) == (~(A & B)) - De Morgan's Law
02087   if (Value *Op0NotVal = dyn_castNotVal(Op0))
02088     if (Value *Op1NotVal = dyn_castNotVal(Op1))
02089       if (Op0->hasOneUse() && Op1->hasOneUse()) {
02090         Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
02091                                         I.getName()+".demorgan");
02092         return BinaryOperator::CreateNot(And);
02093       }
02094 
02095   // Canonicalize xor to the RHS.
02096   bool SwappedForXor = false;
02097   if (match(Op0, m_Xor(m_Value(), m_Value()))) {
02098     std::swap(Op0, Op1);
02099     SwappedForXor = true;
02100   }
02101 
02102   // A | ( A ^ B) -> A |  B
02103   // A | (~A ^ B) -> A | ~B
02104   // (A & B) | (A ^ B)
02105   if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
02106     if (Op0 == A || Op0 == B)
02107       return BinaryOperator::CreateOr(A, B);
02108 
02109     if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
02110         match(Op0, m_And(m_Specific(B), m_Specific(A))))
02111       return BinaryOperator::CreateOr(A, B);
02112 
02113     if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
02114       Value *Not = Builder->CreateNot(B, B->getName()+".not");
02115       return BinaryOperator::CreateOr(Not, Op0);
02116     }
02117     if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
02118       Value *Not = Builder->CreateNot(A, A->getName()+".not");
02119       return BinaryOperator::CreateOr(Not, Op0);
02120     }
02121   }
02122 
02123   // A | ~(A | B) -> A | ~B
02124   // A | ~(A ^ B) -> A | ~B
02125   if (match(Op1, m_Not(m_Value(A))))
02126     if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
02127       if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
02128           Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
02129                                B->getOpcode() == Instruction::Xor)) {
02130         Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
02131                                                  B->getOperand(0);
02132         Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
02133         return BinaryOperator::CreateOr(Not, Op0);
02134       }
02135 
02136   if (SwappedForXor)
02137     std::swap(Op0, Op1);
02138 
02139   if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
02140     if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
02141       if (Value *Res = FoldOrOfICmps(LHS, RHS))
02142         return ReplaceInstUsesWith(I, Res);
02143 
02144   // (fcmp uno x, c) | (fcmp uno y, c)  -> (fcmp uno x, y)
02145   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
02146     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
02147       if (Value *Res = FoldOrOfFCmps(LHS, RHS))
02148         return ReplaceInstUsesWith(I, Res);
02149 
02150   // fold (or (cast A), (cast B)) -> (cast (or A, B))
02151   if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
02152     CastInst *Op1C = dyn_cast<CastInst>(Op1);
02153     if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
02154       Type *SrcTy = Op0C->getOperand(0)->getType();
02155       if (SrcTy == Op1C->getOperand(0)->getType() &&
02156           SrcTy->isIntOrIntVectorTy()) {
02157         Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
02158 
02159         if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
02160             // Only do this if the casts both really cause code to be
02161             // generated.
02162             ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
02163             ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
02164           Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
02165           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
02166         }
02167 
02168         // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
02169         // cast is otherwise not optimizable.  This happens for vector sexts.
02170         if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
02171           if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
02172             if (Value *Res = FoldOrOfICmps(LHS, RHS))
02173               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
02174 
02175         // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
02176         // cast is otherwise not optimizable.  This happens for vector sexts.
02177         if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
02178           if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
02179             if (Value *Res = FoldOrOfFCmps(LHS, RHS))
02180               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
02181       }
02182     }
02183   }
02184 
02185   // or(sext(A), B) -> A ? -1 : B where A is an i1
02186   // or(A, sext(B)) -> B ? -1 : A where B is an i1
02187   if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
02188     return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
02189   if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
02190     return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
02191 
02192   // Note: If we've gotten to the point of visiting the outer OR, then the
02193   // inner one couldn't be simplified.  If it was a constant, then it won't
02194   // be simplified by a later pass either, so we try swapping the inner/outer
02195   // ORs in the hopes that we'll be able to simplify it this way.
02196   // (X|C) | V --> (X|V) | C
02197   if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
02198       match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
02199     Value *Inner = Builder->CreateOr(A, Op1);
02200     Inner->takeName(Op0);
02201     return BinaryOperator::CreateOr(Inner, C1);
02202   }
02203 
02204   // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
02205   // Since this OR statement hasn't been optimized further yet, we hope
02206   // that this transformation will allow the new ORs to be optimized.
02207   {
02208     Value *X = nullptr, *Y = nullptr;
02209     if (Op0->hasOneUse() && Op1->hasOneUse() &&
02210         match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
02211         match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
02212       Value *orTrue = Builder->CreateOr(A, C);
02213       Value *orFalse = Builder->CreateOr(B, D);
02214       return SelectInst::Create(X, orTrue, orFalse);
02215     }
02216   }
02217 
02218   return Changed ? &I : nullptr;
02219 }
02220 
02221 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
02222   bool Changed = SimplifyAssociativeOrCommutative(I);
02223   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
02224 
02225   if (Value *V = SimplifyVectorOp(I))
02226     return ReplaceInstUsesWith(I, V);
02227 
02228   if (Value *V = SimplifyXorInst(Op0, Op1, DL))
02229     return ReplaceInstUsesWith(I, V);
02230 
02231   // (A&B)^(A&C) -> A&(B^C) etc
02232   if (Value *V = SimplifyUsingDistributiveLaws(I))
02233     return ReplaceInstUsesWith(I, V);
02234 
02235   // See if we can simplify any instructions used by the instruction whose sole
02236   // purpose is to compute bits we don't care about.
02237   if (SimplifyDemandedInstructionBits(I))
02238     return &I;
02239 
02240   // Is this a ~ operation?
02241   if (Value *NotOp = dyn_castNotVal(&I)) {
02242     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
02243       if (Op0I->getOpcode() == Instruction::And ||
02244           Op0I->getOpcode() == Instruction::Or) {
02245         // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
02246         // ~(~X | Y) === (X & ~Y) - De Morgan's Law
02247         if (dyn_castNotVal(Op0I->getOperand(1)))
02248           Op0I->swapOperands();
02249         if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
02250           Value *NotY =
02251             Builder->CreateNot(Op0I->getOperand(1),
02252                                Op0I->getOperand(1)->getName()+".not");
02253           if (Op0I->getOpcode() == Instruction::And)
02254             return BinaryOperator::CreateOr(Op0NotVal, NotY);
02255           return BinaryOperator::CreateAnd(Op0NotVal, NotY);
02256         }
02257 
02258         // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
02259         // ~(X | Y) === (~X & ~Y) - De Morgan's Law
02260         if (isFreeToInvert(Op0I->getOperand(0)) &&
02261             isFreeToInvert(Op0I->getOperand(1))) {
02262           Value *NotX =
02263             Builder->CreateNot(Op0I->getOperand(0), "notlhs");
02264           Value *NotY =
02265             Builder->CreateNot(Op0I->getOperand(1), "notrhs");
02266           if (Op0I->getOpcode() == Instruction::And)
02267             return BinaryOperator::CreateOr(NotX, NotY);
02268           return BinaryOperator::CreateAnd(NotX, NotY);
02269         }
02270 
02271       } else if (Op0I->getOpcode() == Instruction::AShr) {
02272         // ~(~X >>s Y) --> (X >>s Y)
02273         if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
02274           return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
02275       }
02276     }
02277   }
02278 
02279 
02280   if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
02281     if (RHS->isOne() && Op0->hasOneUse())
02282       // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
02283       if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
02284         return CmpInst::Create(CI->getOpcode(),
02285                                CI->getInversePredicate(),
02286                                CI->getOperand(0), CI->getOperand(1));
02287 
02288     // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
02289     if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
02290       if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
02291         if (CI->hasOneUse() && Op0C->hasOneUse()) {
02292           Instruction::CastOps Opcode = Op0C->getOpcode();
02293           if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
02294               (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
02295                                             Op0C->getDestTy()))) {
02296             CI->setPredicate(CI->getInversePredicate());
02297             return CastInst::Create(Opcode, CI, Op0C->getType());
02298           }
02299         }
02300       }
02301     }
02302 
02303     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
02304       // ~(c-X) == X-c-1 == X+(-c-1)
02305       if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
02306         if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
02307           Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
02308           Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
02309                                       ConstantInt::get(I.getType(), 1));
02310           return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
02311         }
02312 
02313       if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
02314         if (Op0I->getOpcode() == Instruction::Add) {
02315           // ~(X-c) --> (-c-1)-X
02316           if (RHS->isAllOnesValue()) {
02317             Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
02318             return BinaryOperator::CreateSub(
02319                            ConstantExpr::getSub(NegOp0CI,
02320                                       ConstantInt::get(I.getType(), 1)),
02321                                       Op0I->getOperand(0));
02322           } else if (RHS->getValue().isSignBit()) {
02323             // (X + C) ^ signbit -> (X + C + signbit)
02324             Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
02325             return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
02326 
02327           }
02328         } else if (Op0I->getOpcode() == Instruction::Or) {
02329           // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
02330           if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
02331             Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
02332             // Anything in both C1 and C2 is known to be zero, remove it from
02333             // NewRHS.
02334             Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
02335             NewRHS = ConstantExpr::getAnd(NewRHS,
02336                                        ConstantExpr::getNot(CommonBits));
02337             Worklist.Add(Op0I);
02338             I.setOperand(0, Op0I->getOperand(0));
02339             I.setOperand(1, NewRHS);
02340             return &I;
02341           }
02342         } else if (Op0I->getOpcode() == Instruction::LShr) {
02343           // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
02344           // E1 = "X ^ C1"
02345           BinaryOperator *E1;
02346           ConstantInt *C1;
02347           if (Op0I->hasOneUse() &&
02348               (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
02349               E1->getOpcode() == Instruction::Xor &&
02350               (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
02351             // fold (C1 >> C2) ^ C3
02352             ConstantInt *C2 = Op0CI, *C3 = RHS;
02353             APInt FoldConst = C1->getValue().lshr(C2->getValue());
02354             FoldConst ^= C3->getValue();
02355             // Prepare the two operands.
02356             Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
02357             Opnd0->takeName(Op0I);
02358             cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
02359             Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
02360 
02361             return BinaryOperator::CreateXor(Opnd0, FoldVal);
02362           }
02363         }
02364       }
02365     }
02366 
02367     // Try to fold constant and into select arguments.
02368     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
02369       if (Instruction *R = FoldOpIntoSelect(I, SI))
02370         return R;
02371     if (isa<PHINode>(Op0))
02372       if (Instruction *NV = FoldOpIntoPhi(I))
02373         return NV;
02374   }
02375 
02376   BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
02377   if (Op1I) {
02378     Value *A, *B;
02379     if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
02380       if (A == Op0) {              // B^(B|A) == (A|B)^B
02381         Op1I->swapOperands();
02382         I.swapOperands();
02383         std::swap(Op0, Op1);
02384       } else if (B == Op0) {       // B^(A|B) == (A|B)^B
02385         I.swapOperands();     // Simplified below.
02386         std::swap(Op0, Op1);
02387       }
02388     } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
02389                Op1I->hasOneUse()){
02390       if (A == Op0) {                                      // A^(A&B) -> A^(B&A)
02391         Op1I->swapOperands();
02392         std::swap(A, B);
02393       }
02394       if (B == Op0) {                                      // A^(B&A) -> (B&A)^A
02395         I.swapOperands();     // Simplified below.
02396         std::swap(Op0, Op1);
02397       }
02398     }
02399   }
02400 
02401   BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
02402   if (Op0I) {
02403     Value *A, *B;
02404     if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
02405         Op0I->hasOneUse()) {
02406       if (A == Op1)                                  // (B|A)^B == (A|B)^B
02407         std::swap(A, B);
02408       if (B == Op1)                                  // (A|B)^B == A & ~B
02409         return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
02410     } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
02411                Op0I->hasOneUse()){
02412       if (A == Op1)                                        // (A&B)^A -> (B&A)^A
02413         std::swap(A, B);
02414       if (B == Op1 &&                                      // (B&A)^A == ~B & A
02415           !isa<ConstantInt>(Op1)) {  // Canonical form is (B&C)^C
02416         return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
02417       }
02418     }
02419   }
02420 
02421   // (X >> Z) ^ (Y >> Z)  -> (X^Y) >> Z  for all shifts.
02422   if (Op0I && Op1I && Op0I->isShift() &&
02423       Op0I->getOpcode() == Op1I->getOpcode() &&
02424       Op0I->getOperand(1) == Op1I->getOperand(1) &&
02425       (Op0I->hasOneUse() || Op1I->hasOneUse())) {
02426     Value *NewOp =
02427       Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
02428                          Op0I->getName());
02429     return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
02430                                   Op1I->getOperand(1));
02431   }
02432 
02433   if (Op0I && Op1I) {
02434     Value *A, *B, *C, *D;
02435     // (A & B)^(A | B) -> A ^ B
02436     if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
02437         match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
02438       if ((A == C && B == D) || (A == D && B == C))
02439         return BinaryOperator::CreateXor(A, B);
02440     }
02441     // (A | B)^(A & B) -> A ^ B
02442     if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
02443         match(Op1I, m_And(m_Value(C), m_Value(D)))) {
02444       if ((A == C && B == D) || (A == D && B == C))
02445         return BinaryOperator::CreateXor(A, B);
02446     }
02447   }
02448 
02449   // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
02450   if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
02451     if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
02452       if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
02453         if (LHS->getOperand(0) == RHS->getOperand(1) &&
02454             LHS->getOperand(1) == RHS->getOperand(0))
02455           LHS->swapOperands();
02456         if (LHS->getOperand(0) == RHS->getOperand(0) &&
02457             LHS->getOperand(1) == RHS->getOperand(1)) {
02458           Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
02459           unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
02460           bool isSigned = LHS->isSigned() || RHS->isSigned();
02461           return ReplaceInstUsesWith(I,
02462                                getNewICmpValue(isSigned, Code, Op0, Op1,
02463                                                Builder));
02464         }
02465       }
02466 
02467   // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
02468   if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
02469     if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
02470       if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
02471         Type *SrcTy = Op0C->getOperand(0)->getType();
02472         if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
02473             // Only do this if the casts both really cause code to be generated.
02474             ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
02475                                I.getType()) &&
02476             ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
02477                                I.getType())) {
02478           Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
02479                                             Op1C->getOperand(0), I.getName());
02480           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
02481         }
02482       }
02483   }
02484 
02485   return Changed ? &I : nullptr;
02486 }