1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visitAnd, visitOr, and visitXor functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/Analysis/InstructionSimplify.h"
16 #include "llvm/IR/ConstantRange.h"
17 #include "llvm/IR/Intrinsics.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Transforms/Utils/CmpInstAnalysis.h"
20 using namespace llvm;
21 using namespace PatternMatch;
22 
23 #define DEBUG_TYPE "instcombine"
24 
dyn_castNotVal(Value * V)25 static inline Value *dyn_castNotVal(Value *V) {
26   // If this is not(not(x)) don't return that this is a not: we want the two
27   // not's to be folded first.
28   if (BinaryOperator::isNot(V)) {
29     Value *Operand = BinaryOperator::getNotArgument(V);
30     if (!IsFreeToInvert(Operand, Operand->hasOneUse()))
31       return Operand;
32   }
33 
34   // Constants can be considered to be not'ed values...
35   if (ConstantInt *C = dyn_cast<ConstantInt>(V))
36     return ConstantInt::get(C->getType(), ~C->getValue());
37   return nullptr;
38 }
39 
40 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
41 /// predicate into a three bit mask. It also returns whether it is an ordered
42 /// predicate by reference.
getFCmpCode(FCmpInst::Predicate CC,bool & isOrdered)43 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
44   isOrdered = false;
45   switch (CC) {
46   case FCmpInst::FCMP_ORD: isOrdered = true; return 0;  // 000
47   case FCmpInst::FCMP_UNO:                   return 0;  // 000
48   case FCmpInst::FCMP_OGT: isOrdered = true; return 1;  // 001
49   case FCmpInst::FCMP_UGT:                   return 1;  // 001
50   case FCmpInst::FCMP_OEQ: isOrdered = true; return 2;  // 010
51   case FCmpInst::FCMP_UEQ:                   return 2;  // 010
52   case FCmpInst::FCMP_OGE: isOrdered = true; return 3;  // 011
53   case FCmpInst::FCMP_UGE:                   return 3;  // 011
54   case FCmpInst::FCMP_OLT: isOrdered = true; return 4;  // 100
55   case FCmpInst::FCMP_ULT:                   return 4;  // 100
56   case FCmpInst::FCMP_ONE: isOrdered = true; return 5;  // 101
57   case FCmpInst::FCMP_UNE:                   return 5;  // 101
58   case FCmpInst::FCMP_OLE: isOrdered = true; return 6;  // 110
59   case FCmpInst::FCMP_ULE:                   return 6;  // 110
60     // True -> 7
61   default:
62     // Not expecting FCMP_FALSE and FCMP_TRUE;
63     llvm_unreachable("Unexpected FCmp predicate!");
64   }
65 }
66 
67 /// getNewICmpValue - This is the complement of getICmpCode, which turns an
68 /// opcode and two operands into either a constant true or false, or a brand
69 /// new ICmp instruction. The sign is passed in to determine which kind
70 /// of predicate to use in the new icmp instruction.
getNewICmpValue(bool Sign,unsigned Code,Value * LHS,Value * RHS,InstCombiner::BuilderTy * Builder)71 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
72                               InstCombiner::BuilderTy *Builder) {
73   ICmpInst::Predicate NewPred;
74   if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
75     return NewConstant;
76   return Builder->CreateICmp(NewPred, LHS, RHS);
77 }
78 
79 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
80 /// opcode and two operands into either a FCmp instruction. isordered is passed
81 /// in to determine which kind of predicate to use in the new fcmp instruction.
getFCmpValue(bool isordered,unsigned code,Value * LHS,Value * RHS,InstCombiner::BuilderTy * Builder)82 static Value *getFCmpValue(bool isordered, unsigned code,
83                            Value *LHS, Value *RHS,
84                            InstCombiner::BuilderTy *Builder) {
85   CmpInst::Predicate Pred;
86   switch (code) {
87   default: llvm_unreachable("Illegal FCmp code!");
88   case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
89   case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
90   case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
91   case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
92   case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
93   case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
94   case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
95   case 7:
96     if (!isordered) return ConstantInt::getTrue(LHS->getContext());
97     Pred = FCmpInst::FCMP_ORD; break;
98   }
99   return Builder->CreateFCmp(Pred, LHS, RHS);
100 }
101 
102 /// \brief Transform BITWISE_OP(BSWAP(A),BSWAP(B)) to BSWAP(BITWISE_OP(A, B))
103 /// \param I Binary operator to transform.
104 /// \return Pointer to node that must replace the original binary operator, or
105 ///         null pointer if no transformation was made.
SimplifyBSwap(BinaryOperator & I)106 Value *InstCombiner::SimplifyBSwap(BinaryOperator &I) {
107   IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
108 
109   // Can't do vectors.
110   if (I.getType()->isVectorTy()) return nullptr;
111 
112   // Can only do bitwise ops.
113   unsigned Op = I.getOpcode();
114   if (Op != Instruction::And && Op != Instruction::Or &&
115       Op != Instruction::Xor)
116     return nullptr;
117 
118   Value *OldLHS = I.getOperand(0);
119   Value *OldRHS = I.getOperand(1);
120   ConstantInt *ConstLHS = dyn_cast<ConstantInt>(OldLHS);
121   ConstantInt *ConstRHS = dyn_cast<ConstantInt>(OldRHS);
122   IntrinsicInst *IntrLHS = dyn_cast<IntrinsicInst>(OldLHS);
123   IntrinsicInst *IntrRHS = dyn_cast<IntrinsicInst>(OldRHS);
124   bool IsBswapLHS = (IntrLHS && IntrLHS->getIntrinsicID() == Intrinsic::bswap);
125   bool IsBswapRHS = (IntrRHS && IntrRHS->getIntrinsicID() == Intrinsic::bswap);
126 
127   if (!IsBswapLHS && !IsBswapRHS)
128     return nullptr;
129 
130   if (!IsBswapLHS && !ConstLHS)
131     return nullptr;
132 
133   if (!IsBswapRHS && !ConstRHS)
134     return nullptr;
135 
136   /// OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
137   /// OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
138   Value *NewLHS = IsBswapLHS ? IntrLHS->getOperand(0) :
139                   Builder->getInt(ConstLHS->getValue().byteSwap());
140 
141   Value *NewRHS = IsBswapRHS ? IntrRHS->getOperand(0) :
142                   Builder->getInt(ConstRHS->getValue().byteSwap());
143 
144   Value *BinOp = nullptr;
145   if (Op == Instruction::And)
146     BinOp = Builder->CreateAnd(NewLHS, NewRHS);
147   else if (Op == Instruction::Or)
148     BinOp = Builder->CreateOr(NewLHS, NewRHS);
149   else //if (Op == Instruction::Xor)
150     BinOp = Builder->CreateXor(NewLHS, NewRHS);
151 
152   Module *M = I.getParent()->getParent()->getParent();
153   Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
154   return Builder->CreateCall(F, BinOp);
155 }
156 
157 // OptAndOp - This handles expressions of the form ((val OP C1) & C2).  Where
158 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'.  Op is
159 // guaranteed to be a binary operator.
OptAndOp(Instruction * Op,ConstantInt * OpRHS,ConstantInt * AndRHS,BinaryOperator & TheAnd)160 Instruction *InstCombiner::OptAndOp(Instruction *Op,
161                                     ConstantInt *OpRHS,
162                                     ConstantInt *AndRHS,
163                                     BinaryOperator &TheAnd) {
164   Value *X = Op->getOperand(0);
165   Constant *Together = nullptr;
166   if (!Op->isShift())
167     Together = ConstantExpr::getAnd(AndRHS, OpRHS);
168 
169   switch (Op->getOpcode()) {
170   case Instruction::Xor:
171     if (Op->hasOneUse()) {
172       // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
173       Value *And = Builder->CreateAnd(X, AndRHS);
174       And->takeName(Op);
175       return BinaryOperator::CreateXor(And, Together);
176     }
177     break;
178   case Instruction::Or:
179     if (Op->hasOneUse()){
180       if (Together != OpRHS) {
181         // (X | C1) & C2 --> (X | (C1&C2)) & C2
182         Value *Or = Builder->CreateOr(X, Together);
183         Or->takeName(Op);
184         return BinaryOperator::CreateAnd(Or, AndRHS);
185       }
186 
187       ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
188       if (TogetherCI && !TogetherCI->isZero()){
189         // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
190         // NOTE: This reduces the number of bits set in the & mask, which
191         // can expose opportunities for store narrowing.
192         Together = ConstantExpr::getXor(AndRHS, Together);
193         Value *And = Builder->CreateAnd(X, Together);
194         And->takeName(Op);
195         return BinaryOperator::CreateOr(And, OpRHS);
196       }
197     }
198 
199     break;
200   case Instruction::Add:
201     if (Op->hasOneUse()) {
202       // Adding a one to a single bit bit-field should be turned into an XOR
203       // of the bit.  First thing to check is to see if this AND is with a
204       // single bit constant.
205       const APInt &AndRHSV = AndRHS->getValue();
206 
207       // If there is only one bit set.
208       if (AndRHSV.isPowerOf2()) {
209         // Ok, at this point, we know that we are masking the result of the
210         // ADD down to exactly one bit.  If the constant we are adding has
211         // no bits set below this bit, then we can eliminate the ADD.
212         const APInt& AddRHS = OpRHS->getValue();
213 
214         // Check to see if any bits below the one bit set in AndRHSV are set.
215         if ((AddRHS & (AndRHSV-1)) == 0) {
216           // If not, the only thing that can effect the output of the AND is
217           // the bit specified by AndRHSV.  If that bit is set, the effect of
218           // the XOR is to toggle the bit.  If it is clear, then the ADD has
219           // no effect.
220           if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
221             TheAnd.setOperand(0, X);
222             return &TheAnd;
223           } else {
224             // Pull the XOR out of the AND.
225             Value *NewAnd = Builder->CreateAnd(X, AndRHS);
226             NewAnd->takeName(Op);
227             return BinaryOperator::CreateXor(NewAnd, AndRHS);
228           }
229         }
230       }
231     }
232     break;
233 
234   case Instruction::Shl: {
235     // We know that the AND will not produce any of the bits shifted in, so if
236     // the anded constant includes them, clear them now!
237     //
238     uint32_t BitWidth = AndRHS->getType()->getBitWidth();
239     uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
240     APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
241     ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
242 
243     if (CI->getValue() == ShlMask)
244       // Masking out bits that the shift already masks.
245       return ReplaceInstUsesWith(TheAnd, Op);   // No need for the and.
246 
247     if (CI != AndRHS) {                  // Reducing bits set in and.
248       TheAnd.setOperand(1, CI);
249       return &TheAnd;
250     }
251     break;
252   }
253   case Instruction::LShr: {
254     // We know that the AND will not produce any of the bits shifted in, so if
255     // the anded constant includes them, clear them now!  This only applies to
256     // unsigned shifts, because a signed shr may bring in set bits!
257     //
258     uint32_t BitWidth = AndRHS->getType()->getBitWidth();
259     uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
260     APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
261     ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
262 
263     if (CI->getValue() == ShrMask)
264       // Masking out bits that the shift already masks.
265       return ReplaceInstUsesWith(TheAnd, Op);
266 
267     if (CI != AndRHS) {
268       TheAnd.setOperand(1, CI);  // Reduce bits set in and cst.
269       return &TheAnd;
270     }
271     break;
272   }
273   case Instruction::AShr:
274     // Signed shr.
275     // See if this is shifting in some sign extension, then masking it out
276     // with an and.
277     if (Op->hasOneUse()) {
278       uint32_t BitWidth = AndRHS->getType()->getBitWidth();
279       uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
280       APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
281       Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
282       if (C == AndRHS) {          // Masking out bits shifted in.
283         // (Val ashr C1) & C2 -> (Val lshr C1) & C2
284         // Make the argument unsigned.
285         Value *ShVal = Op->getOperand(0);
286         ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
287         return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
288       }
289     }
290     break;
291   }
292   return nullptr;
293 }
294 
295 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
296 /// (V < Lo || V >= Hi).  In practice, we emit the more efficient
297 /// (V-Lo) \<u Hi-Lo.  This method expects that Lo <= Hi. isSigned indicates
298 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
299 /// insert new instructions.
InsertRangeTest(Value * V,Constant * Lo,Constant * Hi,bool isSigned,bool Inside)300 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
301                                      bool isSigned, bool Inside) {
302   assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
303             ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
304          "Lo is not <= Hi in range emission code!");
305 
306   if (Inside) {
307     if (Lo == Hi)  // Trivially false.
308       return Builder->getFalse();
309 
310     // V >= Min && V < Hi --> V < Hi
311     if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
312       ICmpInst::Predicate pred = (isSigned ?
313         ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
314       return Builder->CreateICmp(pred, V, Hi);
315     }
316 
317     // Emit V-Lo <u Hi-Lo
318     Constant *NegLo = ConstantExpr::getNeg(Lo);
319     Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
320     Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
321     return Builder->CreateICmpULT(Add, UpperBound);
322   }
323 
324   if (Lo == Hi)  // Trivially true.
325     return Builder->getTrue();
326 
327   // V < Min || V >= Hi -> V > Hi-1
328   Hi = SubOne(cast<ConstantInt>(Hi));
329   if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
330     ICmpInst::Predicate pred = (isSigned ?
331         ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
332     return Builder->CreateICmp(pred, V, Hi);
333   }
334 
335   // Emit V-Lo >u Hi-1-Lo
336   // Note that Hi has already had one subtracted from it, above.
337   ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
338   Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
339   Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
340   return Builder->CreateICmpUGT(Add, LowerBound);
341 }
342 
343 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
344 // any number of 0s on either side.  The 1s are allowed to wrap from LSB to
345 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.  0x0F0F0000 is
346 // not, since all 1s are not contiguous.
isRunOfOnes(ConstantInt * Val,uint32_t & MB,uint32_t & ME)347 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
348   const APInt& V = Val->getValue();
349   uint32_t BitWidth = Val->getType()->getBitWidth();
350   if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
351 
352   // look for the first zero bit after the run of ones
353   MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
354   // look for the first non-zero bit
355   ME = V.getActiveBits();
356   return true;
357 }
358 
359 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
360 /// where isSub determines whether the operator is a sub.  If we can fold one of
361 /// the following xforms:
362 ///
363 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
364 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
365 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
366 ///
367 /// return (A +/- B).
368 ///
FoldLogicalPlusAnd(Value * LHS,Value * RHS,ConstantInt * Mask,bool isSub,Instruction & I)369 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
370                                         ConstantInt *Mask, bool isSub,
371                                         Instruction &I) {
372   Instruction *LHSI = dyn_cast<Instruction>(LHS);
373   if (!LHSI || LHSI->getNumOperands() != 2 ||
374       !isa<ConstantInt>(LHSI->getOperand(1))) return nullptr;
375 
376   ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
377 
378   switch (LHSI->getOpcode()) {
379   default: return nullptr;
380   case Instruction::And:
381     if (ConstantExpr::getAnd(N, Mask) == Mask) {
382       // If the AndRHS is a power of two minus one (0+1+), this is simple.
383       if ((Mask->getValue().countLeadingZeros() +
384            Mask->getValue().countPopulation()) ==
385           Mask->getValue().getBitWidth())
386         break;
387 
388       // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
389       // part, we don't need any explicit masks to take them out of A.  If that
390       // is all N is, ignore it.
391       uint32_t MB = 0, ME = 0;
392       if (isRunOfOnes(Mask, MB, ME)) {  // begin/end bit of run, inclusive
393         uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
394         APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
395         if (MaskedValueIsZero(RHS, Mask, 0, &I))
396           break;
397       }
398     }
399     return nullptr;
400   case Instruction::Or:
401   case Instruction::Xor:
402     // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
403     if ((Mask->getValue().countLeadingZeros() +
404          Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
405         && ConstantExpr::getAnd(N, Mask)->isNullValue())
406       break;
407     return nullptr;
408   }
409 
410   if (isSub)
411     return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
412   return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
413 }
414 
415 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
416 /// One of A and B is considered the mask, the other the value. This is
417 /// described as the "AMask" or "BMask" part of the enum. If the enum
418 /// contains only "Mask", then both A and B can be considered masks.
419 /// If A is the mask, then it was proven, that (A & C) == C. This
420 /// is trivial if C == A, or C == 0. If both A and C are constants, this
421 /// proof is also easy.
422 /// For the following explanations we assume that A is the mask.
423 /// The part "AllOnes" declares, that the comparison is true only
424 /// if (A & B) == A, or all bits of A are set in B.
425 ///   Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
426 /// The part "AllZeroes" declares, that the comparison is true only
427 /// if (A & B) == 0, or all bits of A are cleared in B.
428 ///   Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
429 /// The part "Mixed" declares, that (A & B) == C and C might or might not
430 /// contain any number of one bits and zero bits.
431 ///   Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
432 /// The Part "Not" means, that in above descriptions "==" should be replaced
433 /// by "!=".
434 ///   Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
435 /// If the mask A contains a single bit, then the following is equivalent:
436 ///    (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
437 ///    (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
438 enum MaskedICmpType {
439   FoldMskICmp_AMask_AllOnes           =     1,
440   FoldMskICmp_AMask_NotAllOnes        =     2,
441   FoldMskICmp_BMask_AllOnes           =     4,
442   FoldMskICmp_BMask_NotAllOnes        =     8,
443   FoldMskICmp_Mask_AllZeroes          =    16,
444   FoldMskICmp_Mask_NotAllZeroes       =    32,
445   FoldMskICmp_AMask_Mixed             =    64,
446   FoldMskICmp_AMask_NotMixed          =   128,
447   FoldMskICmp_BMask_Mixed             =   256,
448   FoldMskICmp_BMask_NotMixed          =   512
449 };
450 
451 /// return the set of pattern classes (from MaskedICmpType)
452 /// that (icmp SCC (A & B), C) satisfies
getTypeOfMaskedICmp(Value * A,Value * B,Value * C,ICmpInst::Predicate SCC)453 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
454                                     ICmpInst::Predicate SCC)
455 {
456   ConstantInt *ACst = dyn_cast<ConstantInt>(A);
457   ConstantInt *BCst = dyn_cast<ConstantInt>(B);
458   ConstantInt *CCst = dyn_cast<ConstantInt>(C);
459   bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
460   bool icmp_abit = (ACst && !ACst->isZero() &&
461                     ACst->getValue().isPowerOf2());
462   bool icmp_bbit = (BCst && !BCst->isZero() &&
463                     BCst->getValue().isPowerOf2());
464   unsigned result = 0;
465   if (CCst && CCst->isZero()) {
466     // if C is zero, then both A and B qualify as mask
467     result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
468                           FoldMskICmp_Mask_AllZeroes |
469                           FoldMskICmp_AMask_Mixed |
470                           FoldMskICmp_BMask_Mixed)
471                        : (FoldMskICmp_Mask_NotAllZeroes |
472                           FoldMskICmp_Mask_NotAllZeroes |
473                           FoldMskICmp_AMask_NotMixed |
474                           FoldMskICmp_BMask_NotMixed));
475     if (icmp_abit)
476       result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
477                             FoldMskICmp_AMask_NotMixed)
478                          : (FoldMskICmp_AMask_AllOnes |
479                             FoldMskICmp_AMask_Mixed));
480     if (icmp_bbit)
481       result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
482                             FoldMskICmp_BMask_NotMixed)
483                          : (FoldMskICmp_BMask_AllOnes |
484                             FoldMskICmp_BMask_Mixed));
485     return result;
486   }
487   if (A == C) {
488     result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
489                           FoldMskICmp_AMask_Mixed)
490                        : (FoldMskICmp_AMask_NotAllOnes |
491                           FoldMskICmp_AMask_NotMixed));
492     if (icmp_abit)
493       result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
494                             FoldMskICmp_AMask_NotMixed)
495                          : (FoldMskICmp_Mask_AllZeroes |
496                             FoldMskICmp_AMask_Mixed));
497   } else if (ACst && CCst &&
498              ConstantExpr::getAnd(ACst, CCst) == CCst) {
499     result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
500                        : FoldMskICmp_AMask_NotMixed);
501   }
502   if (B == C) {
503     result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
504                           FoldMskICmp_BMask_Mixed)
505                        : (FoldMskICmp_BMask_NotAllOnes |
506                           FoldMskICmp_BMask_NotMixed));
507     if (icmp_bbit)
508       result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
509                             FoldMskICmp_BMask_NotMixed)
510                          : (FoldMskICmp_Mask_AllZeroes |
511                             FoldMskICmp_BMask_Mixed));
512   } else if (BCst && CCst &&
513              ConstantExpr::getAnd(BCst, CCst) == CCst) {
514     result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
515                        : FoldMskICmp_BMask_NotMixed);
516   }
517   return result;
518 }
519 
520 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
521 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
522 /// is adjacent to the corresponding normal flag (recording ==), this just
523 /// involves swapping those bits over.
conjugateICmpMask(unsigned Mask)524 static unsigned conjugateICmpMask(unsigned Mask) {
525   unsigned NewMask;
526   NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes |
527                      FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed |
528                      FoldMskICmp_BMask_Mixed))
529             << 1;
530 
531   NewMask |=
532       (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes |
533                FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed |
534                FoldMskICmp_BMask_NotMixed))
535       >> 1;
536 
537   return NewMask;
538 }
539 
540 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
541 /// if possible. The returned predicate is either == or !=. Returns false if
542 /// decomposition fails.
decomposeBitTestICmp(const ICmpInst * I,ICmpInst::Predicate & Pred,Value * & X,Value * & Y,Value * & Z)543 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
544                                  Value *&X, Value *&Y, Value *&Z) {
545   ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1));
546   if (!C)
547     return false;
548 
549   switch (I->getPredicate()) {
550   default:
551     return false;
552   case ICmpInst::ICMP_SLT:
553     // X < 0 is equivalent to (X & SignBit) != 0.
554     if (!C->isZero())
555       return false;
556     Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
557     Pred = ICmpInst::ICMP_NE;
558     break;
559   case ICmpInst::ICMP_SGT:
560     // X > -1 is equivalent to (X & SignBit) == 0.
561     if (!C->isAllOnesValue())
562       return false;
563     Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
564     Pred = ICmpInst::ICMP_EQ;
565     break;
566   case ICmpInst::ICMP_ULT:
567     // X <u 2^n is equivalent to (X & ~(2^n-1)) == 0.
568     if (!C->getValue().isPowerOf2())
569       return false;
570     Y = ConstantInt::get(I->getContext(), -C->getValue());
571     Pred = ICmpInst::ICMP_EQ;
572     break;
573   case ICmpInst::ICMP_UGT:
574     // X >u 2^n-1 is equivalent to (X & ~(2^n-1)) != 0.
575     if (!(C->getValue() + 1).isPowerOf2())
576       return false;
577     Y = ConstantInt::get(I->getContext(), ~C->getValue());
578     Pred = ICmpInst::ICMP_NE;
579     break;
580   }
581 
582   X = I->getOperand(0);
583   Z = ConstantInt::getNullValue(C->getType());
584   return true;
585 }
586 
587 /// foldLogOpOfMaskedICmpsHelper:
588 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
589 /// return the set of pattern classes (from MaskedICmpType)
590 /// that both LHS and RHS satisfy
foldLogOpOfMaskedICmpsHelper(Value * & A,Value * & B,Value * & C,Value * & D,Value * & E,ICmpInst * LHS,ICmpInst * RHS,ICmpInst::Predicate & LHSCC,ICmpInst::Predicate & RHSCC)591 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
592                                              Value*& B, Value*& C,
593                                              Value*& D, Value*& E,
594                                              ICmpInst *LHS, ICmpInst *RHS,
595                                              ICmpInst::Predicate &LHSCC,
596                                              ICmpInst::Predicate &RHSCC) {
597   if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
598   // vectors are not (yet?) supported
599   if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
600 
601   // Here comes the tricky part:
602   // LHS might be of the form L11 & L12 == X, X == L21 & L22,
603   // and L11 & L12 == L21 & L22. The same goes for RHS.
604   // Now we must find those components L** and R**, that are equal, so
605   // that we can extract the parameters A, B, C, D, and E for the canonical
606   // above.
607   Value *L1 = LHS->getOperand(0);
608   Value *L2 = LHS->getOperand(1);
609   Value *L11,*L12,*L21,*L22;
610   // Check whether the icmp can be decomposed into a bit test.
611   if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
612     L21 = L22 = L1 = nullptr;
613   } else {
614     // Look for ANDs in the LHS icmp.
615     if (!L1->getType()->isIntegerTy()) {
616       // You can icmp pointers, for example. They really aren't masks.
617       L11 = L12 = nullptr;
618     } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
619       // Any icmp can be viewed as being trivially masked; if it allows us to
620       // remove one, it's worth it.
621       L11 = L1;
622       L12 = Constant::getAllOnesValue(L1->getType());
623     }
624 
625     if (!L2->getType()->isIntegerTy()) {
626       // You can icmp pointers, for example. They really aren't masks.
627       L21 = L22 = nullptr;
628     } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
629       L21 = L2;
630       L22 = Constant::getAllOnesValue(L2->getType());
631     }
632   }
633 
634   // Bail if LHS was a icmp that can't be decomposed into an equality.
635   if (!ICmpInst::isEquality(LHSCC))
636     return 0;
637 
638   Value *R1 = RHS->getOperand(0);
639   Value *R2 = RHS->getOperand(1);
640   Value *R11,*R12;
641   bool ok = false;
642   if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
643     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
644       A = R11; D = R12;
645     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
646       A = R12; D = R11;
647     } else {
648       return 0;
649     }
650     E = R2; R1 = nullptr; ok = true;
651   } else if (R1->getType()->isIntegerTy()) {
652     if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
653       // As before, model no mask as a trivial mask if it'll let us do an
654       // optimization.
655       R11 = R1;
656       R12 = Constant::getAllOnesValue(R1->getType());
657     }
658 
659     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
660       A = R11; D = R12; E = R2; ok = true;
661     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
662       A = R12; D = R11; E = R2; ok = true;
663     }
664   }
665 
666   // Bail if RHS was a icmp that can't be decomposed into an equality.
667   if (!ICmpInst::isEquality(RHSCC))
668     return 0;
669 
670   // Look for ANDs in on the right side of the RHS icmp.
671   if (!ok && R2->getType()->isIntegerTy()) {
672     if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
673       R11 = R2;
674       R12 = Constant::getAllOnesValue(R2->getType());
675     }
676 
677     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
678       A = R11; D = R12; E = R1; ok = true;
679     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
680       A = R12; D = R11; E = R1; ok = true;
681     } else {
682       return 0;
683     }
684   }
685   if (!ok)
686     return 0;
687 
688   if (L11 == A) {
689     B = L12; C = L2;
690   } else if (L12 == A) {
691     B = L11; C = L2;
692   } else if (L21 == A) {
693     B = L22; C = L1;
694   } else if (L22 == A) {
695     B = L21; C = L1;
696   }
697 
698   unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
699   unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
700   return left_type & right_type;
701 }
702 /// foldLogOpOfMaskedICmps:
703 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
704 /// into a single (icmp(A & X) ==/!= Y)
foldLogOpOfMaskedICmps(ICmpInst * LHS,ICmpInst * RHS,bool IsAnd,llvm::InstCombiner::BuilderTy * Builder)705 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
706                                      llvm::InstCombiner::BuilderTy *Builder) {
707   Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
708   ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
709   unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
710                                                LHSCC, RHSCC);
711   if (mask == 0) return nullptr;
712   assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
713          "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
714 
715   // In full generality:
716   //     (icmp (A & B) Op C) | (icmp (A & D) Op E)
717   // ==  ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
718   //
719   // If the latter can be converted into (icmp (A & X) Op Y) then the former is
720   // equivalent to (icmp (A & X) !Op Y).
721   //
722   // Therefore, we can pretend for the rest of this function that we're dealing
723   // with the conjunction, provided we flip the sense of any comparisons (both
724   // input and output).
725 
726   // In most cases we're going to produce an EQ for the "&&" case.
727   ICmpInst::Predicate NEWCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
728   if (!IsAnd) {
729     // Convert the masking analysis into its equivalent with negated
730     // comparisons.
731     mask = conjugateICmpMask(mask);
732   }
733 
734   if (mask & FoldMskICmp_Mask_AllZeroes) {
735     // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
736     // -> (icmp eq (A & (B|D)), 0)
737     Value *newOr = Builder->CreateOr(B, D);
738     Value *newAnd = Builder->CreateAnd(A, newOr);
739     // we can't use C as zero, because we might actually handle
740     //   (icmp ne (A & B), B) & (icmp ne (A & D), D)
741     // with B and D, having a single bit set
742     Value *zero = Constant::getNullValue(A->getType());
743     return Builder->CreateICmp(NEWCC, newAnd, zero);
744   }
745   if (mask & FoldMskICmp_BMask_AllOnes) {
746     // (icmp eq (A & B), B) & (icmp eq (A & D), D)
747     // -> (icmp eq (A & (B|D)), (B|D))
748     Value *newOr = Builder->CreateOr(B, D);
749     Value *newAnd = Builder->CreateAnd(A, newOr);
750     return Builder->CreateICmp(NEWCC, newAnd, newOr);
751   }
752   if (mask & FoldMskICmp_AMask_AllOnes) {
753     // (icmp eq (A & B), A) & (icmp eq (A & D), A)
754     // -> (icmp eq (A & (B&D)), A)
755     Value *newAnd1 = Builder->CreateAnd(B, D);
756     Value *newAnd = Builder->CreateAnd(A, newAnd1);
757     return Builder->CreateICmp(NEWCC, newAnd, A);
758   }
759 
760   // Remaining cases assume at least that B and D are constant, and depend on
761   // their actual values. This isn't strictly, necessary, just a "handle the
762   // easy cases for now" decision.
763   ConstantInt *BCst = dyn_cast<ConstantInt>(B);
764   if (!BCst) return nullptr;
765   ConstantInt *DCst = dyn_cast<ConstantInt>(D);
766   if (!DCst) return nullptr;
767 
768   if (mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) {
769     // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
770     // (icmp ne (A & B), B) & (icmp ne (A & D), D)
771     //     -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
772     // Only valid if one of the masks is a superset of the other (check "B&D" is
773     // the same as either B or D).
774     APInt NewMask = BCst->getValue() & DCst->getValue();
775 
776     if (NewMask == BCst->getValue())
777       return LHS;
778     else if (NewMask == DCst->getValue())
779       return RHS;
780   }
781   if (mask & FoldMskICmp_AMask_NotAllOnes) {
782     // (icmp ne (A & B), B) & (icmp ne (A & D), D)
783     //     -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
784     // Only valid if one of the masks is a superset of the other (check "B|D" is
785     // the same as either B or D).
786     APInt NewMask = BCst->getValue() | DCst->getValue();
787 
788     if (NewMask == BCst->getValue())
789       return LHS;
790     else if (NewMask == DCst->getValue())
791       return RHS;
792   }
793   if (mask & FoldMskICmp_BMask_Mixed) {
794     // (icmp eq (A & B), C) & (icmp eq (A & D), E)
795     // We already know that B & C == C && D & E == E.
796     // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
797     // C and E, which are shared by both the mask B and the mask D, don't
798     // contradict, then we can transform to
799     // -> (icmp eq (A & (B|D)), (C|E))
800     // Currently, we only handle the case of B, C, D, and E being constant.
801     // we can't simply use C and E, because we might actually handle
802     //   (icmp ne (A & B), B) & (icmp eq (A & D), D)
803     // with B and D, having a single bit set
804     ConstantInt *CCst = dyn_cast<ConstantInt>(C);
805     if (!CCst) return nullptr;
806     ConstantInt *ECst = dyn_cast<ConstantInt>(E);
807     if (!ECst) return nullptr;
808     if (LHSCC != NEWCC)
809       CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
810     if (RHSCC != NEWCC)
811       ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
812     // if there is a conflict we should actually return a false for the
813     // whole construct
814     if (((BCst->getValue() & DCst->getValue()) &
815          (CCst->getValue() ^ ECst->getValue())) != 0)
816       return ConstantInt::get(LHS->getType(), !IsAnd);
817     Value *newOr1 = Builder->CreateOr(B, D);
818     Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
819     Value *newAnd = Builder->CreateAnd(A, newOr1);
820     return Builder->CreateICmp(NEWCC, newAnd, newOr2);
821   }
822   return nullptr;
823 }
824 
825 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
826 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
827 /// If \p Inverted is true then the check is for the inverted range, e.g.
828 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
simplifyRangeCheck(ICmpInst * Cmp0,ICmpInst * Cmp1,bool Inverted)829 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
830                                         bool Inverted) {
831   // Check the lower range comparison, e.g. x >= 0
832   // InstCombine already ensured that if there is a constant it's on the RHS.
833   ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
834   if (!RangeStart)
835     return nullptr;
836 
837   ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
838                                Cmp0->getPredicate());
839 
840   // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
841   if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
842         (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
843     return nullptr;
844 
845   ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
846                                Cmp1->getPredicate());
847 
848   Value *Input = Cmp0->getOperand(0);
849   Value *RangeEnd;
850   if (Cmp1->getOperand(0) == Input) {
851     // For the upper range compare we have: icmp x, n
852     RangeEnd = Cmp1->getOperand(1);
853   } else if (Cmp1->getOperand(1) == Input) {
854     // For the upper range compare we have: icmp n, x
855     RangeEnd = Cmp1->getOperand(0);
856     Pred1 = ICmpInst::getSwappedPredicate(Pred1);
857   } else {
858     return nullptr;
859   }
860 
861   // Check the upper range comparison, e.g. x < n
862   ICmpInst::Predicate NewPred;
863   switch (Pred1) {
864     case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
865     case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
866     default: return nullptr;
867   }
868 
869   // This simplification is only valid if the upper range is not negative.
870   bool IsNegative, IsNotNegative;
871   ComputeSignBit(RangeEnd, IsNotNegative, IsNegative, /*Depth=*/0, Cmp1);
872   if (!IsNotNegative)
873     return nullptr;
874 
875   if (Inverted)
876     NewPred = ICmpInst::getInversePredicate(NewPred);
877 
878   return Builder->CreateICmp(NewPred, Input, RangeEnd);
879 }
880 
881 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
FoldAndOfICmps(ICmpInst * LHS,ICmpInst * RHS)882 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
883   ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
884 
885   // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
886   if (PredicatesFoldable(LHSCC, RHSCC)) {
887     if (LHS->getOperand(0) == RHS->getOperand(1) &&
888         LHS->getOperand(1) == RHS->getOperand(0))
889       LHS->swapOperands();
890     if (LHS->getOperand(0) == RHS->getOperand(0) &&
891         LHS->getOperand(1) == RHS->getOperand(1)) {
892       Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
893       unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
894       bool isSigned = LHS->isSigned() || RHS->isSigned();
895       return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
896     }
897   }
898 
899   // handle (roughly):  (icmp eq (A & B), C) & (icmp eq (A & D), E)
900   if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
901     return V;
902 
903   // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
904   if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
905     return V;
906 
907   // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
908   if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
909     return V;
910 
911   // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
912   Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
913   ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
914   ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
915   if (!LHSCst || !RHSCst) return nullptr;
916 
917   if (LHSCst == RHSCst && LHSCC == RHSCC) {
918     // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
919     // where C is a power of 2
920     if (LHSCC == ICmpInst::ICMP_ULT &&
921         LHSCst->getValue().isPowerOf2()) {
922       Value *NewOr = Builder->CreateOr(Val, Val2);
923       return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
924     }
925 
926     // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
927     if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
928       Value *NewOr = Builder->CreateOr(Val, Val2);
929       return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
930     }
931   }
932 
933   // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
934   // where CMAX is the all ones value for the truncated type,
935   // iff the lower bits of C2 and CA are zero.
936   if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
937       LHS->hasOneUse() && RHS->hasOneUse()) {
938     Value *V;
939     ConstantInt *AndCst, *SmallCst = nullptr, *BigCst = nullptr;
940 
941     // (trunc x) == C1 & (and x, CA) == C2
942     // (and x, CA) == C2 & (trunc x) == C1
943     if (match(Val2, m_Trunc(m_Value(V))) &&
944         match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
945       SmallCst = RHSCst;
946       BigCst = LHSCst;
947     } else if (match(Val, m_Trunc(m_Value(V))) &&
948                match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
949       SmallCst = LHSCst;
950       BigCst = RHSCst;
951     }
952 
953     if (SmallCst && BigCst) {
954       unsigned BigBitSize = BigCst->getType()->getBitWidth();
955       unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
956 
957       // Check that the low bits are zero.
958       APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
959       if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
960         Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
961         APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
962         Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
963         return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
964       }
965     }
966   }
967 
968   // From here on, we only handle:
969   //    (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
970   if (Val != Val2) return nullptr;
971 
972   // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
973   if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
974       RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
975       LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
976       RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
977     return nullptr;
978 
979   // Make a constant range that's the intersection of the two icmp ranges.
980   // If the intersection is empty, we know that the result is false.
981   ConstantRange LHSRange =
982       ConstantRange::makeAllowedICmpRegion(LHSCC, LHSCst->getValue());
983   ConstantRange RHSRange =
984       ConstantRange::makeAllowedICmpRegion(RHSCC, RHSCst->getValue());
985 
986   if (LHSRange.intersectWith(RHSRange).isEmptySet())
987     return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
988 
989   // We can't fold (ugt x, C) & (sgt x, C2).
990   if (!PredicatesFoldable(LHSCC, RHSCC))
991     return nullptr;
992 
993   // Ensure that the larger constant is on the RHS.
994   bool ShouldSwap;
995   if (CmpInst::isSigned(LHSCC) ||
996       (ICmpInst::isEquality(LHSCC) &&
997        CmpInst::isSigned(RHSCC)))
998     ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
999   else
1000     ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1001 
1002   if (ShouldSwap) {
1003     std::swap(LHS, RHS);
1004     std::swap(LHSCst, RHSCst);
1005     std::swap(LHSCC, RHSCC);
1006   }
1007 
1008   // At this point, we know we have two icmp instructions
1009   // comparing a value against two constants and and'ing the result
1010   // together.  Because of the above check, we know that we only have
1011   // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
1012   // (from the icmp folding check above), that the two constants
1013   // are not equal and that the larger constant is on the RHS
1014   assert(LHSCst != RHSCst && "Compares not folded above?");
1015 
1016   switch (LHSCC) {
1017   default: llvm_unreachable("Unknown integer condition code!");
1018   case ICmpInst::ICMP_EQ:
1019     switch (RHSCC) {
1020     default: llvm_unreachable("Unknown integer condition code!");
1021     case ICmpInst::ICMP_NE:         // (X == 13 & X != 15) -> X == 13
1022     case ICmpInst::ICMP_ULT:        // (X == 13 & X <  15) -> X == 13
1023     case ICmpInst::ICMP_SLT:        // (X == 13 & X <  15) -> X == 13
1024       return LHS;
1025     }
1026   case ICmpInst::ICMP_NE:
1027     switch (RHSCC) {
1028     default: llvm_unreachable("Unknown integer condition code!");
1029     case ICmpInst::ICMP_ULT:
1030       if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
1031         return Builder->CreateICmpULT(Val, LHSCst);
1032       if (LHSCst->isNullValue())    // (X !=  0 & X u< 14) -> X-1 u< 13
1033         return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
1034       break;                        // (X != 13 & X u< 15) -> no change
1035     case ICmpInst::ICMP_SLT:
1036       if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
1037         return Builder->CreateICmpSLT(Val, LHSCst);
1038       break;                        // (X != 13 & X s< 15) -> no change
1039     case ICmpInst::ICMP_EQ:         // (X != 13 & X == 15) -> X == 15
1040     case ICmpInst::ICMP_UGT:        // (X != 13 & X u> 15) -> X u> 15
1041     case ICmpInst::ICMP_SGT:        // (X != 13 & X s> 15) -> X s> 15
1042       return RHS;
1043     case ICmpInst::ICMP_NE:
1044       // Special case to get the ordering right when the values wrap around
1045       // zero.
1046       if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue())
1047         std::swap(LHSCst, RHSCst);
1048       if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
1049         Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1050         Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1051         return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1),
1052                                       Val->getName()+".cmp");
1053       }
1054       break;                        // (X != 13 & X != 15) -> no change
1055     }
1056     break;
1057   case ICmpInst::ICMP_ULT:
1058     switch (RHSCC) {
1059     default: llvm_unreachable("Unknown integer condition code!");
1060     case ICmpInst::ICMP_EQ:         // (X u< 13 & X == 15) -> false
1061     case ICmpInst::ICMP_UGT:        // (X u< 13 & X u> 15) -> false
1062       return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1063     case ICmpInst::ICMP_SGT:        // (X u< 13 & X s> 15) -> no change
1064       break;
1065     case ICmpInst::ICMP_NE:         // (X u< 13 & X != 15) -> X u< 13
1066     case ICmpInst::ICMP_ULT:        // (X u< 13 & X u< 15) -> X u< 13
1067       return LHS;
1068     case ICmpInst::ICMP_SLT:        // (X u< 13 & X s< 15) -> no change
1069       break;
1070     }
1071     break;
1072   case ICmpInst::ICMP_SLT:
1073     switch (RHSCC) {
1074     default: llvm_unreachable("Unknown integer condition code!");
1075     case ICmpInst::ICMP_UGT:        // (X s< 13 & X u> 15) -> no change
1076       break;
1077     case ICmpInst::ICMP_NE:         // (X s< 13 & X != 15) -> X < 13
1078     case ICmpInst::ICMP_SLT:        // (X s< 13 & X s< 15) -> X < 13
1079       return LHS;
1080     case ICmpInst::ICMP_ULT:        // (X s< 13 & X u< 15) -> no change
1081       break;
1082     }
1083     break;
1084   case ICmpInst::ICMP_UGT:
1085     switch (RHSCC) {
1086     default: llvm_unreachable("Unknown integer condition code!");
1087     case ICmpInst::ICMP_EQ:         // (X u> 13 & X == 15) -> X == 15
1088     case ICmpInst::ICMP_UGT:        // (X u> 13 & X u> 15) -> X u> 15
1089       return RHS;
1090     case ICmpInst::ICMP_SGT:        // (X u> 13 & X s> 15) -> no change
1091       break;
1092     case ICmpInst::ICMP_NE:
1093       if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
1094         return Builder->CreateICmp(LHSCC, Val, RHSCst);
1095       break;                        // (X u> 13 & X != 15) -> no change
1096     case ICmpInst::ICMP_ULT:        // (X u> 13 & X u< 15) -> (X-14) <u 1
1097       return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
1098     case ICmpInst::ICMP_SLT:        // (X u> 13 & X s< 15) -> no change
1099       break;
1100     }
1101     break;
1102   case ICmpInst::ICMP_SGT:
1103     switch (RHSCC) {
1104     default: llvm_unreachable("Unknown integer condition code!");
1105     case ICmpInst::ICMP_EQ:         // (X s> 13 & X == 15) -> X == 15
1106     case ICmpInst::ICMP_SGT:        // (X s> 13 & X s> 15) -> X s> 15
1107       return RHS;
1108     case ICmpInst::ICMP_UGT:        // (X s> 13 & X u> 15) -> no change
1109       break;
1110     case ICmpInst::ICMP_NE:
1111       if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
1112         return Builder->CreateICmp(LHSCC, Val, RHSCst);
1113       break;                        // (X s> 13 & X != 15) -> no change
1114     case ICmpInst::ICMP_SLT:        // (X s> 13 & X s< 15) -> (X-14) s< 1
1115       return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
1116     case ICmpInst::ICMP_ULT:        // (X s> 13 & X u< 15) -> no change
1117       break;
1118     }
1119     break;
1120   }
1121 
1122   return nullptr;
1123 }
1124 
1125 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp).  NOTE: Unlike the rest of
1126 /// instcombine, this returns a Value which should already be inserted into the
1127 /// function.
FoldAndOfFCmps(FCmpInst * LHS,FCmpInst * RHS)1128 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1129   if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
1130       RHS->getPredicate() == FCmpInst::FCMP_ORD) {
1131     if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
1132       return nullptr;
1133 
1134     // (fcmp ord x, c) & (fcmp ord y, c)  -> (fcmp ord x, y)
1135     if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1136       if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1137         // If either of the constants are nans, then the whole thing returns
1138         // false.
1139         if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1140           return Builder->getFalse();
1141         return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1142       }
1143 
1144     // Handle vector zeros.  This occurs because the canonical form of
1145     // "fcmp ord x,x" is "fcmp ord x, 0".
1146     if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1147         isa<ConstantAggregateZero>(RHS->getOperand(1)))
1148       return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1149     return nullptr;
1150   }
1151 
1152   Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1153   Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1154   FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1155 
1156 
1157   if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1158     // Swap RHS operands to match LHS.
1159     Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1160     std::swap(Op1LHS, Op1RHS);
1161   }
1162 
1163   if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1164     // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1165     if (Op0CC == Op1CC)
1166       return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1167     if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
1168       return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1169     if (Op0CC == FCmpInst::FCMP_TRUE)
1170       return RHS;
1171     if (Op1CC == FCmpInst::FCMP_TRUE)
1172       return LHS;
1173 
1174     bool Op0Ordered;
1175     bool Op1Ordered;
1176     unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1177     unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1178     // uno && ord -> false
1179     if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered)
1180         return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1181     if (Op1Pred == 0) {
1182       std::swap(LHS, RHS);
1183       std::swap(Op0Pred, Op1Pred);
1184       std::swap(Op0Ordered, Op1Ordered);
1185     }
1186     if (Op0Pred == 0) {
1187       // uno && ueq -> uno && (uno || eq) -> uno
1188       // ord && olt -> ord && (ord && lt) -> olt
1189       if (!Op0Ordered && (Op0Ordered == Op1Ordered))
1190         return LHS;
1191       if (Op0Ordered && (Op0Ordered == Op1Ordered))
1192         return RHS;
1193 
1194       // uno && oeq -> uno && (ord && eq) -> false
1195       if (!Op0Ordered)
1196         return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1197       // ord && ueq -> ord && (uno || eq) -> oeq
1198       return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
1199     }
1200   }
1201 
1202   return nullptr;
1203 }
1204 
visitAnd(BinaryOperator & I)1205 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1206   bool Changed = SimplifyAssociativeOrCommutative(I);
1207   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1208 
1209   if (Value *V = SimplifyVectorOp(I))
1210     return ReplaceInstUsesWith(I, V);
1211 
1212   if (Value *V = SimplifyAndInst(Op0, Op1, DL, TLI, DT, AC))
1213     return ReplaceInstUsesWith(I, V);
1214 
1215   // (A|B)&(A|C) -> A|(B&C) etc
1216   if (Value *V = SimplifyUsingDistributiveLaws(I))
1217     return ReplaceInstUsesWith(I, V);
1218 
1219   // See if we can simplify any instructions used by the instruction whose sole
1220   // purpose is to compute bits we don't care about.
1221   if (SimplifyDemandedInstructionBits(I))
1222     return &I;
1223 
1224   if (Value *V = SimplifyBSwap(I))
1225     return ReplaceInstUsesWith(I, V);
1226 
1227   if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1228     const APInt &AndRHSMask = AndRHS->getValue();
1229 
1230     // Optimize a variety of ((val OP C1) & C2) combinations...
1231     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1232       Value *Op0LHS = Op0I->getOperand(0);
1233       Value *Op0RHS = Op0I->getOperand(1);
1234       switch (Op0I->getOpcode()) {
1235       default: break;
1236       case Instruction::Xor:
1237       case Instruction::Or: {
1238         // If the mask is only needed on one incoming arm, push it up.
1239         if (!Op0I->hasOneUse()) break;
1240 
1241         APInt NotAndRHS(~AndRHSMask);
1242         if (MaskedValueIsZero(Op0LHS, NotAndRHS, 0, &I)) {
1243           // Not masking anything out for the LHS, move to RHS.
1244           Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1245                                              Op0RHS->getName()+".masked");
1246           return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1247         }
1248         if (!isa<Constant>(Op0RHS) &&
1249             MaskedValueIsZero(Op0RHS, NotAndRHS, 0, &I)) {
1250           // Not masking anything out for the RHS, move to LHS.
1251           Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1252                                              Op0LHS->getName()+".masked");
1253           return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1254         }
1255 
1256         break;
1257       }
1258       case Instruction::Add:
1259         // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1260         // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1261         // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1262         if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1263           return BinaryOperator::CreateAnd(V, AndRHS);
1264         if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1265           return BinaryOperator::CreateAnd(V, AndRHS);  // Add commutes
1266         break;
1267 
1268       case Instruction::Sub:
1269         // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1270         // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1271         // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1272         if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1273           return BinaryOperator::CreateAnd(V, AndRHS);
1274 
1275         // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1276         // has 1's for all bits that the subtraction with A might affect.
1277         if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1278           uint32_t BitWidth = AndRHSMask.getBitWidth();
1279           uint32_t Zeros = AndRHSMask.countLeadingZeros();
1280           APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1281 
1282           if (MaskedValueIsZero(Op0LHS, Mask, 0, &I)) {
1283             Value *NewNeg = Builder->CreateNeg(Op0RHS);
1284             return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1285           }
1286         }
1287         break;
1288 
1289       case Instruction::Shl:
1290       case Instruction::LShr:
1291         // (1 << x) & 1 --> zext(x == 0)
1292         // (1 >> x) & 1 --> zext(x == 0)
1293         if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1294           Value *NewICmp =
1295             Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1296           return new ZExtInst(NewICmp, I.getType());
1297         }
1298         break;
1299       }
1300 
1301       if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1302         if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1303           return Res;
1304     }
1305 
1306     // If this is an integer truncation, and if the source is an 'and' with
1307     // immediate, transform it.  This frequently occurs for bitfield accesses.
1308     {
1309       Value *X = nullptr; ConstantInt *YC = nullptr;
1310       if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1311         // Change: and (trunc (and X, YC) to T), C2
1312         // into  : and (trunc X to T), trunc(YC) & C2
1313         // This will fold the two constants together, which may allow
1314         // other simplifications.
1315         Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1316         Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1317         C3 = ConstantExpr::getAnd(C3, AndRHS);
1318         return BinaryOperator::CreateAnd(NewCast, C3);
1319       }
1320     }
1321 
1322     // Try to fold constant and into select arguments.
1323     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1324       if (Instruction *R = FoldOpIntoSelect(I, SI))
1325         return R;
1326     if (isa<PHINode>(Op0))
1327       if (Instruction *NV = FoldOpIntoPhi(I))
1328         return NV;
1329   }
1330 
1331 
1332   // (~A & ~B) == (~(A | B)) - De Morgan's Law
1333   if (Value *Op0NotVal = dyn_castNotVal(Op0))
1334     if (Value *Op1NotVal = dyn_castNotVal(Op1))
1335       if (Op0->hasOneUse() && Op1->hasOneUse()) {
1336         Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1337                                       I.getName()+".demorgan");
1338         return BinaryOperator::CreateNot(Or);
1339       }
1340 
1341   {
1342     Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
1343     // (A|B) & ~(A&B) -> A^B
1344     if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1345         match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1346         ((A == C && B == D) || (A == D && B == C)))
1347       return BinaryOperator::CreateXor(A, B);
1348 
1349     // ~(A&B) & (A|B) -> A^B
1350     if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1351         match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1352         ((A == C && B == D) || (A == D && B == C)))
1353       return BinaryOperator::CreateXor(A, B);
1354 
1355     // A&(A^B) => A & ~B
1356     {
1357       Value *tmpOp0 = Op0;
1358       Value *tmpOp1 = Op1;
1359       if (Op0->hasOneUse() &&
1360           match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1361         if (A == Op1 || B == Op1 ) {
1362           tmpOp1 = Op0;
1363           tmpOp0 = Op1;
1364           // Simplify below
1365         }
1366       }
1367 
1368       if (tmpOp1->hasOneUse() &&
1369           match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
1370         if (B == tmpOp0) {
1371           std::swap(A, B);
1372         }
1373         // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1374         // A is originally -1 (or a vector of -1 and undefs), then we enter
1375         // an endless loop. By checking that A is non-constant we ensure that
1376         // we will never get to the loop.
1377         if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1378           return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1379       }
1380     }
1381 
1382     // (A&((~A)|B)) -> A&B
1383     if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1384         match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1385       return BinaryOperator::CreateAnd(A, Op1);
1386     if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1387         match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1388       return BinaryOperator::CreateAnd(A, Op0);
1389 
1390     // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1391     if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1392       if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1393         if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
1394           return BinaryOperator::CreateAnd(Op0, Builder->CreateNot(C));
1395 
1396     // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1397     if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1398       if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1399         if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
1400           return BinaryOperator::CreateAnd(Op1, Builder->CreateNot(C));
1401 
1402     // (A | B) & ((~A) ^ B) -> (A & B)
1403     if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1404         match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
1405       return BinaryOperator::CreateAnd(A, B);
1406 
1407     // ((~A) ^ B) & (A | B) -> (A & B)
1408     if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1409         match(Op1, m_Or(m_Specific(A), m_Specific(B))))
1410       return BinaryOperator::CreateAnd(A, B);
1411   }
1412 
1413   {
1414     ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1415     ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1416     if (LHS && RHS)
1417       if (Value *Res = FoldAndOfICmps(LHS, RHS))
1418         return ReplaceInstUsesWith(I, Res);
1419 
1420     // TODO: Make this recursive; it's a little tricky because an arbitrary
1421     // number of 'and' instructions might have to be created.
1422     Value *X, *Y;
1423     if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1424       if (auto *Cmp = dyn_cast<ICmpInst>(X))
1425         if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1426           return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1427       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1428         if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1429           return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1430     }
1431     if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1432       if (auto *Cmp = dyn_cast<ICmpInst>(X))
1433         if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1434           return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1435       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1436         if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1437           return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1438     }
1439   }
1440 
1441   // If and'ing two fcmp, try combine them into one.
1442   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1443     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1444       if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1445         return ReplaceInstUsesWith(I, Res);
1446 
1447 
1448   // fold (and (cast A), (cast B)) -> (cast (and A, B))
1449   if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1450     if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1451       Type *SrcTy = Op0C->getOperand(0)->getType();
1452       if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1453           SrcTy == Op1C->getOperand(0)->getType() &&
1454           SrcTy->isIntOrIntVectorTy()) {
1455         Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1456 
1457         // Only do this if the casts both really cause code to be generated.
1458         if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1459             ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1460           Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1461           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1462         }
1463 
1464         // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1465         // cast is otherwise not optimizable.  This happens for vector sexts.
1466         if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1467           if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1468             if (Value *Res = FoldAndOfICmps(LHS, RHS))
1469               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1470 
1471         // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1472         // cast is otherwise not optimizable.  This happens for vector sexts.
1473         if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1474           if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1475             if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1476               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1477       }
1478     }
1479 
1480   {
1481     Value *X = nullptr;
1482     bool OpsSwapped = false;
1483     // Canonicalize SExt or Not to the LHS
1484     if (match(Op1, m_SExt(m_Value())) ||
1485         match(Op1, m_Not(m_Value()))) {
1486       std::swap(Op0, Op1);
1487       OpsSwapped = true;
1488     }
1489 
1490     // Fold (and (sext bool to A), B) --> (select bool, B, 0)
1491     if (match(Op0, m_SExt(m_Value(X))) &&
1492         X->getType()->getScalarType()->isIntegerTy(1)) {
1493       Value *Zero = Constant::getNullValue(Op1->getType());
1494       return SelectInst::Create(X, Op1, Zero);
1495     }
1496 
1497     // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
1498     if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
1499         X->getType()->getScalarType()->isIntegerTy(1)) {
1500       Value *Zero = Constant::getNullValue(Op0->getType());
1501       return SelectInst::Create(X, Zero, Op1);
1502     }
1503 
1504     if (OpsSwapped)
1505       std::swap(Op0, Op1);
1506   }
1507 
1508   return Changed ? &I : nullptr;
1509 }
1510 
1511 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1512 /// capable of providing pieces of a bswap.  The subexpression provides pieces
1513 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1514 /// the expression came from the corresponding "byte swapped" byte in some other
1515 /// value.  For example, if the current subexpression is "(shl i32 %X, 24)" then
1516 /// we know that the expression deposits the low byte of %X into the high byte
1517 /// of the bswap result and that all other bytes are zero.  This expression is
1518 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1519 /// match.
1520 ///
1521 /// This function returns true if the match was unsuccessful and false if so.
1522 /// On entry to the function the "OverallLeftShift" is a signed integer value
1523 /// indicating the number of bytes that the subexpression is later shifted.  For
1524 /// example, if the expression is later right shifted by 16 bits, the
1525 /// OverallLeftShift value would be -2 on entry.  This is used to specify which
1526 /// byte of ByteValues is actually being set.
1527 ///
1528 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1529 /// byte is masked to zero by a user.  For example, in (X & 255), X will be
1530 /// processed with a bytemask of 1.  Because bytemask is 32-bits, this limits
1531 /// this function to working on up to 32-byte (256 bit) values.  ByteMask is
1532 /// always in the local (OverallLeftShift) coordinate space.
1533 ///
CollectBSwapParts(Value * V,int OverallLeftShift,uint32_t ByteMask,SmallVectorImpl<Value * > & ByteValues)1534 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1535                               SmallVectorImpl<Value *> &ByteValues) {
1536   if (Instruction *I = dyn_cast<Instruction>(V)) {
1537     // If this is an or instruction, it may be an inner node of the bswap.
1538     if (I->getOpcode() == Instruction::Or) {
1539       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1540                                ByteValues) ||
1541              CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1542                                ByteValues);
1543     }
1544 
1545     // If this is a logical shift by a constant multiple of 8, recurse with
1546     // OverallLeftShift and ByteMask adjusted.
1547     if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1548       unsigned ShAmt =
1549         cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1550       // Ensure the shift amount is defined and of a byte value.
1551       if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1552         return true;
1553 
1554       unsigned ByteShift = ShAmt >> 3;
1555       if (I->getOpcode() == Instruction::Shl) {
1556         // X << 2 -> collect(X, +2)
1557         OverallLeftShift += ByteShift;
1558         ByteMask >>= ByteShift;
1559       } else {
1560         // X >>u 2 -> collect(X, -2)
1561         OverallLeftShift -= ByteShift;
1562         ByteMask <<= ByteShift;
1563         ByteMask &= (~0U >> (32-ByteValues.size()));
1564       }
1565 
1566       if (OverallLeftShift >= (int)ByteValues.size()) return true;
1567       if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1568 
1569       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1570                                ByteValues);
1571     }
1572 
1573     // If this is a logical 'and' with a mask that clears bytes, clear the
1574     // corresponding bytes in ByteMask.
1575     if (I->getOpcode() == Instruction::And &&
1576         isa<ConstantInt>(I->getOperand(1))) {
1577       // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1578       unsigned NumBytes = ByteValues.size();
1579       APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1580       const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1581 
1582       for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1583         // If this byte is masked out by a later operation, we don't care what
1584         // the and mask is.
1585         if ((ByteMask & (1 << i)) == 0)
1586           continue;
1587 
1588         // If the AndMask is all zeros for this byte, clear the bit.
1589         APInt MaskB = AndMask & Byte;
1590         if (MaskB == 0) {
1591           ByteMask &= ~(1U << i);
1592           continue;
1593         }
1594 
1595         // If the AndMask is not all ones for this byte, it's not a bytezap.
1596         if (MaskB != Byte)
1597           return true;
1598 
1599         // Otherwise, this byte is kept.
1600       }
1601 
1602       return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1603                                ByteValues);
1604     }
1605   }
1606 
1607   // Okay, we got to something that isn't a shift, 'or' or 'and'.  This must be
1608   // the input value to the bswap.  Some observations: 1) if more than one byte
1609   // is demanded from this input, then it could not be successfully assembled
1610   // into a byteswap.  At least one of the two bytes would not be aligned with
1611   // their ultimate destination.
1612   if (!isPowerOf2_32(ByteMask)) return true;
1613   unsigned InputByteNo = countTrailingZeros(ByteMask);
1614 
1615   // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1616   // is demanded, it needs to go into byte 0 of the result.  This means that the
1617   // byte needs to be shifted until it lands in the right byte bucket.  The
1618   // shift amount depends on the position: if the byte is coming from the high
1619   // part of the value (e.g. byte 3) then it must be shifted right.  If from the
1620   // low part, it must be shifted left.
1621   unsigned DestByteNo = InputByteNo + OverallLeftShift;
1622   if (ByteValues.size()-1-DestByteNo != InputByteNo)
1623     return true;
1624 
1625   // If the destination byte value is already defined, the values are or'd
1626   // together, which isn't a bswap (unless it's an or of the same bits).
1627   if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1628     return true;
1629   ByteValues[DestByteNo] = V;
1630   return false;
1631 }
1632 
1633 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1634 /// If so, insert the new bswap intrinsic and return it.
MatchBSwap(BinaryOperator & I)1635 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1636   IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1637   if (!ITy || ITy->getBitWidth() % 16 ||
1638       // ByteMask only allows up to 32-byte values.
1639       ITy->getBitWidth() > 32*8)
1640     return nullptr;   // Can only bswap pairs of bytes.  Can't do vectors.
1641 
1642   /// ByteValues - For each byte of the result, we keep track of which value
1643   /// defines each byte.
1644   SmallVector<Value*, 8> ByteValues;
1645   ByteValues.resize(ITy->getBitWidth()/8);
1646 
1647   // Try to find all the pieces corresponding to the bswap.
1648   uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1649   if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1650     return nullptr;
1651 
1652   // Check to see if all of the bytes come from the same value.
1653   Value *V = ByteValues[0];
1654   if (!V) return nullptr;  // Didn't find a byte?  Must be zero.
1655 
1656   // Check to make sure that all of the bytes come from the same value.
1657   for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1658     if (ByteValues[i] != V)
1659       return nullptr;
1660   Module *M = I.getParent()->getParent()->getParent();
1661   Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
1662   return CallInst::Create(F, V);
1663 }
1664 
1665 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D).  Check
1666 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1667 /// we can simplify this expression to "cond ? C : D or B".
MatchSelectFromAndOr(Value * A,Value * B,Value * C,Value * D)1668 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1669                                          Value *C, Value *D) {
1670   // If A is not a select of -1/0, this cannot match.
1671   Value *Cond = nullptr;
1672   if (!match(A, m_SExt(m_Value(Cond))) ||
1673       !Cond->getType()->isIntegerTy(1))
1674     return nullptr;
1675 
1676   // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1677   if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1678     return SelectInst::Create(Cond, C, B);
1679   if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1680     return SelectInst::Create(Cond, C, B);
1681 
1682   // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1683   if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1684     return SelectInst::Create(Cond, C, D);
1685   if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1686     return SelectInst::Create(Cond, C, D);
1687   return nullptr;
1688 }
1689 
1690 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
FoldOrOfICmps(ICmpInst * LHS,ICmpInst * RHS,Instruction * CxtI)1691 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1692                                    Instruction *CxtI) {
1693   ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1694 
1695   // Fold (iszero(A & K1) | iszero(A & K2)) ->  (A & (K1 | K2)) != (K1 | K2)
1696   // if K1 and K2 are a one-bit mask.
1697   ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1698   ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1699 
1700   if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() &&
1701       RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1702 
1703     BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0));
1704     BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0));
1705     if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() &&
1706         LAnd->getOpcode() == Instruction::And &&
1707         RAnd->getOpcode() == Instruction::And) {
1708 
1709       Value *Mask = nullptr;
1710       Value *Masked = nullptr;
1711       if (LAnd->getOperand(0) == RAnd->getOperand(0) &&
1712           isKnownToBeAPowerOfTwo(LAnd->getOperand(1), DL, false, 0, AC, CxtI,
1713                                  DT) &&
1714           isKnownToBeAPowerOfTwo(RAnd->getOperand(1), DL, false, 0, AC, CxtI,
1715                                  DT)) {
1716         Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1));
1717         Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask);
1718       } else if (LAnd->getOperand(1) == RAnd->getOperand(1) &&
1719                  isKnownToBeAPowerOfTwo(LAnd->getOperand(0), DL, false, 0, AC,
1720                                         CxtI, DT) &&
1721                  isKnownToBeAPowerOfTwo(RAnd->getOperand(0), DL, false, 0, AC,
1722                                         CxtI, DT)) {
1723         Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0));
1724         Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask);
1725       }
1726 
1727       if (Masked)
1728         return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask);
1729     }
1730   }
1731 
1732   // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
1733   //                   -->  (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
1734   // The original condition actually refers to the following two ranges:
1735   // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
1736   // We can fold these two ranges if:
1737   // 1) C1 and C2 is unsigned greater than C3.
1738   // 2) The two ranges are separated.
1739   // 3) C1 ^ C2 is one-bit mask.
1740   // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
1741   // This implies all values in the two ranges differ by exactly one bit.
1742 
1743   if ((LHSCC == ICmpInst::ICMP_ULT || LHSCC == ICmpInst::ICMP_ULE) &&
1744       LHSCC == RHSCC && LHSCst && RHSCst && LHS->hasOneUse() &&
1745       RHS->hasOneUse() && LHSCst->getType() == RHSCst->getType() &&
1746       LHSCst->getValue() == (RHSCst->getValue())) {
1747 
1748     Value *LAdd = LHS->getOperand(0);
1749     Value *RAdd = RHS->getOperand(0);
1750 
1751     Value *LAddOpnd, *RAddOpnd;
1752     ConstantInt *LAddCst, *RAddCst;
1753     if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddCst))) &&
1754         match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddCst))) &&
1755         LAddCst->getValue().ugt(LHSCst->getValue()) &&
1756         RAddCst->getValue().ugt(LHSCst->getValue())) {
1757 
1758       APInt DiffCst = LAddCst->getValue() ^ RAddCst->getValue();
1759       if (LAddOpnd == RAddOpnd && DiffCst.isPowerOf2()) {
1760         ConstantInt *MaxAddCst = nullptr;
1761         if (LAddCst->getValue().ult(RAddCst->getValue()))
1762           MaxAddCst = RAddCst;
1763         else
1764           MaxAddCst = LAddCst;
1765 
1766         APInt RRangeLow = -RAddCst->getValue();
1767         APInt RRangeHigh = RRangeLow + LHSCst->getValue();
1768         APInt LRangeLow = -LAddCst->getValue();
1769         APInt LRangeHigh = LRangeLow + LHSCst->getValue();
1770         APInt LowRangeDiff = RRangeLow ^ LRangeLow;
1771         APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
1772         APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
1773                                                    : RRangeLow - LRangeLow;
1774 
1775         if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
1776             RangeDiff.ugt(LHSCst->getValue())) {
1777           Value *MaskCst = ConstantInt::get(LAddCst->getType(), ~DiffCst);
1778 
1779           Value *NewAnd = Builder->CreateAnd(LAddOpnd, MaskCst);
1780           Value *NewAdd = Builder->CreateAdd(NewAnd, MaxAddCst);
1781           return (Builder->CreateICmp(LHS->getPredicate(), NewAdd, LHSCst));
1782         }
1783       }
1784     }
1785   }
1786 
1787   // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1788   if (PredicatesFoldable(LHSCC, RHSCC)) {
1789     if (LHS->getOperand(0) == RHS->getOperand(1) &&
1790         LHS->getOperand(1) == RHS->getOperand(0))
1791       LHS->swapOperands();
1792     if (LHS->getOperand(0) == RHS->getOperand(0) &&
1793         LHS->getOperand(1) == RHS->getOperand(1)) {
1794       Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1795       unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1796       bool isSigned = LHS->isSigned() || RHS->isSigned();
1797       return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1798     }
1799   }
1800 
1801   // handle (roughly):
1802   // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1803   if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
1804     return V;
1805 
1806   Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1807   if (LHS->hasOneUse() || RHS->hasOneUse()) {
1808     // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
1809     // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
1810     Value *A = nullptr, *B = nullptr;
1811     if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
1812       B = Val;
1813       if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
1814         A = Val2;
1815       else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
1816         A = RHS->getOperand(1);
1817     }
1818     // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
1819     // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
1820     else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1821       B = Val2;
1822       if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
1823         A = Val;
1824       else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
1825         A = LHS->getOperand(1);
1826     }
1827     if (A && B)
1828       return Builder->CreateICmp(
1829           ICmpInst::ICMP_UGE,
1830           Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
1831   }
1832 
1833   // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
1834   if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
1835     return V;
1836 
1837   // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
1838   if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
1839     return V;
1840 
1841   // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1842   if (!LHSCst || !RHSCst) return nullptr;
1843 
1844   if (LHSCst == RHSCst && LHSCC == RHSCC) {
1845     // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1846     if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1847       Value *NewOr = Builder->CreateOr(Val, Val2);
1848       return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1849     }
1850   }
1851 
1852   // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1853   //   iff C2 + CA == C1.
1854   if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1855     ConstantInt *AddCst;
1856     if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1857       if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1858         return Builder->CreateICmpULE(Val, LHSCst);
1859   }
1860 
1861   // From here on, we only handle:
1862   //    (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1863   if (Val != Val2) return nullptr;
1864 
1865   // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1866   if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1867       RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1868       LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1869       RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1870     return nullptr;
1871 
1872   // We can't fold (ugt x, C) | (sgt x, C2).
1873   if (!PredicatesFoldable(LHSCC, RHSCC))
1874     return nullptr;
1875 
1876   // Ensure that the larger constant is on the RHS.
1877   bool ShouldSwap;
1878   if (CmpInst::isSigned(LHSCC) ||
1879       (ICmpInst::isEquality(LHSCC) &&
1880        CmpInst::isSigned(RHSCC)))
1881     ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1882   else
1883     ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1884 
1885   if (ShouldSwap) {
1886     std::swap(LHS, RHS);
1887     std::swap(LHSCst, RHSCst);
1888     std::swap(LHSCC, RHSCC);
1889   }
1890 
1891   // At this point, we know we have two icmp instructions
1892   // comparing a value against two constants and or'ing the result
1893   // together.  Because of the above check, we know that we only have
1894   // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1895   // icmp folding check above), that the two constants are not
1896   // equal.
1897   assert(LHSCst != RHSCst && "Compares not folded above?");
1898 
1899   switch (LHSCC) {
1900   default: llvm_unreachable("Unknown integer condition code!");
1901   case ICmpInst::ICMP_EQ:
1902     switch (RHSCC) {
1903     default: llvm_unreachable("Unknown integer condition code!");
1904     case ICmpInst::ICMP_EQ:
1905       if (LHS->getOperand(0) == RHS->getOperand(0)) {
1906         // if LHSCst and RHSCst differ only by one bit:
1907         // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
1908         assert(LHSCst->getValue().ule(LHSCst->getValue()));
1909 
1910         APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
1911         if (Xor.isPowerOf2()) {
1912           Value *NegCst = Builder->getInt(~Xor);
1913           Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
1914           return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
1915         }
1916       }
1917 
1918       if (LHSCst == SubOne(RHSCst)) {
1919         // (X == 13 | X == 14) -> X-13 <u 2
1920         Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1921         Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1922         AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1923         return Builder->CreateICmpULT(Add, AddCST);
1924       }
1925 
1926       break;                         // (X == 13 | X == 15) -> no change
1927     case ICmpInst::ICMP_UGT:         // (X == 13 | X u> 14) -> no change
1928     case ICmpInst::ICMP_SGT:         // (X == 13 | X s> 14) -> no change
1929       break;
1930     case ICmpInst::ICMP_NE:          // (X == 13 | X != 15) -> X != 15
1931     case ICmpInst::ICMP_ULT:         // (X == 13 | X u< 15) -> X u< 15
1932     case ICmpInst::ICMP_SLT:         // (X == 13 | X s< 15) -> X s< 15
1933       return RHS;
1934     }
1935     break;
1936   case ICmpInst::ICMP_NE:
1937     switch (RHSCC) {
1938     default: llvm_unreachable("Unknown integer condition code!");
1939     case ICmpInst::ICMP_EQ:          // (X != 13 | X == 15) -> X != 13
1940     case ICmpInst::ICMP_UGT:         // (X != 13 | X u> 15) -> X != 13
1941     case ICmpInst::ICMP_SGT:         // (X != 13 | X s> 15) -> X != 13
1942       return LHS;
1943     case ICmpInst::ICMP_NE:          // (X != 13 | X != 15) -> true
1944     case ICmpInst::ICMP_ULT:         // (X != 13 | X u< 15) -> true
1945     case ICmpInst::ICMP_SLT:         // (X != 13 | X s< 15) -> true
1946       return Builder->getTrue();
1947     }
1948   case ICmpInst::ICMP_ULT:
1949     switch (RHSCC) {
1950     default: llvm_unreachable("Unknown integer condition code!");
1951     case ICmpInst::ICMP_EQ:         // (X u< 13 | X == 14) -> no change
1952       break;
1953     case ICmpInst::ICMP_UGT:        // (X u< 13 | X u> 15) -> (X-13) u> 2
1954       // If RHSCst is [us]MAXINT, it is always false.  Not handling
1955       // this can cause overflow.
1956       if (RHSCst->isMaxValue(false))
1957         return LHS;
1958       return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1959     case ICmpInst::ICMP_SGT:        // (X u< 13 | X s> 15) -> no change
1960       break;
1961     case ICmpInst::ICMP_NE:         // (X u< 13 | X != 15) -> X != 15
1962     case ICmpInst::ICMP_ULT:        // (X u< 13 | X u< 15) -> X u< 15
1963       return RHS;
1964     case ICmpInst::ICMP_SLT:        // (X u< 13 | X s< 15) -> no change
1965       break;
1966     }
1967     break;
1968   case ICmpInst::ICMP_SLT:
1969     switch (RHSCC) {
1970     default: llvm_unreachable("Unknown integer condition code!");
1971     case ICmpInst::ICMP_EQ:         // (X s< 13 | X == 14) -> no change
1972       break;
1973     case ICmpInst::ICMP_SGT:        // (X s< 13 | X s> 15) -> (X-13) s> 2
1974       // If RHSCst is [us]MAXINT, it is always false.  Not handling
1975       // this can cause overflow.
1976       if (RHSCst->isMaxValue(true))
1977         return LHS;
1978       return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1979     case ICmpInst::ICMP_UGT:        // (X s< 13 | X u> 15) -> no change
1980       break;
1981     case ICmpInst::ICMP_NE:         // (X s< 13 | X != 15) -> X != 15
1982     case ICmpInst::ICMP_SLT:        // (X s< 13 | X s< 15) -> X s< 15
1983       return RHS;
1984     case ICmpInst::ICMP_ULT:        // (X s< 13 | X u< 15) -> no change
1985       break;
1986     }
1987     break;
1988   case ICmpInst::ICMP_UGT:
1989     switch (RHSCC) {
1990     default: llvm_unreachable("Unknown integer condition code!");
1991     case ICmpInst::ICMP_EQ:         // (X u> 13 | X == 15) -> X u> 13
1992     case ICmpInst::ICMP_UGT:        // (X u> 13 | X u> 15) -> X u> 13
1993       return LHS;
1994     case ICmpInst::ICMP_SGT:        // (X u> 13 | X s> 15) -> no change
1995       break;
1996     case ICmpInst::ICMP_NE:         // (X u> 13 | X != 15) -> true
1997     case ICmpInst::ICMP_ULT:        // (X u> 13 | X u< 15) -> true
1998       return Builder->getTrue();
1999     case ICmpInst::ICMP_SLT:        // (X u> 13 | X s< 15) -> no change
2000       break;
2001     }
2002     break;
2003   case ICmpInst::ICMP_SGT:
2004     switch (RHSCC) {
2005     default: llvm_unreachable("Unknown integer condition code!");
2006     case ICmpInst::ICMP_EQ:         // (X s> 13 | X == 15) -> X > 13
2007     case ICmpInst::ICMP_SGT:        // (X s> 13 | X s> 15) -> X > 13
2008       return LHS;
2009     case ICmpInst::ICMP_UGT:        // (X s> 13 | X u> 15) -> no change
2010       break;
2011     case ICmpInst::ICMP_NE:         // (X s> 13 | X != 15) -> true
2012     case ICmpInst::ICMP_SLT:        // (X s> 13 | X s< 15) -> true
2013       return Builder->getTrue();
2014     case ICmpInst::ICMP_ULT:        // (X s> 13 | X u< 15) -> no change
2015       break;
2016     }
2017     break;
2018   }
2019   return nullptr;
2020 }
2021 
2022 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp).  NOTE: Unlike the rest of
2023 /// instcombine, this returns a Value which should already be inserted into the
2024 /// function.
FoldOrOfFCmps(FCmpInst * LHS,FCmpInst * RHS)2025 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
2026   if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
2027       RHS->getPredicate() == FCmpInst::FCMP_UNO &&
2028       LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
2029     if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
2030       if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
2031         // If either of the constants are nans, then the whole thing returns
2032         // true.
2033         if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
2034           return Builder->getTrue();
2035 
2036         // Otherwise, no need to compare the two constants, compare the
2037         // rest.
2038         return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2039       }
2040 
2041     // Handle vector zeros.  This occurs because the canonical form of
2042     // "fcmp uno x,x" is "fcmp uno x, 0".
2043     if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
2044         isa<ConstantAggregateZero>(RHS->getOperand(1)))
2045       return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2046 
2047     return nullptr;
2048   }
2049 
2050   Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
2051   Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
2052   FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
2053 
2054   if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
2055     // Swap RHS operands to match LHS.
2056     Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
2057     std::swap(Op1LHS, Op1RHS);
2058   }
2059   if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
2060     // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
2061     if (Op0CC == Op1CC)
2062       return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
2063     if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
2064       return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
2065     if (Op0CC == FCmpInst::FCMP_FALSE)
2066       return RHS;
2067     if (Op1CC == FCmpInst::FCMP_FALSE)
2068       return LHS;
2069     bool Op0Ordered;
2070     bool Op1Ordered;
2071     unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
2072     unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
2073     if (Op0Ordered == Op1Ordered) {
2074       // If both are ordered or unordered, return a new fcmp with
2075       // or'ed predicates.
2076       return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
2077     }
2078   }
2079   return nullptr;
2080 }
2081 
2082 /// FoldOrWithConstants - This helper function folds:
2083 ///
2084 ///     ((A | B) & C1) | (B & C2)
2085 ///
2086 /// into:
2087 ///
2088 ///     (A & C1) | B
2089 ///
2090 /// when the XOR of the two constants is "all ones" (-1).
FoldOrWithConstants(BinaryOperator & I,Value * Op,Value * A,Value * B,Value * C)2091 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
2092                                                Value *A, Value *B, Value *C) {
2093   ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2094   if (!CI1) return nullptr;
2095 
2096   Value *V1 = nullptr;
2097   ConstantInt *CI2 = nullptr;
2098   if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr;
2099 
2100   APInt Xor = CI1->getValue() ^ CI2->getValue();
2101   if (!Xor.isAllOnesValue()) return nullptr;
2102 
2103   if (V1 == A || V1 == B) {
2104     Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
2105     return BinaryOperator::CreateOr(NewOp, V1);
2106   }
2107 
2108   return nullptr;
2109 }
2110 
2111 /// \brief This helper function folds:
2112 ///
2113 ///     ((A | B) & C1) ^ (B & C2)
2114 ///
2115 /// into:
2116 ///
2117 ///     (A & C1) ^ B
2118 ///
2119 /// when the XOR of the two constants is "all ones" (-1).
FoldXorWithConstants(BinaryOperator & I,Value * Op,Value * A,Value * B,Value * C)2120 Instruction *InstCombiner::FoldXorWithConstants(BinaryOperator &I, Value *Op,
2121                                                 Value *A, Value *B, Value *C) {
2122   ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2123   if (!CI1)
2124     return nullptr;
2125 
2126   Value *V1 = nullptr;
2127   ConstantInt *CI2 = nullptr;
2128   if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2))))
2129     return nullptr;
2130 
2131   APInt Xor = CI1->getValue() ^ CI2->getValue();
2132   if (!Xor.isAllOnesValue())
2133     return nullptr;
2134 
2135   if (V1 == A || V1 == B) {
2136     Value *NewOp = Builder->CreateAnd(V1 == A ? B : A, CI1);
2137     return BinaryOperator::CreateXor(NewOp, V1);
2138   }
2139 
2140   return nullptr;
2141 }
2142 
visitOr(BinaryOperator & I)2143 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2144   bool Changed = SimplifyAssociativeOrCommutative(I);
2145   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2146 
2147   if (Value *V = SimplifyVectorOp(I))
2148     return ReplaceInstUsesWith(I, V);
2149 
2150   if (Value *V = SimplifyOrInst(Op0, Op1, DL, TLI, DT, AC))
2151     return ReplaceInstUsesWith(I, V);
2152 
2153   // (A&B)|(A&C) -> A&(B|C) etc
2154   if (Value *V = SimplifyUsingDistributiveLaws(I))
2155     return ReplaceInstUsesWith(I, V);
2156 
2157   // See if we can simplify any instructions used by the instruction whose sole
2158   // purpose is to compute bits we don't care about.
2159   if (SimplifyDemandedInstructionBits(I))
2160     return &I;
2161 
2162   if (Value *V = SimplifyBSwap(I))
2163     return ReplaceInstUsesWith(I, V);
2164 
2165   if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2166     ConstantInt *C1 = nullptr; Value *X = nullptr;
2167     // (X & C1) | C2 --> (X | C2) & (C1|C2)
2168     // iff (C1 & C2) == 0.
2169     if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
2170         (RHS->getValue() & C1->getValue()) != 0 &&
2171         Op0->hasOneUse()) {
2172       Value *Or = Builder->CreateOr(X, RHS);
2173       Or->takeName(Op0);
2174       return BinaryOperator::CreateAnd(Or,
2175                              Builder->getInt(RHS->getValue() | C1->getValue()));
2176     }
2177 
2178     // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2179     if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
2180         Op0->hasOneUse()) {
2181       Value *Or = Builder->CreateOr(X, RHS);
2182       Or->takeName(Op0);
2183       return BinaryOperator::CreateXor(Or,
2184                             Builder->getInt(C1->getValue() & ~RHS->getValue()));
2185     }
2186 
2187     // Try to fold constant and into select arguments.
2188     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2189       if (Instruction *R = FoldOpIntoSelect(I, SI))
2190         return R;
2191 
2192     if (isa<PHINode>(Op0))
2193       if (Instruction *NV = FoldOpIntoPhi(I))
2194         return NV;
2195   }
2196 
2197   Value *A = nullptr, *B = nullptr;
2198   ConstantInt *C1 = nullptr, *C2 = nullptr;
2199 
2200   // (A | B) | C  and  A | (B | C)                  -> bswap if possible.
2201   // (A >> B) | (C << D)  and  (A << B) | (B >> C)  -> bswap if possible.
2202   if (match(Op0, m_Or(m_Value(), m_Value())) ||
2203       match(Op1, m_Or(m_Value(), m_Value())) ||
2204       (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
2205        match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
2206     if (Instruction *BSwap = MatchBSwap(I))
2207       return BSwap;
2208   }
2209 
2210   // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2211   if (Op0->hasOneUse() &&
2212       match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2213       MaskedValueIsZero(Op1, C1->getValue(), 0, &I)) {
2214     Value *NOr = Builder->CreateOr(A, Op1);
2215     NOr->takeName(Op0);
2216     return BinaryOperator::CreateXor(NOr, C1);
2217   }
2218 
2219   // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2220   if (Op1->hasOneUse() &&
2221       match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2222       MaskedValueIsZero(Op0, C1->getValue(), 0, &I)) {
2223     Value *NOr = Builder->CreateOr(A, Op0);
2224     NOr->takeName(Op0);
2225     return BinaryOperator::CreateXor(NOr, C1);
2226   }
2227 
2228   // ((~A & B) | A) -> (A | B)
2229   if (match(Op0, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2230       match(Op1, m_Specific(A)))
2231     return BinaryOperator::CreateOr(A, B);
2232 
2233   // ((A & B) | ~A) -> (~A | B)
2234   if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2235       match(Op1, m_Not(m_Specific(A))))
2236     return BinaryOperator::CreateOr(Builder->CreateNot(A), B);
2237 
2238   // (A & (~B)) | (A ^ B) -> (A ^ B)
2239   if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2240       match(Op1, m_Xor(m_Specific(A), m_Specific(B))))
2241     return BinaryOperator::CreateXor(A, B);
2242 
2243   // (A ^ B) | ( A & (~B)) -> (A ^ B)
2244   if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2245       match(Op1, m_And(m_Specific(A), m_Not(m_Specific(B)))))
2246     return BinaryOperator::CreateXor(A, B);
2247 
2248   // (A & C)|(B & D)
2249   Value *C = nullptr, *D = nullptr;
2250   if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2251       match(Op1, m_And(m_Value(B), m_Value(D)))) {
2252     Value *V1 = nullptr, *V2 = nullptr;
2253     C1 = dyn_cast<ConstantInt>(C);
2254     C2 = dyn_cast<ConstantInt>(D);
2255     if (C1 && C2) {  // (A & C1)|(B & C2)
2256       if ((C1->getValue() & C2->getValue()) == 0) {
2257         // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2258         // iff (C1&C2) == 0 and (N&~C1) == 0
2259         if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2260             ((V1 == B &&
2261               MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2262              (V2 == B &&
2263               MaskedValueIsZero(V1, ~C1->getValue(), 0, &I))))  // (N|V)
2264           return BinaryOperator::CreateAnd(A,
2265                                 Builder->getInt(C1->getValue()|C2->getValue()));
2266         // Or commutes, try both ways.
2267         if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2268             ((V1 == A &&
2269               MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2270              (V2 == A &&
2271               MaskedValueIsZero(V1, ~C2->getValue(), 0, &I))))  // (N|V)
2272           return BinaryOperator::CreateAnd(B,
2273                                 Builder->getInt(C1->getValue()|C2->getValue()));
2274 
2275         // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2276         // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2277         ConstantInt *C3 = nullptr, *C4 = nullptr;
2278         if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2279             (C3->getValue() & ~C1->getValue()) == 0 &&
2280             match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2281             (C4->getValue() & ~C2->getValue()) == 0) {
2282           V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2283           return BinaryOperator::CreateAnd(V2,
2284                                 Builder->getInt(C1->getValue()|C2->getValue()));
2285         }
2286       }
2287     }
2288 
2289     // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) ->  C0 ? A : B, and commuted variants.
2290     // Don't do this for vector select idioms, the code generator doesn't handle
2291     // them well yet.
2292     if (!I.getType()->isVectorTy()) {
2293       if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
2294         return Match;
2295       if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
2296         return Match;
2297       if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
2298         return Match;
2299       if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
2300         return Match;
2301     }
2302 
2303     // ((A&~B)|(~A&B)) -> A^B
2304     if ((match(C, m_Not(m_Specific(D))) &&
2305          match(B, m_Not(m_Specific(A)))))
2306       return BinaryOperator::CreateXor(A, D);
2307     // ((~B&A)|(~A&B)) -> A^B
2308     if ((match(A, m_Not(m_Specific(D))) &&
2309          match(B, m_Not(m_Specific(C)))))
2310       return BinaryOperator::CreateXor(C, D);
2311     // ((A&~B)|(B&~A)) -> A^B
2312     if ((match(C, m_Not(m_Specific(B))) &&
2313          match(D, m_Not(m_Specific(A)))))
2314       return BinaryOperator::CreateXor(A, B);
2315     // ((~B&A)|(B&~A)) -> A^B
2316     if ((match(A, m_Not(m_Specific(B))) &&
2317          match(D, m_Not(m_Specific(C)))))
2318       return BinaryOperator::CreateXor(C, B);
2319 
2320     // ((A|B)&1)|(B&-2) -> (A&1) | B
2321     if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
2322         match(A, m_Or(m_Specific(B), m_Value(V1)))) {
2323       Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
2324       if (Ret) return Ret;
2325     }
2326     // (B&-2)|((A|B)&1) -> (A&1) | B
2327     if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
2328         match(B, m_Or(m_Value(V1), m_Specific(A)))) {
2329       Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
2330       if (Ret) return Ret;
2331     }
2332     // ((A^B)&1)|(B&-2) -> (A&1) ^ B
2333     if (match(A, m_Xor(m_Value(V1), m_Specific(B))) ||
2334         match(A, m_Xor(m_Specific(B), m_Value(V1)))) {
2335       Instruction *Ret = FoldXorWithConstants(I, Op1, V1, B, C);
2336       if (Ret) return Ret;
2337     }
2338     // (B&-2)|((A^B)&1) -> (A&1) ^ B
2339     if (match(B, m_Xor(m_Specific(A), m_Value(V1))) ||
2340         match(B, m_Xor(m_Value(V1), m_Specific(A)))) {
2341       Instruction *Ret = FoldXorWithConstants(I, Op0, A, V1, D);
2342       if (Ret) return Ret;
2343     }
2344   }
2345 
2346   // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2347   if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2348     if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2349       if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
2350         return BinaryOperator::CreateOr(Op0, C);
2351 
2352   // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2353   if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2354     if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2355       if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
2356         return BinaryOperator::CreateOr(Op1, C);
2357 
2358   // ((B | C) & A) | B -> B | (A & C)
2359   if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2360     return BinaryOperator::CreateOr(Op1, Builder->CreateAnd(A, C));
2361 
2362   // (~A | ~B) == (~(A & B)) - De Morgan's Law
2363   if (Value *Op0NotVal = dyn_castNotVal(Op0))
2364     if (Value *Op1NotVal = dyn_castNotVal(Op1))
2365       if (Op0->hasOneUse() && Op1->hasOneUse()) {
2366         Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
2367                                         I.getName()+".demorgan");
2368         return BinaryOperator::CreateNot(And);
2369       }
2370 
2371   // Canonicalize xor to the RHS.
2372   bool SwappedForXor = false;
2373   if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2374     std::swap(Op0, Op1);
2375     SwappedForXor = true;
2376   }
2377 
2378   // A | ( A ^ B) -> A |  B
2379   // A | (~A ^ B) -> A | ~B
2380   // (A & B) | (A ^ B)
2381   if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2382     if (Op0 == A || Op0 == B)
2383       return BinaryOperator::CreateOr(A, B);
2384 
2385     if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2386         match(Op0, m_And(m_Specific(B), m_Specific(A))))
2387       return BinaryOperator::CreateOr(A, B);
2388 
2389     if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2390       Value *Not = Builder->CreateNot(B, B->getName()+".not");
2391       return BinaryOperator::CreateOr(Not, Op0);
2392     }
2393     if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2394       Value *Not = Builder->CreateNot(A, A->getName()+".not");
2395       return BinaryOperator::CreateOr(Not, Op0);
2396     }
2397   }
2398 
2399   // A | ~(A | B) -> A | ~B
2400   // A | ~(A ^ B) -> A | ~B
2401   if (match(Op1, m_Not(m_Value(A))))
2402     if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2403       if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2404           Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2405                                B->getOpcode() == Instruction::Xor)) {
2406         Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2407                                                  B->getOperand(0);
2408         Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
2409         return BinaryOperator::CreateOr(Not, Op0);
2410       }
2411 
2412   // (A & B) | ((~A) ^ B) -> (~A ^ B)
2413   if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2414       match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
2415     return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2416 
2417   // ((~A) ^ B) | (A & B) -> (~A ^ B)
2418   if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
2419       match(Op1, m_And(m_Specific(A), m_Specific(B))))
2420     return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2421 
2422   if (SwappedForXor)
2423     std::swap(Op0, Op1);
2424 
2425   {
2426     ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2427     ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2428     if (LHS && RHS)
2429       if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2430         return ReplaceInstUsesWith(I, Res);
2431 
2432     // TODO: Make this recursive; it's a little tricky because an arbitrary
2433     // number of 'or' instructions might have to be created.
2434     Value *X, *Y;
2435     if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2436       if (auto *Cmp = dyn_cast<ICmpInst>(X))
2437         if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2438           return ReplaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2439       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2440         if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2441           return ReplaceInstUsesWith(I, Builder->CreateOr(Res, X));
2442     }
2443     if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2444       if (auto *Cmp = dyn_cast<ICmpInst>(X))
2445         if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2446           return ReplaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2447       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2448         if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2449           return ReplaceInstUsesWith(I, Builder->CreateOr(Res, X));
2450     }
2451   }
2452 
2453   // (fcmp uno x, c) | (fcmp uno y, c)  -> (fcmp uno x, y)
2454   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2455     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2456       if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2457         return ReplaceInstUsesWith(I, Res);
2458 
2459   // fold (or (cast A), (cast B)) -> (cast (or A, B))
2460   if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2461     CastInst *Op1C = dyn_cast<CastInst>(Op1);
2462     if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
2463       Type *SrcTy = Op0C->getOperand(0)->getType();
2464       if (SrcTy == Op1C->getOperand(0)->getType() &&
2465           SrcTy->isIntOrIntVectorTy()) {
2466         Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
2467 
2468         if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
2469             // Only do this if the casts both really cause code to be
2470             // generated.
2471             ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
2472             ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
2473           Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
2474           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2475         }
2476 
2477         // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
2478         // cast is otherwise not optimizable.  This happens for vector sexts.
2479         if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
2480           if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
2481             if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2482               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2483 
2484         // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
2485         // cast is otherwise not optimizable.  This happens for vector sexts.
2486         if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
2487           if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
2488             if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2489               return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2490       }
2491     }
2492   }
2493 
2494   // or(sext(A), B) -> A ? -1 : B where A is an i1
2495   // or(A, sext(B)) -> B ? -1 : A where B is an i1
2496   if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2497     return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2498   if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2499     return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2500 
2501   // Note: If we've gotten to the point of visiting the outer OR, then the
2502   // inner one couldn't be simplified.  If it was a constant, then it won't
2503   // be simplified by a later pass either, so we try swapping the inner/outer
2504   // ORs in the hopes that we'll be able to simplify it this way.
2505   // (X|C) | V --> (X|V) | C
2506   if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2507       match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2508     Value *Inner = Builder->CreateOr(A, Op1);
2509     Inner->takeName(Op0);
2510     return BinaryOperator::CreateOr(Inner, C1);
2511   }
2512 
2513   // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2514   // Since this OR statement hasn't been optimized further yet, we hope
2515   // that this transformation will allow the new ORs to be optimized.
2516   {
2517     Value *X = nullptr, *Y = nullptr;
2518     if (Op0->hasOneUse() && Op1->hasOneUse() &&
2519         match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2520         match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2521       Value *orTrue = Builder->CreateOr(A, C);
2522       Value *orFalse = Builder->CreateOr(B, D);
2523       return SelectInst::Create(X, orTrue, orFalse);
2524     }
2525   }
2526 
2527   return Changed ? &I : nullptr;
2528 }
2529 
visitXor(BinaryOperator & I)2530 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2531   bool Changed = SimplifyAssociativeOrCommutative(I);
2532   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2533 
2534   if (Value *V = SimplifyVectorOp(I))
2535     return ReplaceInstUsesWith(I, V);
2536 
2537   if (Value *V = SimplifyXorInst(Op0, Op1, DL, TLI, DT, AC))
2538     return ReplaceInstUsesWith(I, V);
2539 
2540   // (A&B)^(A&C) -> A&(B^C) etc
2541   if (Value *V = SimplifyUsingDistributiveLaws(I))
2542     return ReplaceInstUsesWith(I, V);
2543 
2544   // See if we can simplify any instructions used by the instruction whose sole
2545   // purpose is to compute bits we don't care about.
2546   if (SimplifyDemandedInstructionBits(I))
2547     return &I;
2548 
2549   if (Value *V = SimplifyBSwap(I))
2550     return ReplaceInstUsesWith(I, V);
2551 
2552   // Is this a ~ operation?
2553   if (Value *NotOp = dyn_castNotVal(&I)) {
2554     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2555       if (Op0I->getOpcode() == Instruction::And ||
2556           Op0I->getOpcode() == Instruction::Or) {
2557         // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2558         // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2559         if (dyn_castNotVal(Op0I->getOperand(1)))
2560           Op0I->swapOperands();
2561         if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2562           Value *NotY =
2563             Builder->CreateNot(Op0I->getOperand(1),
2564                                Op0I->getOperand(1)->getName()+".not");
2565           if (Op0I->getOpcode() == Instruction::And)
2566             return BinaryOperator::CreateOr(Op0NotVal, NotY);
2567           return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2568         }
2569 
2570         // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2571         // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2572         if (IsFreeToInvert(Op0I->getOperand(0),
2573                            Op0I->getOperand(0)->hasOneUse()) &&
2574             IsFreeToInvert(Op0I->getOperand(1),
2575                            Op0I->getOperand(1)->hasOneUse())) {
2576           Value *NotX =
2577             Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2578           Value *NotY =
2579             Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2580           if (Op0I->getOpcode() == Instruction::And)
2581             return BinaryOperator::CreateOr(NotX, NotY);
2582           return BinaryOperator::CreateAnd(NotX, NotY);
2583         }
2584 
2585       } else if (Op0I->getOpcode() == Instruction::AShr) {
2586         // ~(~X >>s Y) --> (X >>s Y)
2587         if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2588           return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2589       }
2590     }
2591   }
2592 
2593   if (Constant *RHS = dyn_cast<Constant>(Op1)) {
2594     if (RHS->isAllOnesValue() && Op0->hasOneUse())
2595       // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2596       if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2597         return CmpInst::Create(CI->getOpcode(),
2598                                CI->getInversePredicate(),
2599                                CI->getOperand(0), CI->getOperand(1));
2600   }
2601 
2602   if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2603     // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2604     if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2605       if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2606         if (CI->hasOneUse() && Op0C->hasOneUse()) {
2607           Instruction::CastOps Opcode = Op0C->getOpcode();
2608           if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2609               (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
2610                                             Op0C->getDestTy()))) {
2611             CI->setPredicate(CI->getInversePredicate());
2612             return CastInst::Create(Opcode, CI, Op0C->getType());
2613           }
2614         }
2615       }
2616     }
2617 
2618     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2619       // ~(c-X) == X-c-1 == X+(-c-1)
2620       if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2621         if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2622           Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2623           Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2624                                       ConstantInt::get(I.getType(), 1));
2625           return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2626         }
2627 
2628       if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2629         if (Op0I->getOpcode() == Instruction::Add) {
2630           // ~(X-c) --> (-c-1)-X
2631           if (RHS->isAllOnesValue()) {
2632             Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2633             return BinaryOperator::CreateSub(
2634                            ConstantExpr::getSub(NegOp0CI,
2635                                       ConstantInt::get(I.getType(), 1)),
2636                                       Op0I->getOperand(0));
2637           } else if (RHS->getValue().isSignBit()) {
2638             // (X + C) ^ signbit -> (X + C + signbit)
2639             Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
2640             return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2641 
2642           }
2643         } else if (Op0I->getOpcode() == Instruction::Or) {
2644           // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2645           if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue(),
2646                                 0, &I)) {
2647             Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2648             // Anything in both C1 and C2 is known to be zero, remove it from
2649             // NewRHS.
2650             Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2651             NewRHS = ConstantExpr::getAnd(NewRHS,
2652                                        ConstantExpr::getNot(CommonBits));
2653             Worklist.Add(Op0I);
2654             I.setOperand(0, Op0I->getOperand(0));
2655             I.setOperand(1, NewRHS);
2656             return &I;
2657           }
2658         } else if (Op0I->getOpcode() == Instruction::LShr) {
2659           // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2660           // E1 = "X ^ C1"
2661           BinaryOperator *E1;
2662           ConstantInt *C1;
2663           if (Op0I->hasOneUse() &&
2664               (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2665               E1->getOpcode() == Instruction::Xor &&
2666               (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2667             // fold (C1 >> C2) ^ C3
2668             ConstantInt *C2 = Op0CI, *C3 = RHS;
2669             APInt FoldConst = C1->getValue().lshr(C2->getValue());
2670             FoldConst ^= C3->getValue();
2671             // Prepare the two operands.
2672             Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
2673             Opnd0->takeName(Op0I);
2674             cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2675             Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2676 
2677             return BinaryOperator::CreateXor(Opnd0, FoldVal);
2678           }
2679         }
2680       }
2681     }
2682 
2683     // Try to fold constant and into select arguments.
2684     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2685       if (Instruction *R = FoldOpIntoSelect(I, SI))
2686         return R;
2687     if (isa<PHINode>(Op0))
2688       if (Instruction *NV = FoldOpIntoPhi(I))
2689         return NV;
2690   }
2691 
2692   BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2693   if (Op1I) {
2694     Value *A, *B;
2695     if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2696       if (A == Op0) {              // B^(B|A) == (A|B)^B
2697         Op1I->swapOperands();
2698         I.swapOperands();
2699         std::swap(Op0, Op1);
2700       } else if (B == Op0) {       // B^(A|B) == (A|B)^B
2701         I.swapOperands();     // Simplified below.
2702         std::swap(Op0, Op1);
2703       }
2704     } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2705                Op1I->hasOneUse()){
2706       if (A == Op0) {                                      // A^(A&B) -> A^(B&A)
2707         Op1I->swapOperands();
2708         std::swap(A, B);
2709       }
2710       if (B == Op0) {                                      // A^(B&A) -> (B&A)^A
2711         I.swapOperands();     // Simplified below.
2712         std::swap(Op0, Op1);
2713       }
2714     }
2715   }
2716 
2717   BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2718   if (Op0I) {
2719     Value *A, *B;
2720     if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2721         Op0I->hasOneUse()) {
2722       if (A == Op1)                                  // (B|A)^B == (A|B)^B
2723         std::swap(A, B);
2724       if (B == Op1)                                  // (A|B)^B == A & ~B
2725         return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2726     } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2727                Op0I->hasOneUse()){
2728       if (A == Op1)                                        // (A&B)^A -> (B&A)^A
2729         std::swap(A, B);
2730       if (B == Op1 &&                                      // (B&A)^A == ~B & A
2731           !isa<ConstantInt>(Op1)) {  // Canonical form is (B&C)^C
2732         return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2733       }
2734     }
2735   }
2736 
2737   if (Op0I && Op1I) {
2738     Value *A, *B, *C, *D;
2739     // (A & B)^(A | B) -> A ^ B
2740     if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2741         match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2742       if ((A == C && B == D) || (A == D && B == C))
2743         return BinaryOperator::CreateXor(A, B);
2744     }
2745     // (A | B)^(A & B) -> A ^ B
2746     if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2747         match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2748       if ((A == C && B == D) || (A == D && B == C))
2749         return BinaryOperator::CreateXor(A, B);
2750     }
2751     // (A | ~B) ^ (~A | B) -> A ^ B
2752     if (match(Op0I, m_Or(m_Value(A), m_Not(m_Value(B)))) &&
2753         match(Op1I, m_Or(m_Not(m_Specific(A)), m_Specific(B)))) {
2754       return BinaryOperator::CreateXor(A, B);
2755     }
2756     // (~A | B) ^ (A | ~B) -> A ^ B
2757     if (match(Op0I, m_Or(m_Not(m_Value(A)), m_Value(B))) &&
2758         match(Op1I, m_Or(m_Specific(A), m_Not(m_Specific(B))))) {
2759       return BinaryOperator::CreateXor(A, B);
2760     }
2761     // (A & ~B) ^ (~A & B) -> A ^ B
2762     if (match(Op0I, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2763         match(Op1I, m_And(m_Not(m_Specific(A)), m_Specific(B)))) {
2764       return BinaryOperator::CreateXor(A, B);
2765     }
2766     // (~A & B) ^ (A & ~B) -> A ^ B
2767     if (match(Op0I, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2768         match(Op1I, m_And(m_Specific(A), m_Not(m_Specific(B))))) {
2769       return BinaryOperator::CreateXor(A, B);
2770     }
2771     // (A ^ C)^(A | B) -> ((~A) & B) ^ C
2772     if (match(Op0I, m_Xor(m_Value(D), m_Value(C))) &&
2773         match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2774       if (D == A)
2775         return BinaryOperator::CreateXor(
2776             Builder->CreateAnd(Builder->CreateNot(A), B), C);
2777       if (D == B)
2778         return BinaryOperator::CreateXor(
2779             Builder->CreateAnd(Builder->CreateNot(B), A), C);
2780     }
2781     // (A | B)^(A ^ C) -> ((~A) & B) ^ C
2782     if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2783         match(Op1I, m_Xor(m_Value(D), m_Value(C)))) {
2784       if (D == A)
2785         return BinaryOperator::CreateXor(
2786             Builder->CreateAnd(Builder->CreateNot(A), B), C);
2787       if (D == B)
2788         return BinaryOperator::CreateXor(
2789             Builder->CreateAnd(Builder->CreateNot(B), A), C);
2790     }
2791     // (A & B) ^ (A ^ B) -> (A | B)
2792     if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2793         match(Op1I, m_Xor(m_Specific(A), m_Specific(B))))
2794       return BinaryOperator::CreateOr(A, B);
2795     // (A ^ B) ^ (A & B) -> (A | B)
2796     if (match(Op0I, m_Xor(m_Value(A), m_Value(B))) &&
2797         match(Op1I, m_And(m_Specific(A), m_Specific(B))))
2798       return BinaryOperator::CreateOr(A, B);
2799   }
2800 
2801   Value *A = nullptr, *B = nullptr;
2802   // (A & ~B) ^ (~A) -> ~(A & B)
2803   if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2804       match(Op1, m_Not(m_Specific(A))))
2805     return BinaryOperator::CreateNot(Builder->CreateAnd(A, B));
2806 
2807   // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2808   if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2809     if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2810       if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2811         if (LHS->getOperand(0) == RHS->getOperand(1) &&
2812             LHS->getOperand(1) == RHS->getOperand(0))
2813           LHS->swapOperands();
2814         if (LHS->getOperand(0) == RHS->getOperand(0) &&
2815             LHS->getOperand(1) == RHS->getOperand(1)) {
2816           Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2817           unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2818           bool isSigned = LHS->isSigned() || RHS->isSigned();
2819           return ReplaceInstUsesWith(I,
2820                                getNewICmpValue(isSigned, Code, Op0, Op1,
2821                                                Builder));
2822         }
2823       }
2824 
2825   // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2826   if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2827     if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2828       if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2829         Type *SrcTy = Op0C->getOperand(0)->getType();
2830         if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2831             // Only do this if the casts both really cause code to be generated.
2832             ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2833                                I.getType()) &&
2834             ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2835                                I.getType())) {
2836           Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2837                                             Op1C->getOperand(0), I.getName());
2838           return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2839         }
2840       }
2841   }
2842 
2843   return Changed ? &I : nullptr;
2844 }
2845