• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
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 contains routines that help analyze properties that chains of
11 // computations have.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/MemoryBuiltins.h"
21 #include "llvm/Analysis/Loads.h"
22 #include "llvm/Analysis/LoopInfo.h"
23 #include "llvm/Analysis/VectorUtils.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/ConstantRange.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/GlobalAlias.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/Metadata.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/PatternMatch.h"
38 #include "llvm/IR/Statepoint.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/MathExtras.h"
41 #include <algorithm>
42 #include <array>
43 #include <cstring>
44 using namespace llvm;
45 using namespace llvm::PatternMatch;
46 
47 const unsigned MaxDepth = 6;
48 
49 // Controls the number of uses of the value searched for possible
50 // dominating comparisons.
51 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
52                                               cl::Hidden, cl::init(20));
53 
54 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
55 /// 0). For vector types, returns the element type's bitwidth.
getBitWidth(Type * Ty,const DataLayout & DL)56 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
57   if (unsigned BitWidth = Ty->getScalarSizeInBits())
58     return BitWidth;
59 
60   return DL.getPointerTypeSizeInBits(Ty);
61 }
62 
63 namespace {
64 // Simplifying using an assume can only be done in a particular control-flow
65 // context (the context instruction provides that context). If an assume and
66 // the context instruction are not in the same block then the DT helps in
67 // figuring out if we can use it.
68 struct Query {
69   const DataLayout &DL;
70   AssumptionCache *AC;
71   const Instruction *CxtI;
72   const DominatorTree *DT;
73 
74   /// Set of assumptions that should be excluded from further queries.
75   /// This is because of the potential for mutual recursion to cause
76   /// computeKnownBits to repeatedly visit the same assume intrinsic. The
77   /// classic case of this is assume(x = y), which will attempt to determine
78   /// bits in x from bits in y, which will attempt to determine bits in y from
79   /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
80   /// isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
81   /// isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so
82   /// on.
83   std::array<const Value*, MaxDepth> Excluded;
84   unsigned NumExcluded;
85 
Query__anon920e3c5e0111::Query86   Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
87         const DominatorTree *DT)
88       : DL(DL), AC(AC), CxtI(CxtI), DT(DT), NumExcluded(0) {}
89 
Query__anon920e3c5e0111::Query90   Query(const Query &Q, const Value *NewExcl)
91       : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), NumExcluded(Q.NumExcluded) {
92     Excluded = Q.Excluded;
93     Excluded[NumExcluded++] = NewExcl;
94     assert(NumExcluded <= Excluded.size());
95   }
96 
isExcluded__anon920e3c5e0111::Query97   bool isExcluded(const Value *Value) const {
98     if (NumExcluded == 0)
99       return false;
100     auto End = Excluded.begin() + NumExcluded;
101     return std::find(Excluded.begin(), End, Value) != End;
102   }
103 };
104 } // end anonymous namespace
105 
106 // Given the provided Value and, potentially, a context instruction, return
107 // the preferred context instruction (if any).
safeCxtI(const Value * V,const Instruction * CxtI)108 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
109   // If we've been provided with a context instruction, then use that (provided
110   // it has been inserted).
111   if (CxtI && CxtI->getParent())
112     return CxtI;
113 
114   // If the value is really an already-inserted instruction, then use that.
115   CxtI = dyn_cast<Instruction>(V);
116   if (CxtI && CxtI->getParent())
117     return CxtI;
118 
119   return nullptr;
120 }
121 
122 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
123                              unsigned Depth, const Query &Q);
124 
computeKnownBits(Value * V,APInt & KnownZero,APInt & KnownOne,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)125 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
126                             const DataLayout &DL, unsigned Depth,
127                             AssumptionCache *AC, const Instruction *CxtI,
128                             const DominatorTree *DT) {
129   ::computeKnownBits(V, KnownZero, KnownOne, Depth,
130                      Query(DL, AC, safeCxtI(V, CxtI), DT));
131 }
132 
haveNoCommonBitsSet(Value * LHS,Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)133 bool llvm::haveNoCommonBitsSet(Value *LHS, Value *RHS, const DataLayout &DL,
134                                AssumptionCache *AC, const Instruction *CxtI,
135                                const DominatorTree *DT) {
136   assert(LHS->getType() == RHS->getType() &&
137          "LHS and RHS should have the same type");
138   assert(LHS->getType()->isIntOrIntVectorTy() &&
139          "LHS and RHS should be integers");
140   IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
141   APInt LHSKnownZero(IT->getBitWidth(), 0), LHSKnownOne(IT->getBitWidth(), 0);
142   APInt RHSKnownZero(IT->getBitWidth(), 0), RHSKnownOne(IT->getBitWidth(), 0);
143   computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, 0, AC, CxtI, DT);
144   computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, 0, AC, CxtI, DT);
145   return (LHSKnownZero | RHSKnownZero).isAllOnesValue();
146 }
147 
148 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
149                            unsigned Depth, const Query &Q);
150 
ComputeSignBit(Value * V,bool & KnownZero,bool & KnownOne,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)151 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
152                           const DataLayout &DL, unsigned Depth,
153                           AssumptionCache *AC, const Instruction *CxtI,
154                           const DominatorTree *DT) {
155   ::ComputeSignBit(V, KnownZero, KnownOne, Depth,
156                    Query(DL, AC, safeCxtI(V, CxtI), DT));
157 }
158 
159 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
160                                    const Query &Q);
161 
isKnownToBeAPowerOfTwo(Value * V,const DataLayout & DL,bool OrZero,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)162 bool llvm::isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, bool OrZero,
163                                   unsigned Depth, AssumptionCache *AC,
164                                   const Instruction *CxtI,
165                                   const DominatorTree *DT) {
166   return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
167                                   Query(DL, AC, safeCxtI(V, CxtI), DT));
168 }
169 
170 static bool isKnownNonZero(Value *V, unsigned Depth, const Query &Q);
171 
isKnownNonZero(Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)172 bool llvm::isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
173                           AssumptionCache *AC, const Instruction *CxtI,
174                           const DominatorTree *DT) {
175   return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
176 }
177 
isKnownNonNegative(Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)178 bool llvm::isKnownNonNegative(Value *V, const DataLayout &DL, unsigned Depth,
179                               AssumptionCache *AC, const Instruction *CxtI,
180                               const DominatorTree *DT) {
181   bool NonNegative, Negative;
182   ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
183   return NonNegative;
184 }
185 
isKnownPositive(Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)186 bool llvm::isKnownPositive(Value *V, const DataLayout &DL, unsigned Depth,
187                            AssumptionCache *AC, const Instruction *CxtI,
188                            const DominatorTree *DT) {
189   if (auto *CI = dyn_cast<ConstantInt>(V))
190     return CI->getValue().isStrictlyPositive();
191 
192   // TODO: We'd doing two recursive queries here.  We should factor this such
193   // that only a single query is needed.
194   return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) &&
195     isKnownNonZero(V, DL, Depth, AC, CxtI, DT);
196 }
197 
isKnownNegative(Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)198 bool llvm::isKnownNegative(Value *V, const DataLayout &DL, unsigned Depth,
199                            AssumptionCache *AC, const Instruction *CxtI,
200                            const DominatorTree *DT) {
201   bool NonNegative, Negative;
202   ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
203   return Negative;
204 }
205 
206 static bool isKnownNonEqual(Value *V1, Value *V2, const Query &Q);
207 
isKnownNonEqual(Value * V1,Value * V2,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)208 bool llvm::isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
209                           AssumptionCache *AC, const Instruction *CxtI,
210                           const DominatorTree *DT) {
211   return ::isKnownNonEqual(V1, V2, Query(DL, AC,
212                                          safeCxtI(V1, safeCxtI(V2, CxtI)),
213                                          DT));
214 }
215 
216 static bool MaskedValueIsZero(Value *V, const APInt &Mask, unsigned Depth,
217                               const Query &Q);
218 
MaskedValueIsZero(Value * V,const APInt & Mask,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)219 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
220                              unsigned Depth, AssumptionCache *AC,
221                              const Instruction *CxtI, const DominatorTree *DT) {
222   return ::MaskedValueIsZero(V, Mask, Depth,
223                              Query(DL, AC, safeCxtI(V, CxtI), DT));
224 }
225 
226 static unsigned ComputeNumSignBits(Value *V, unsigned Depth, const Query &Q);
227 
ComputeNumSignBits(Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)228 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL,
229                                   unsigned Depth, AssumptionCache *AC,
230                                   const Instruction *CxtI,
231                                   const DominatorTree *DT) {
232   return ::ComputeNumSignBits(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
233 }
234 
computeKnownBitsAddSub(bool Add,Value * Op0,Value * Op1,bool NSW,APInt & KnownZero,APInt & KnownOne,APInt & KnownZero2,APInt & KnownOne2,unsigned Depth,const Query & Q)235 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
236                                    APInt &KnownZero, APInt &KnownOne,
237                                    APInt &KnownZero2, APInt &KnownOne2,
238                                    unsigned Depth, const Query &Q) {
239   if (!Add) {
240     if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
241       // We know that the top bits of C-X are clear if X contains less bits
242       // than C (i.e. no wrap-around can happen).  For example, 20-X is
243       // positive if we can prove that X is >= 0 and < 16.
244       if (!CLHS->getValue().isNegative()) {
245         unsigned BitWidth = KnownZero.getBitWidth();
246         unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
247         // NLZ can't be BitWidth with no sign bit
248         APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
249         computeKnownBits(Op1, KnownZero2, KnownOne2, Depth + 1, Q);
250 
251         // If all of the MaskV bits are known to be zero, then we know the
252         // output top bits are zero, because we now know that the output is
253         // from [0-C].
254         if ((KnownZero2 & MaskV) == MaskV) {
255           unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
256           // Top bits known zero.
257           KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
258         }
259       }
260     }
261   }
262 
263   unsigned BitWidth = KnownZero.getBitWidth();
264 
265   // If an initial sequence of bits in the result is not needed, the
266   // corresponding bits in the operands are not needed.
267   APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
268   computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, Depth + 1, Q);
269   computeKnownBits(Op1, KnownZero2, KnownOne2, Depth + 1, Q);
270 
271   // Carry in a 1 for a subtract, rather than a 0.
272   APInt CarryIn(BitWidth, 0);
273   if (!Add) {
274     // Sum = LHS + ~RHS + 1
275     std::swap(KnownZero2, KnownOne2);
276     CarryIn.setBit(0);
277   }
278 
279   APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
280   APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
281 
282   // Compute known bits of the carry.
283   APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
284   APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
285 
286   // Compute set of known bits (where all three relevant bits are known).
287   APInt LHSKnown = LHSKnownZero | LHSKnownOne;
288   APInt RHSKnown = KnownZero2 | KnownOne2;
289   APInt CarryKnown = CarryKnownZero | CarryKnownOne;
290   APInt Known = LHSKnown & RHSKnown & CarryKnown;
291 
292   assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
293          "known bits of sum differ");
294 
295   // Compute known bits of the result.
296   KnownZero = ~PossibleSumOne & Known;
297   KnownOne = PossibleSumOne & Known;
298 
299   // Are we still trying to solve for the sign bit?
300   if (!Known.isNegative()) {
301     if (NSW) {
302       // Adding two non-negative numbers, or subtracting a negative number from
303       // a non-negative one, can't wrap into negative.
304       if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
305         KnownZero |= APInt::getSignBit(BitWidth);
306       // Adding two negative numbers, or subtracting a non-negative number from
307       // a negative one, can't wrap into non-negative.
308       else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
309         KnownOne |= APInt::getSignBit(BitWidth);
310     }
311   }
312 }
313 
computeKnownBitsMul(Value * Op0,Value * Op1,bool NSW,APInt & KnownZero,APInt & KnownOne,APInt & KnownZero2,APInt & KnownOne2,unsigned Depth,const Query & Q)314 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
315                                 APInt &KnownZero, APInt &KnownOne,
316                                 APInt &KnownZero2, APInt &KnownOne2,
317                                 unsigned Depth, const Query &Q) {
318   unsigned BitWidth = KnownZero.getBitWidth();
319   computeKnownBits(Op1, KnownZero, KnownOne, Depth + 1, Q);
320   computeKnownBits(Op0, KnownZero2, KnownOne2, Depth + 1, Q);
321 
322   bool isKnownNegative = false;
323   bool isKnownNonNegative = false;
324   // If the multiplication is known not to overflow, compute the sign bit.
325   if (NSW) {
326     if (Op0 == Op1) {
327       // The product of a number with itself is non-negative.
328       isKnownNonNegative = true;
329     } else {
330       bool isKnownNonNegativeOp1 = KnownZero.isNegative();
331       bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
332       bool isKnownNegativeOp1 = KnownOne.isNegative();
333       bool isKnownNegativeOp0 = KnownOne2.isNegative();
334       // The product of two numbers with the same sign is non-negative.
335       isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
336         (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
337       // The product of a negative number and a non-negative number is either
338       // negative or zero.
339       if (!isKnownNonNegative)
340         isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
341                            isKnownNonZero(Op0, Depth, Q)) ||
342                           (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
343                            isKnownNonZero(Op1, Depth, Q));
344     }
345   }
346 
347   // If low bits are zero in either operand, output low known-0 bits.
348   // Also compute a conservative estimate for high known-0 bits.
349   // More trickiness is possible, but this is sufficient for the
350   // interesting case of alignment computation.
351   KnownOne.clearAllBits();
352   unsigned TrailZ = KnownZero.countTrailingOnes() +
353                     KnownZero2.countTrailingOnes();
354   unsigned LeadZ =  std::max(KnownZero.countLeadingOnes() +
355                              KnownZero2.countLeadingOnes(),
356                              BitWidth) - BitWidth;
357 
358   TrailZ = std::min(TrailZ, BitWidth);
359   LeadZ = std::min(LeadZ, BitWidth);
360   KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
361               APInt::getHighBitsSet(BitWidth, LeadZ);
362 
363   // Only make use of no-wrap flags if we failed to compute the sign bit
364   // directly.  This matters if the multiplication always overflows, in
365   // which case we prefer to follow the result of the direct computation,
366   // though as the program is invoking undefined behaviour we can choose
367   // whatever we like here.
368   if (isKnownNonNegative && !KnownOne.isNegative())
369     KnownZero.setBit(BitWidth - 1);
370   else if (isKnownNegative && !KnownZero.isNegative())
371     KnownOne.setBit(BitWidth - 1);
372 }
373 
computeKnownBitsFromRangeMetadata(const MDNode & Ranges,APInt & KnownZero,APInt & KnownOne)374 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
375                                              APInt &KnownZero,
376                                              APInt &KnownOne) {
377   unsigned BitWidth = KnownZero.getBitWidth();
378   unsigned NumRanges = Ranges.getNumOperands() / 2;
379   assert(NumRanges >= 1);
380 
381   KnownZero.setAllBits();
382   KnownOne.setAllBits();
383 
384   for (unsigned i = 0; i < NumRanges; ++i) {
385     ConstantInt *Lower =
386         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
387     ConstantInt *Upper =
388         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
389     ConstantRange Range(Lower->getValue(), Upper->getValue());
390 
391     // The first CommonPrefixBits of all values in Range are equal.
392     unsigned CommonPrefixBits =
393         (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
394 
395     APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
396     KnownOne &= Range.getUnsignedMax() & Mask;
397     KnownZero &= ~Range.getUnsignedMax() & Mask;
398   }
399 }
400 
isEphemeralValueOf(Instruction * I,const Value * E)401 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
402   SmallVector<const Value *, 16> WorkSet(1, I);
403   SmallPtrSet<const Value *, 32> Visited;
404   SmallPtrSet<const Value *, 16> EphValues;
405 
406   // The instruction defining an assumption's condition itself is always
407   // considered ephemeral to that assumption (even if it has other
408   // non-ephemeral users). See r246696's test case for an example.
409   if (std::find(I->op_begin(), I->op_end(), E) != I->op_end())
410     return true;
411 
412   while (!WorkSet.empty()) {
413     const Value *V = WorkSet.pop_back_val();
414     if (!Visited.insert(V).second)
415       continue;
416 
417     // If all uses of this value are ephemeral, then so is this value.
418     if (std::all_of(V->user_begin(), V->user_end(),
419                     [&](const User *U) { return EphValues.count(U); })) {
420       if (V == E)
421         return true;
422 
423       EphValues.insert(V);
424       if (const User *U = dyn_cast<User>(V))
425         for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
426              J != JE; ++J) {
427           if (isSafeToSpeculativelyExecute(*J))
428             WorkSet.push_back(*J);
429         }
430     }
431   }
432 
433   return false;
434 }
435 
436 // Is this an intrinsic that cannot be speculated but also cannot trap?
isAssumeLikeIntrinsic(const Instruction * I)437 static bool isAssumeLikeIntrinsic(const Instruction *I) {
438   if (const CallInst *CI = dyn_cast<CallInst>(I))
439     if (Function *F = CI->getCalledFunction())
440       switch (F->getIntrinsicID()) {
441       default: break;
442       // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
443       case Intrinsic::assume:
444       case Intrinsic::dbg_declare:
445       case Intrinsic::dbg_value:
446       case Intrinsic::invariant_start:
447       case Intrinsic::invariant_end:
448       case Intrinsic::lifetime_start:
449       case Intrinsic::lifetime_end:
450       case Intrinsic::objectsize:
451       case Intrinsic::ptr_annotation:
452       case Intrinsic::var_annotation:
453         return true;
454       }
455 
456   return false;
457 }
458 
isValidAssumeForContext(Value * V,const Instruction * CxtI,const DominatorTree * DT)459 static bool isValidAssumeForContext(Value *V, const Instruction *CxtI,
460                                     const DominatorTree *DT) {
461   Instruction *Inv = cast<Instruction>(V);
462 
463   // There are two restrictions on the use of an assume:
464   //  1. The assume must dominate the context (or the control flow must
465   //     reach the assume whenever it reaches the context).
466   //  2. The context must not be in the assume's set of ephemeral values
467   //     (otherwise we will use the assume to prove that the condition
468   //     feeding the assume is trivially true, thus causing the removal of
469   //     the assume).
470 
471   if (DT) {
472     if (DT->dominates(Inv, CxtI)) {
473       return true;
474     } else if (Inv->getParent() == CxtI->getParent()) {
475       // The context comes first, but they're both in the same block. Make sure
476       // there is nothing in between that might interrupt the control flow.
477       for (BasicBlock::const_iterator I =
478              std::next(BasicBlock::const_iterator(CxtI)),
479                                       IE(Inv); I != IE; ++I)
480         if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
481           return false;
482 
483       return !isEphemeralValueOf(Inv, CxtI);
484     }
485 
486     return false;
487   }
488 
489   // When we don't have a DT, we do a limited search...
490   if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
491     return true;
492   } else if (Inv->getParent() == CxtI->getParent()) {
493     // Search forward from the assume until we reach the context (or the end
494     // of the block); the common case is that the assume will come first.
495     for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
496          IE = Inv->getParent()->end(); I != IE; ++I)
497       if (&*I == CxtI)
498         return true;
499 
500     // The context must come first...
501     for (BasicBlock::const_iterator I =
502            std::next(BasicBlock::const_iterator(CxtI)),
503                                     IE(Inv); I != IE; ++I)
504       if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
505         return false;
506 
507     return !isEphemeralValueOf(Inv, CxtI);
508   }
509 
510   return false;
511 }
512 
isValidAssumeForContext(const Instruction * I,const Instruction * CxtI,const DominatorTree * DT)513 bool llvm::isValidAssumeForContext(const Instruction *I,
514                                    const Instruction *CxtI,
515                                    const DominatorTree *DT) {
516   return ::isValidAssumeForContext(const_cast<Instruction *>(I), CxtI, DT);
517 }
518 
computeKnownBitsFromAssume(Value * V,APInt & KnownZero,APInt & KnownOne,unsigned Depth,const Query & Q)519 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
520                                        APInt &KnownOne, unsigned Depth,
521                                        const Query &Q) {
522   // Use of assumptions is context-sensitive. If we don't have a context, we
523   // cannot use them!
524   if (!Q.AC || !Q.CxtI)
525     return;
526 
527   unsigned BitWidth = KnownZero.getBitWidth();
528 
529   for (auto &AssumeVH : Q.AC->assumptions()) {
530     if (!AssumeVH)
531       continue;
532     CallInst *I = cast<CallInst>(AssumeVH);
533     assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
534            "Got assumption for the wrong function!");
535     if (Q.isExcluded(I))
536       continue;
537 
538     // Warning: This loop can end up being somewhat performance sensetive.
539     // We're running this loop for once for each value queried resulting in a
540     // runtime of ~O(#assumes * #values).
541 
542     assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
543            "must be an assume intrinsic");
544 
545     Value *Arg = I->getArgOperand(0);
546 
547     if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
548       assert(BitWidth == 1 && "assume operand is not i1?");
549       KnownZero.clearAllBits();
550       KnownOne.setAllBits();
551       return;
552     }
553 
554     // The remaining tests are all recursive, so bail out if we hit the limit.
555     if (Depth == MaxDepth)
556       continue;
557 
558     Value *A, *B;
559     auto m_V = m_CombineOr(m_Specific(V),
560                            m_CombineOr(m_PtrToInt(m_Specific(V)),
561                            m_BitCast(m_Specific(V))));
562 
563     CmpInst::Predicate Pred;
564     ConstantInt *C;
565     // assume(v = a)
566     if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
567         Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
568       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
569       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
570       KnownZero |= RHSKnownZero;
571       KnownOne  |= RHSKnownOne;
572     // assume(v & b = a)
573     } else if (match(Arg,
574                      m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
575                Pred == ICmpInst::ICMP_EQ &&
576                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
577       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
578       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
579       APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
580       computeKnownBits(B, MaskKnownZero, MaskKnownOne, Depth+1, Query(Q, I));
581 
582       // For those bits in the mask that are known to be one, we can propagate
583       // known bits from the RHS to V.
584       KnownZero |= RHSKnownZero & MaskKnownOne;
585       KnownOne  |= RHSKnownOne  & MaskKnownOne;
586     // assume(~(v & b) = a)
587     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
588                                    m_Value(A))) &&
589                Pred == ICmpInst::ICMP_EQ &&
590                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
591       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
592       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
593       APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
594       computeKnownBits(B, MaskKnownZero, MaskKnownOne, Depth+1, Query(Q, I));
595 
596       // For those bits in the mask that are known to be one, we can propagate
597       // inverted known bits from the RHS to V.
598       KnownZero |= RHSKnownOne  & MaskKnownOne;
599       KnownOne  |= RHSKnownZero & MaskKnownOne;
600     // assume(v | b = a)
601     } else if (match(Arg,
602                      m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
603                Pred == ICmpInst::ICMP_EQ &&
604                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
605       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
606       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
607       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
608       computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I));
609 
610       // For those bits in B that are known to be zero, we can propagate known
611       // bits from the RHS to V.
612       KnownZero |= RHSKnownZero & BKnownZero;
613       KnownOne  |= RHSKnownOne  & BKnownZero;
614     // assume(~(v | b) = a)
615     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
616                                    m_Value(A))) &&
617                Pred == ICmpInst::ICMP_EQ &&
618                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
619       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
620       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
621       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
622       computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I));
623 
624       // For those bits in B that are known to be zero, we can propagate
625       // inverted known bits from the RHS to V.
626       KnownZero |= RHSKnownOne  & BKnownZero;
627       KnownOne  |= RHSKnownZero & BKnownZero;
628     // assume(v ^ b = a)
629     } else if (match(Arg,
630                      m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
631                Pred == ICmpInst::ICMP_EQ &&
632                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
633       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
634       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
635       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
636       computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I));
637 
638       // For those bits in B that are known to be zero, we can propagate known
639       // bits from the RHS to V. For those bits in B that are known to be one,
640       // we can propagate inverted known bits from the RHS to V.
641       KnownZero |= RHSKnownZero & BKnownZero;
642       KnownOne  |= RHSKnownOne  & BKnownZero;
643       KnownZero |= RHSKnownOne  & BKnownOne;
644       KnownOne  |= RHSKnownZero & BKnownOne;
645     // assume(~(v ^ b) = a)
646     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
647                                    m_Value(A))) &&
648                Pred == ICmpInst::ICMP_EQ &&
649                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
650       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
651       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
652       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
653       computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I));
654 
655       // For those bits in B that are known to be zero, we can propagate
656       // inverted known bits from the RHS to V. For those bits in B that are
657       // known to be one, we can propagate known bits from the RHS to V.
658       KnownZero |= RHSKnownOne  & BKnownZero;
659       KnownOne  |= RHSKnownZero & BKnownZero;
660       KnownZero |= RHSKnownZero & BKnownOne;
661       KnownOne  |= RHSKnownOne  & BKnownOne;
662     // assume(v << c = a)
663     } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
664                                    m_Value(A))) &&
665                Pred == ICmpInst::ICMP_EQ &&
666                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
667       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
668       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
669       // For those bits in RHS that are known, we can propagate them to known
670       // bits in V shifted to the right by C.
671       KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
672       KnownOne  |= RHSKnownOne.lshr(C->getZExtValue());
673     // assume(~(v << c) = a)
674     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
675                                    m_Value(A))) &&
676                Pred == ICmpInst::ICMP_EQ &&
677                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
678       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
679       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
680       // For those bits in RHS that are known, we can propagate them inverted
681       // to known bits in V shifted to the right by C.
682       KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
683       KnownOne  |= RHSKnownZero.lshr(C->getZExtValue());
684     // assume(v >> c = a)
685     } else if (match(Arg,
686                      m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
687                                                 m_AShr(m_V, m_ConstantInt(C))),
688                               m_Value(A))) &&
689                Pred == ICmpInst::ICMP_EQ &&
690                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
691       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
692       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
693       // For those bits in RHS that are known, we can propagate them to known
694       // bits in V shifted to the right by C.
695       KnownZero |= RHSKnownZero << C->getZExtValue();
696       KnownOne  |= RHSKnownOne  << C->getZExtValue();
697     // assume(~(v >> c) = a)
698     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
699                                              m_LShr(m_V, m_ConstantInt(C)),
700                                              m_AShr(m_V, m_ConstantInt(C)))),
701                                    m_Value(A))) &&
702                Pred == ICmpInst::ICMP_EQ &&
703                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
704       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
705       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
706       // For those bits in RHS that are known, we can propagate them inverted
707       // to known bits in V shifted to the right by C.
708       KnownZero |= RHSKnownOne  << C->getZExtValue();
709       KnownOne  |= RHSKnownZero << C->getZExtValue();
710     // assume(v >=_s c) where c is non-negative
711     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
712                Pred == ICmpInst::ICMP_SGE &&
713                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
714       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
715       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
716 
717       if (RHSKnownZero.isNegative()) {
718         // We know that the sign bit is zero.
719         KnownZero |= APInt::getSignBit(BitWidth);
720       }
721     // assume(v >_s c) where c is at least -1.
722     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
723                Pred == ICmpInst::ICMP_SGT &&
724                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
725       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
726       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
727 
728       if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
729         // We know that the sign bit is zero.
730         KnownZero |= APInt::getSignBit(BitWidth);
731       }
732     // assume(v <=_s c) where c is negative
733     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
734                Pred == ICmpInst::ICMP_SLE &&
735                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
736       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
737       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
738 
739       if (RHSKnownOne.isNegative()) {
740         // We know that the sign bit is one.
741         KnownOne |= APInt::getSignBit(BitWidth);
742       }
743     // assume(v <_s c) where c is non-positive
744     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
745                Pred == ICmpInst::ICMP_SLT &&
746                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
747       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
748       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
749 
750       if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
751         // We know that the sign bit is one.
752         KnownOne |= APInt::getSignBit(BitWidth);
753       }
754     // assume(v <=_u c)
755     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
756                Pred == ICmpInst::ICMP_ULE &&
757                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
758       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
759       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
760 
761       // Whatever high bits in c are zero are known to be zero.
762       KnownZero |=
763         APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
764     // assume(v <_u c)
765     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
766                Pred == ICmpInst::ICMP_ULT &&
767                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
768       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
769       computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I));
770 
771       // Whatever high bits in c are zero are known to be zero (if c is a power
772       // of 2, then one more).
773       if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
774         KnownZero |=
775           APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
776       else
777         KnownZero |=
778           APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
779     }
780   }
781 }
782 
783 // Compute known bits from a shift operator, including those with a
784 // non-constant shift amount. KnownZero and KnownOne are the outputs of this
785 // function. KnownZero2 and KnownOne2 are pre-allocated temporaries with the
786 // same bit width as KnownZero and KnownOne. KZF and KOF are operator-specific
787 // functors that, given the known-zero or known-one bits respectively, and a
788 // shift amount, compute the implied known-zero or known-one bits of the shift
789 // operator's result respectively for that shift amount. The results from calling
790 // KZF and KOF are conservatively combined for all permitted shift amounts.
791 template <typename KZFunctor, typename KOFunctor>
computeKnownBitsFromShiftOperator(Operator * I,APInt & KnownZero,APInt & KnownOne,APInt & KnownZero2,APInt & KnownOne2,unsigned Depth,const Query & Q,KZFunctor KZF,KOFunctor KOF)792 static void computeKnownBitsFromShiftOperator(Operator *I,
793               APInt &KnownZero, APInt &KnownOne,
794               APInt &KnownZero2, APInt &KnownOne2,
795               unsigned Depth, const Query &Q, KZFunctor KZF, KOFunctor KOF) {
796   unsigned BitWidth = KnownZero.getBitWidth();
797 
798   if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
799     unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
800 
801     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
802     KnownZero = KZF(KnownZero, ShiftAmt);
803     KnownOne  = KOF(KnownOne, ShiftAmt);
804     return;
805   }
806 
807   computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q);
808 
809   // Note: We cannot use KnownZero.getLimitedValue() here, because if
810   // BitWidth > 64 and any upper bits are known, we'll end up returning the
811   // limit value (which implies all bits are known).
812   uint64_t ShiftAmtKZ = KnownZero.zextOrTrunc(64).getZExtValue();
813   uint64_t ShiftAmtKO = KnownOne.zextOrTrunc(64).getZExtValue();
814 
815   // It would be more-clearly correct to use the two temporaries for this
816   // calculation. Reusing the APInts here to prevent unnecessary allocations.
817   KnownZero.clearAllBits();
818   KnownOne.clearAllBits();
819 
820   // If we know the shifter operand is nonzero, we can sometimes infer more
821   // known bits. However this is expensive to compute, so be lazy about it and
822   // only compute it when absolutely necessary.
823   Optional<bool> ShifterOperandIsNonZero;
824 
825   // Early exit if we can't constrain any well-defined shift amount.
826   if (!(ShiftAmtKZ & (BitWidth - 1)) && !(ShiftAmtKO & (BitWidth - 1))) {
827     ShifterOperandIsNonZero =
828         isKnownNonZero(I->getOperand(1), Depth + 1, Q);
829     if (!*ShifterOperandIsNonZero)
830       return;
831   }
832 
833   computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q);
834 
835   KnownZero = KnownOne = APInt::getAllOnesValue(BitWidth);
836   for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
837     // Combine the shifted known input bits only for those shift amounts
838     // compatible with its known constraints.
839     if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
840       continue;
841     if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
842       continue;
843     // If we know the shifter is nonzero, we may be able to infer more known
844     // bits. This check is sunk down as far as possible to avoid the expensive
845     // call to isKnownNonZero if the cheaper checks above fail.
846     if (ShiftAmt == 0) {
847       if (!ShifterOperandIsNonZero.hasValue())
848         ShifterOperandIsNonZero =
849             isKnownNonZero(I->getOperand(1), Depth + 1, Q);
850       if (*ShifterOperandIsNonZero)
851         continue;
852     }
853 
854     KnownZero &= KZF(KnownZero2, ShiftAmt);
855     KnownOne  &= KOF(KnownOne2, ShiftAmt);
856   }
857 
858   // If there are no compatible shift amounts, then we've proven that the shift
859   // amount must be >= the BitWidth, and the result is undefined. We could
860   // return anything we'd like, but we need to make sure the sets of known bits
861   // stay disjoint (it should be better for some other code to actually
862   // propagate the undef than to pick a value here using known bits).
863   if ((KnownZero & KnownOne) != 0) {
864     KnownZero.clearAllBits();
865     KnownOne.clearAllBits();
866   }
867 }
868 
computeKnownBitsFromOperator(Operator * I,APInt & KnownZero,APInt & KnownOne,unsigned Depth,const Query & Q)869 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
870                                          APInt &KnownOne, unsigned Depth,
871                                          const Query &Q) {
872   unsigned BitWidth = KnownZero.getBitWidth();
873 
874   APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
875   switch (I->getOpcode()) {
876   default: break;
877   case Instruction::Load:
878     if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
879       computeKnownBitsFromRangeMetadata(*MD, KnownZero, KnownOne);
880     break;
881   case Instruction::And: {
882     // If either the LHS or the RHS are Zero, the result is zero.
883     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q);
884     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q);
885 
886     // Output known-1 bits are only known if set in both the LHS & RHS.
887     KnownOne &= KnownOne2;
888     // Output known-0 are known to be clear if zero in either the LHS | RHS.
889     KnownZero |= KnownZero2;
890 
891     // and(x, add (x, -1)) is a common idiom that always clears the low bit;
892     // here we handle the more general case of adding any odd number by
893     // matching the form add(x, add(x, y)) where y is odd.
894     // TODO: This could be generalized to clearing any bit set in y where the
895     // following bit is known to be unset in y.
896     Value *Y = nullptr;
897     if (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)),
898                                       m_Value(Y))) ||
899         match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)),
900                                       m_Value(Y)))) {
901       APInt KnownZero3(BitWidth, 0), KnownOne3(BitWidth, 0);
902       computeKnownBits(Y, KnownZero3, KnownOne3, Depth + 1, Q);
903       if (KnownOne3.countTrailingOnes() > 0)
904         KnownZero |= APInt::getLowBitsSet(BitWidth, 1);
905     }
906     break;
907   }
908   case Instruction::Or: {
909     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q);
910     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q);
911 
912     // Output known-0 bits are only known if clear in both the LHS & RHS.
913     KnownZero &= KnownZero2;
914     // Output known-1 are known to be set if set in either the LHS | RHS.
915     KnownOne |= KnownOne2;
916     break;
917   }
918   case Instruction::Xor: {
919     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q);
920     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q);
921 
922     // Output known-0 bits are known if clear or set in both the LHS & RHS.
923     APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
924     // Output known-1 are known to be set if set in only one of the LHS, RHS.
925     KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
926     KnownZero = KnownZeroOut;
927     break;
928   }
929   case Instruction::Mul: {
930     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
931     computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
932                         KnownOne, KnownZero2, KnownOne2, Depth, Q);
933     break;
934   }
935   case Instruction::UDiv: {
936     // For the purposes of computing leading zeros we can conservatively
937     // treat a udiv as a logical right shift by the power of 2 known to
938     // be less than the denominator.
939     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q);
940     unsigned LeadZ = KnownZero2.countLeadingOnes();
941 
942     KnownOne2.clearAllBits();
943     KnownZero2.clearAllBits();
944     computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, Depth + 1, Q);
945     unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
946     if (RHSUnknownLeadingOnes != BitWidth)
947       LeadZ = std::min(BitWidth,
948                        LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
949 
950     KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
951     break;
952   }
953   case Instruction::Select:
954     computeKnownBits(I->getOperand(2), KnownZero, KnownOne, Depth + 1, Q);
955     computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, Depth + 1, Q);
956 
957     // Only known if known in both the LHS and RHS.
958     KnownOne &= KnownOne2;
959     KnownZero &= KnownZero2;
960     break;
961   case Instruction::FPTrunc:
962   case Instruction::FPExt:
963   case Instruction::FPToUI:
964   case Instruction::FPToSI:
965   case Instruction::SIToFP:
966   case Instruction::UIToFP:
967     break; // Can't work with floating point.
968   case Instruction::PtrToInt:
969   case Instruction::IntToPtr:
970   case Instruction::AddrSpaceCast: // Pointers could be different sizes.
971     // FALL THROUGH and handle them the same as zext/trunc.
972   case Instruction::ZExt:
973   case Instruction::Trunc: {
974     Type *SrcTy = I->getOperand(0)->getType();
975 
976     unsigned SrcBitWidth;
977     // Note that we handle pointer operands here because of inttoptr/ptrtoint
978     // which fall through here.
979     SrcBitWidth = Q.DL.getTypeSizeInBits(SrcTy->getScalarType());
980 
981     assert(SrcBitWidth && "SrcBitWidth can't be zero");
982     KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
983     KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
984     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
985     KnownZero = KnownZero.zextOrTrunc(BitWidth);
986     KnownOne = KnownOne.zextOrTrunc(BitWidth);
987     // Any top bits are known to be zero.
988     if (BitWidth > SrcBitWidth)
989       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
990     break;
991   }
992   case Instruction::BitCast: {
993     Type *SrcTy = I->getOperand(0)->getType();
994     if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
995         // TODO: For now, not handling conversions like:
996         // (bitcast i64 %x to <2 x i32>)
997         !I->getType()->isVectorTy()) {
998       computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
999       break;
1000     }
1001     break;
1002   }
1003   case Instruction::SExt: {
1004     // Compute the bits in the result that are not present in the input.
1005     unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1006 
1007     KnownZero = KnownZero.trunc(SrcBitWidth);
1008     KnownOne = KnownOne.trunc(SrcBitWidth);
1009     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
1010     KnownZero = KnownZero.zext(BitWidth);
1011     KnownOne = KnownOne.zext(BitWidth);
1012 
1013     // If the sign bit of the input is known set or clear, then we know the
1014     // top bits of the result.
1015     if (KnownZero[SrcBitWidth-1])             // Input sign bit known zero
1016       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1017     else if (KnownOne[SrcBitWidth-1])           // Input sign bit known set
1018       KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1019     break;
1020   }
1021   case Instruction::Shl: {
1022     // (shl X, C1) & C2 == 0   iff   (X & C2 >>u C1) == 0
1023     auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1024       return (KnownZero << ShiftAmt) |
1025              APInt::getLowBitsSet(BitWidth, ShiftAmt); // Low bits known 0.
1026     };
1027 
1028     auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1029       return KnownOne << ShiftAmt;
1030     };
1031 
1032     computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1033                                       KnownZero2, KnownOne2, Depth, Q, KZF,
1034                                       KOF);
1035     break;
1036   }
1037   case Instruction::LShr: {
1038     // (ushr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
1039     auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1040       return APIntOps::lshr(KnownZero, ShiftAmt) |
1041              // High bits known zero.
1042              APInt::getHighBitsSet(BitWidth, ShiftAmt);
1043     };
1044 
1045     auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1046       return APIntOps::lshr(KnownOne, ShiftAmt);
1047     };
1048 
1049     computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1050                                       KnownZero2, KnownOne2, Depth, Q, KZF,
1051                                       KOF);
1052     break;
1053   }
1054   case Instruction::AShr: {
1055     // (ashr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
1056     auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1057       return APIntOps::ashr(KnownZero, ShiftAmt);
1058     };
1059 
1060     auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1061       return APIntOps::ashr(KnownOne, ShiftAmt);
1062     };
1063 
1064     computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1065                                       KnownZero2, KnownOne2, Depth, Q, KZF,
1066                                       KOF);
1067     break;
1068   }
1069   case Instruction::Sub: {
1070     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1071     computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1072                            KnownZero, KnownOne, KnownZero2, KnownOne2, Depth,
1073                            Q);
1074     break;
1075   }
1076   case Instruction::Add: {
1077     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1078     computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1079                            KnownZero, KnownOne, KnownZero2, KnownOne2, Depth,
1080                            Q);
1081     break;
1082   }
1083   case Instruction::SRem:
1084     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1085       APInt RA = Rem->getValue().abs();
1086       if (RA.isPowerOf2()) {
1087         APInt LowBits = RA - 1;
1088         computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1,
1089                          Q);
1090 
1091         // The low bits of the first operand are unchanged by the srem.
1092         KnownZero = KnownZero2 & LowBits;
1093         KnownOne = KnownOne2 & LowBits;
1094 
1095         // If the first operand is non-negative or has all low bits zero, then
1096         // the upper bits are all zero.
1097         if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1098           KnownZero |= ~LowBits;
1099 
1100         // If the first operand is negative and not all low bits are zero, then
1101         // the upper bits are all one.
1102         if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1103           KnownOne |= ~LowBits;
1104 
1105         assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1106       }
1107     }
1108 
1109     // The sign bit is the LHS's sign bit, except when the result of the
1110     // remainder is zero.
1111     if (KnownZero.isNonNegative()) {
1112       APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1113       computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, Depth + 1,
1114                        Q);
1115       // If it's known zero, our sign bit is also zero.
1116       if (LHSKnownZero.isNegative())
1117         KnownZero.setBit(BitWidth - 1);
1118     }
1119 
1120     break;
1121   case Instruction::URem: {
1122     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1123       const APInt &RA = Rem->getValue();
1124       if (RA.isPowerOf2()) {
1125         APInt LowBits = (RA - 1);
1126         computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
1127         KnownZero |= ~LowBits;
1128         KnownOne &= LowBits;
1129         break;
1130       }
1131     }
1132 
1133     // Since the result is less than or equal to either operand, any leading
1134     // zero bits in either operand must also exist in the result.
1135     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
1136     computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, Depth + 1, Q);
1137 
1138     unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1139                                 KnownZero2.countLeadingOnes());
1140     KnownOne.clearAllBits();
1141     KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1142     break;
1143   }
1144 
1145   case Instruction::Alloca: {
1146     AllocaInst *AI = cast<AllocaInst>(I);
1147     unsigned Align = AI->getAlignment();
1148     if (Align == 0)
1149       Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1150 
1151     if (Align > 0)
1152       KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1153     break;
1154   }
1155   case Instruction::GetElementPtr: {
1156     // Analyze all of the subscripts of this getelementptr instruction
1157     // to determine if we can prove known low zero bits.
1158     APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1159     computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, Depth + 1,
1160                      Q);
1161     unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1162 
1163     gep_type_iterator GTI = gep_type_begin(I);
1164     for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1165       Value *Index = I->getOperand(i);
1166       if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1167         // Handle struct member offset arithmetic.
1168 
1169         // Handle case when index is vector zeroinitializer
1170         Constant *CIndex = cast<Constant>(Index);
1171         if (CIndex->isZeroValue())
1172           continue;
1173 
1174         if (CIndex->getType()->isVectorTy())
1175           Index = CIndex->getSplatValue();
1176 
1177         unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1178         const StructLayout *SL = Q.DL.getStructLayout(STy);
1179         uint64_t Offset = SL->getElementOffset(Idx);
1180         TrailZ = std::min<unsigned>(TrailZ,
1181                                     countTrailingZeros(Offset));
1182       } else {
1183         // Handle array index arithmetic.
1184         Type *IndexedTy = GTI.getIndexedType();
1185         if (!IndexedTy->isSized()) {
1186           TrailZ = 0;
1187           break;
1188         }
1189         unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1190         uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1191         LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1192         computeKnownBits(Index, LocalKnownZero, LocalKnownOne, Depth + 1, Q);
1193         TrailZ = std::min(TrailZ,
1194                           unsigned(countTrailingZeros(TypeSize) +
1195                                    LocalKnownZero.countTrailingOnes()));
1196       }
1197     }
1198 
1199     KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1200     break;
1201   }
1202   case Instruction::PHI: {
1203     PHINode *P = cast<PHINode>(I);
1204     // Handle the case of a simple two-predecessor recurrence PHI.
1205     // There's a lot more that could theoretically be done here, but
1206     // this is sufficient to catch some interesting cases.
1207     if (P->getNumIncomingValues() == 2) {
1208       for (unsigned i = 0; i != 2; ++i) {
1209         Value *L = P->getIncomingValue(i);
1210         Value *R = P->getIncomingValue(!i);
1211         Operator *LU = dyn_cast<Operator>(L);
1212         if (!LU)
1213           continue;
1214         unsigned Opcode = LU->getOpcode();
1215         // Check for operations that have the property that if
1216         // both their operands have low zero bits, the result
1217         // will have low zero bits.
1218         if (Opcode == Instruction::Add ||
1219             Opcode == Instruction::Sub ||
1220             Opcode == Instruction::And ||
1221             Opcode == Instruction::Or ||
1222             Opcode == Instruction::Mul) {
1223           Value *LL = LU->getOperand(0);
1224           Value *LR = LU->getOperand(1);
1225           // Find a recurrence.
1226           if (LL == I)
1227             L = LR;
1228           else if (LR == I)
1229             L = LL;
1230           else
1231             break;
1232           // Ok, we have a PHI of the form L op= R. Check for low
1233           // zero bits.
1234           computeKnownBits(R, KnownZero2, KnownOne2, Depth + 1, Q);
1235 
1236           // We need to take the minimum number of known bits
1237           APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1238           computeKnownBits(L, KnownZero3, KnownOne3, Depth + 1, Q);
1239 
1240           KnownZero = APInt::getLowBitsSet(BitWidth,
1241                                            std::min(KnownZero2.countTrailingOnes(),
1242                                                     KnownZero3.countTrailingOnes()));
1243           break;
1244         }
1245       }
1246     }
1247 
1248     // Unreachable blocks may have zero-operand PHI nodes.
1249     if (P->getNumIncomingValues() == 0)
1250       break;
1251 
1252     // Otherwise take the unions of the known bit sets of the operands,
1253     // taking conservative care to avoid excessive recursion.
1254     if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1255       // Skip if every incoming value references to ourself.
1256       if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1257         break;
1258 
1259       KnownZero = APInt::getAllOnesValue(BitWidth);
1260       KnownOne = APInt::getAllOnesValue(BitWidth);
1261       for (Value *IncValue : P->incoming_values()) {
1262         // Skip direct self references.
1263         if (IncValue == P) continue;
1264 
1265         KnownZero2 = APInt(BitWidth, 0);
1266         KnownOne2 = APInt(BitWidth, 0);
1267         // Recurse, but cap the recursion to one level, because we don't
1268         // want to waste time spinning around in loops.
1269         computeKnownBits(IncValue, KnownZero2, KnownOne2, MaxDepth - 1, Q);
1270         KnownZero &= KnownZero2;
1271         KnownOne &= KnownOne2;
1272         // If all bits have been ruled out, there's no need to check
1273         // more operands.
1274         if (!KnownZero && !KnownOne)
1275           break;
1276       }
1277     }
1278     break;
1279   }
1280   case Instruction::Call:
1281   case Instruction::Invoke:
1282     // If range metadata is attached to this call, set known bits from that,
1283     // and then intersect with known bits based on other properties of the
1284     // function.
1285     if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1286       computeKnownBitsFromRangeMetadata(*MD, KnownZero, KnownOne);
1287     if (Value *RV = CallSite(I).getReturnedArgOperand()) {
1288       computeKnownBits(RV, KnownZero2, KnownOne2, Depth + 1, Q);
1289       KnownZero |= KnownZero2;
1290       KnownOne |= KnownOne2;
1291     }
1292     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1293       switch (II->getIntrinsicID()) {
1294       default: break;
1295       case Intrinsic::bswap:
1296         computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q);
1297         KnownZero |= KnownZero2.byteSwap();
1298         KnownOne |= KnownOne2.byteSwap();
1299         break;
1300       case Intrinsic::ctlz:
1301       case Intrinsic::cttz: {
1302         unsigned LowBits = Log2_32(BitWidth)+1;
1303         // If this call is undefined for 0, the result will be less than 2^n.
1304         if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1305           LowBits -= 1;
1306         KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1307         break;
1308       }
1309       case Intrinsic::ctpop: {
1310         computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q);
1311         // We can bound the space the count needs.  Also, bits known to be zero
1312         // can't contribute to the population.
1313         unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation();
1314         unsigned LeadingZeros =
1315           APInt(BitWidth, BitsPossiblySet).countLeadingZeros();
1316         assert(LeadingZeros <= BitWidth);
1317         KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros);
1318         KnownOne &= ~KnownZero;
1319         // TODO: we could bound KnownOne using the lower bound on the number
1320         // of bits which might be set provided by popcnt KnownOne2.
1321         break;
1322       }
1323       case Intrinsic::x86_sse42_crc32_64_64:
1324         KnownZero |= APInt::getHighBitsSet(64, 32);
1325         break;
1326       }
1327     }
1328     break;
1329   case Instruction::ExtractValue:
1330     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1331       ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1332       if (EVI->getNumIndices() != 1) break;
1333       if (EVI->getIndices()[0] == 0) {
1334         switch (II->getIntrinsicID()) {
1335         default: break;
1336         case Intrinsic::uadd_with_overflow:
1337         case Intrinsic::sadd_with_overflow:
1338           computeKnownBitsAddSub(true, II->getArgOperand(0),
1339                                  II->getArgOperand(1), false, KnownZero,
1340                                  KnownOne, KnownZero2, KnownOne2, Depth, Q);
1341           break;
1342         case Intrinsic::usub_with_overflow:
1343         case Intrinsic::ssub_with_overflow:
1344           computeKnownBitsAddSub(false, II->getArgOperand(0),
1345                                  II->getArgOperand(1), false, KnownZero,
1346                                  KnownOne, KnownZero2, KnownOne2, Depth, Q);
1347           break;
1348         case Intrinsic::umul_with_overflow:
1349         case Intrinsic::smul_with_overflow:
1350           computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1351                               KnownZero, KnownOne, KnownZero2, KnownOne2, Depth,
1352                               Q);
1353           break;
1354         }
1355       }
1356     }
1357   }
1358 }
1359 
1360 /// Determine which bits of V are known to be either zero or one and return
1361 /// them in the KnownZero/KnownOne bit sets.
1362 ///
1363 /// NOTE: we cannot consider 'undef' to be "IsZero" here.  The problem is that
1364 /// we cannot optimize based on the assumption that it is zero without changing
1365 /// it to be an explicit zero.  If we don't change it to zero, other code could
1366 /// optimized based on the contradictory assumption that it is non-zero.
1367 /// Because instcombine aggressively folds operations with undef args anyway,
1368 /// this won't lose us code quality.
1369 ///
1370 /// This function is defined on values with integer type, values with pointer
1371 /// type, and vectors of integers.  In the case
1372 /// where V is a vector, known zero, and known one values are the
1373 /// same width as the vector element, and the bit is set only if it is true
1374 /// for all of the elements in the vector.
computeKnownBits(Value * V,APInt & KnownZero,APInt & KnownOne,unsigned Depth,const Query & Q)1375 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1376                       unsigned Depth, const Query &Q) {
1377   assert(V && "No Value?");
1378   assert(Depth <= MaxDepth && "Limit Search Depth");
1379   unsigned BitWidth = KnownZero.getBitWidth();
1380 
1381   assert((V->getType()->isIntOrIntVectorTy() ||
1382           V->getType()->getScalarType()->isPointerTy()) &&
1383          "Not integer or pointer type!");
1384   assert((Q.DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1385          (!V->getType()->isIntOrIntVectorTy() ||
1386           V->getType()->getScalarSizeInBits() == BitWidth) &&
1387          KnownZero.getBitWidth() == BitWidth &&
1388          KnownOne.getBitWidth() == BitWidth &&
1389          "V, KnownOne and KnownZero should have same BitWidth");
1390 
1391   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1392     // We know all of the bits for a constant!
1393     KnownOne = CI->getValue();
1394     KnownZero = ~KnownOne;
1395     return;
1396   }
1397   // Null and aggregate-zero are all-zeros.
1398   if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1399     KnownOne.clearAllBits();
1400     KnownZero = APInt::getAllOnesValue(BitWidth);
1401     return;
1402   }
1403   // Handle a constant vector by taking the intersection of the known bits of
1404   // each element.
1405   if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1406     // We know that CDS must be a vector of integers. Take the intersection of
1407     // each element.
1408     KnownZero.setAllBits(); KnownOne.setAllBits();
1409     APInt Elt(KnownZero.getBitWidth(), 0);
1410     for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1411       Elt = CDS->getElementAsInteger(i);
1412       KnownZero &= ~Elt;
1413       KnownOne &= Elt;
1414     }
1415     return;
1416   }
1417 
1418   if (auto *CV = dyn_cast<ConstantVector>(V)) {
1419     // We know that CV must be a vector of integers. Take the intersection of
1420     // each element.
1421     KnownZero.setAllBits(); KnownOne.setAllBits();
1422     APInt Elt(KnownZero.getBitWidth(), 0);
1423     for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1424       Constant *Element = CV->getAggregateElement(i);
1425       auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1426       if (!ElementCI) {
1427         KnownZero.clearAllBits();
1428         KnownOne.clearAllBits();
1429         return;
1430       }
1431       Elt = ElementCI->getValue();
1432       KnownZero &= ~Elt;
1433       KnownOne &= Elt;
1434     }
1435     return;
1436   }
1437 
1438   // Start out not knowing anything.
1439   KnownZero.clearAllBits(); KnownOne.clearAllBits();
1440 
1441   // Limit search depth.
1442   // All recursive calls that increase depth must come after this.
1443   if (Depth == MaxDepth)
1444     return;
1445 
1446   // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1447   // the bits of its aliasee.
1448   if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1449     if (!GA->isInterposable())
1450       computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, Depth + 1, Q);
1451     return;
1452   }
1453 
1454   if (Operator *I = dyn_cast<Operator>(V))
1455     computeKnownBitsFromOperator(I, KnownZero, KnownOne, Depth, Q);
1456 
1457   // Aligned pointers have trailing zeros - refine KnownZero set
1458   if (V->getType()->isPointerTy()) {
1459     unsigned Align = V->getPointerAlignment(Q.DL);
1460     if (Align)
1461       KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1462   }
1463 
1464   // computeKnownBitsFromAssume strictly refines KnownZero and
1465   // KnownOne. Therefore, we run them after computeKnownBitsFromOperator.
1466 
1467   // Check whether a nearby assume intrinsic can determine some known bits.
1468   computeKnownBitsFromAssume(V, KnownZero, KnownOne, Depth, Q);
1469 
1470   assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1471 }
1472 
1473 /// Determine whether the sign bit is known to be zero or one.
1474 /// Convenience wrapper around computeKnownBits.
ComputeSignBit(Value * V,bool & KnownZero,bool & KnownOne,unsigned Depth,const Query & Q)1475 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1476                     unsigned Depth, const Query &Q) {
1477   unsigned BitWidth = getBitWidth(V->getType(), Q.DL);
1478   if (!BitWidth) {
1479     KnownZero = false;
1480     KnownOne = false;
1481     return;
1482   }
1483   APInt ZeroBits(BitWidth, 0);
1484   APInt OneBits(BitWidth, 0);
1485   computeKnownBits(V, ZeroBits, OneBits, Depth, Q);
1486   KnownOne = OneBits[BitWidth - 1];
1487   KnownZero = ZeroBits[BitWidth - 1];
1488 }
1489 
1490 /// Return true if the given value is known to have exactly one
1491 /// bit set when defined. For vectors return true if every element is known to
1492 /// be a power of two when defined. Supports values with integer or pointer
1493 /// types and vectors of integers.
isKnownToBeAPowerOfTwo(Value * V,bool OrZero,unsigned Depth,const Query & Q)1494 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1495                             const Query &Q) {
1496   if (Constant *C = dyn_cast<Constant>(V)) {
1497     if (C->isNullValue())
1498       return OrZero;
1499 
1500     const APInt *ConstIntOrConstSplatInt;
1501     if (match(C, m_APInt(ConstIntOrConstSplatInt)))
1502       return ConstIntOrConstSplatInt->isPowerOf2();
1503   }
1504 
1505   // 1 << X is clearly a power of two if the one is not shifted off the end.  If
1506   // it is shifted off the end then the result is undefined.
1507   if (match(V, m_Shl(m_One(), m_Value())))
1508     return true;
1509 
1510   // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1511   // bottom.  If it is shifted off the bottom then the result is undefined.
1512   if (match(V, m_LShr(m_SignBit(), m_Value())))
1513     return true;
1514 
1515   // The remaining tests are all recursive, so bail out if we hit the limit.
1516   if (Depth++ == MaxDepth)
1517     return false;
1518 
1519   Value *X = nullptr, *Y = nullptr;
1520   // A shift left or a logical shift right of a power of two is a power of two
1521   // or zero.
1522   if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1523                  match(V, m_LShr(m_Value(X), m_Value()))))
1524     return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1525 
1526   if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1527     return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1528 
1529   if (SelectInst *SI = dyn_cast<SelectInst>(V))
1530     return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1531            isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1532 
1533   if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1534     // A power of two and'd with anything is a power of two or zero.
1535     if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1536         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1537       return true;
1538     // X & (-X) is always a power of two or zero.
1539     if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1540       return true;
1541     return false;
1542   }
1543 
1544   // Adding a power-of-two or zero to the same power-of-two or zero yields
1545   // either the original power-of-two, a larger power-of-two or zero.
1546   if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1547     OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1548     if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1549       if (match(X, m_And(m_Specific(Y), m_Value())) ||
1550           match(X, m_And(m_Value(), m_Specific(Y))))
1551         if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1552           return true;
1553       if (match(Y, m_And(m_Specific(X), m_Value())) ||
1554           match(Y, m_And(m_Value(), m_Specific(X))))
1555         if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1556           return true;
1557 
1558       unsigned BitWidth = V->getType()->getScalarSizeInBits();
1559       APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1560       computeKnownBits(X, LHSZeroBits, LHSOneBits, Depth, Q);
1561 
1562       APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1563       computeKnownBits(Y, RHSZeroBits, RHSOneBits, Depth, Q);
1564       // If i8 V is a power of two or zero:
1565       //  ZeroBits: 1 1 1 0 1 1 1 1
1566       // ~ZeroBits: 0 0 0 1 0 0 0 0
1567       if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1568         // If OrZero isn't set, we cannot give back a zero result.
1569         // Make sure either the LHS or RHS has a bit set.
1570         if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1571           return true;
1572     }
1573   }
1574 
1575   // An exact divide or right shift can only shift off zero bits, so the result
1576   // is a power of two only if the first operand is a power of two and not
1577   // copying a sign bit (sdiv int_min, 2).
1578   if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1579       match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1580     return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1581                                   Depth, Q);
1582   }
1583 
1584   return false;
1585 }
1586 
1587 /// \brief Test whether a GEP's result is known to be non-null.
1588 ///
1589 /// Uses properties inherent in a GEP to try to determine whether it is known
1590 /// to be non-null.
1591 ///
1592 /// Currently this routine does not support vector GEPs.
isGEPKnownNonNull(GEPOperator * GEP,unsigned Depth,const Query & Q)1593 static bool isGEPKnownNonNull(GEPOperator *GEP, unsigned Depth,
1594                               const Query &Q) {
1595   if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1596     return false;
1597 
1598   // FIXME: Support vector-GEPs.
1599   assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1600 
1601   // If the base pointer is non-null, we cannot walk to a null address with an
1602   // inbounds GEP in address space zero.
1603   if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1604     return true;
1605 
1606   // Walk the GEP operands and see if any operand introduces a non-zero offset.
1607   // If so, then the GEP cannot produce a null pointer, as doing so would
1608   // inherently violate the inbounds contract within address space zero.
1609   for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1610        GTI != GTE; ++GTI) {
1611     // Struct types are easy -- they must always be indexed by a constant.
1612     if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1613       ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1614       unsigned ElementIdx = OpC->getZExtValue();
1615       const StructLayout *SL = Q.DL.getStructLayout(STy);
1616       uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1617       if (ElementOffset > 0)
1618         return true;
1619       continue;
1620     }
1621 
1622     // If we have a zero-sized type, the index doesn't matter. Keep looping.
1623     if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1624       continue;
1625 
1626     // Fast path the constant operand case both for efficiency and so we don't
1627     // increment Depth when just zipping down an all-constant GEP.
1628     if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1629       if (!OpC->isZero())
1630         return true;
1631       continue;
1632     }
1633 
1634     // We post-increment Depth here because while isKnownNonZero increments it
1635     // as well, when we pop back up that increment won't persist. We don't want
1636     // to recurse 10k times just because we have 10k GEP operands. We don't
1637     // bail completely out because we want to handle constant GEPs regardless
1638     // of depth.
1639     if (Depth++ >= MaxDepth)
1640       continue;
1641 
1642     if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1643       return true;
1644   }
1645 
1646   return false;
1647 }
1648 
1649 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1650 /// ensure that the value it's attached to is never Value?  'RangeType' is
1651 /// is the type of the value described by the range.
rangeMetadataExcludesValue(MDNode * Ranges,const APInt & Value)1652 static bool rangeMetadataExcludesValue(MDNode* Ranges, const APInt& Value) {
1653   const unsigned NumRanges = Ranges->getNumOperands() / 2;
1654   assert(NumRanges >= 1);
1655   for (unsigned i = 0; i < NumRanges; ++i) {
1656     ConstantInt *Lower =
1657         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1658     ConstantInt *Upper =
1659         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1660     ConstantRange Range(Lower->getValue(), Upper->getValue());
1661     if (Range.contains(Value))
1662       return false;
1663   }
1664   return true;
1665 }
1666 
1667 /// Return true if the given value is known to be non-zero when defined.
1668 /// For vectors return true if every element is known to be non-zero when
1669 /// defined. Supports values with integer or pointer type and vectors of
1670 /// integers.
isKnownNonZero(Value * V,unsigned Depth,const Query & Q)1671 bool isKnownNonZero(Value *V, unsigned Depth, const Query &Q) {
1672   if (auto *C = dyn_cast<Constant>(V)) {
1673     if (C->isNullValue())
1674       return false;
1675     if (isa<ConstantInt>(C))
1676       // Must be non-zero due to null test above.
1677       return true;
1678 
1679     // For constant vectors, check that all elements are undefined or known
1680     // non-zero to determine that the whole vector is known non-zero.
1681     if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1682       for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1683         Constant *Elt = C->getAggregateElement(i);
1684         if (!Elt || Elt->isNullValue())
1685           return false;
1686         if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1687           return false;
1688       }
1689       return true;
1690     }
1691 
1692     return false;
1693   }
1694 
1695   if (auto *I = dyn_cast<Instruction>(V)) {
1696     if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1697       // If the possible ranges don't contain zero, then the value is
1698       // definitely non-zero.
1699       if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1700         const APInt ZeroValue(Ty->getBitWidth(), 0);
1701         if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1702           return true;
1703       }
1704     }
1705   }
1706 
1707   // The remaining tests are all recursive, so bail out if we hit the limit.
1708   if (Depth++ >= MaxDepth)
1709     return false;
1710 
1711   // Check for pointer simplifications.
1712   if (V->getType()->isPointerTy()) {
1713     if (isKnownNonNull(V))
1714       return true;
1715     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1716       if (isGEPKnownNonNull(GEP, Depth, Q))
1717         return true;
1718   }
1719 
1720   unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
1721 
1722   // X | Y != 0 if X != 0 or Y != 0.
1723   Value *X = nullptr, *Y = nullptr;
1724   if (match(V, m_Or(m_Value(X), m_Value(Y))))
1725     return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
1726 
1727   // ext X != 0 if X != 0.
1728   if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1729     return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
1730 
1731   // shl X, Y != 0 if X is odd.  Note that the value of the shift is undefined
1732   // if the lowest bit is shifted off the end.
1733   if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1734     // shl nuw can't remove any non-zero bits.
1735     OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1736     if (BO->hasNoUnsignedWrap())
1737       return isKnownNonZero(X, Depth, Q);
1738 
1739     APInt KnownZero(BitWidth, 0);
1740     APInt KnownOne(BitWidth, 0);
1741     computeKnownBits(X, KnownZero, KnownOne, Depth, Q);
1742     if (KnownOne[0])
1743       return true;
1744   }
1745   // shr X, Y != 0 if X is negative.  Note that the value of the shift is not
1746   // defined if the sign bit is shifted off the end.
1747   else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1748     // shr exact can only shift out zero bits.
1749     PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1750     if (BO->isExact())
1751       return isKnownNonZero(X, Depth, Q);
1752 
1753     bool XKnownNonNegative, XKnownNegative;
1754     ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q);
1755     if (XKnownNegative)
1756       return true;
1757 
1758     // If the shifter operand is a constant, and all of the bits shifted
1759     // out are known to be zero, and X is known non-zero then at least one
1760     // non-zero bit must remain.
1761     if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1762       APInt KnownZero(BitWidth, 0);
1763       APInt KnownOne(BitWidth, 0);
1764       computeKnownBits(X, KnownZero, KnownOne, Depth, Q);
1765 
1766       auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1767       // Is there a known one in the portion not shifted out?
1768       if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1769         return true;
1770       // Are all the bits to be shifted out known zero?
1771       if (KnownZero.countTrailingOnes() >= ShiftVal)
1772         return isKnownNonZero(X, Depth, Q);
1773     }
1774   }
1775   // div exact can only produce a zero if the dividend is zero.
1776   else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1777     return isKnownNonZero(X, Depth, Q);
1778   }
1779   // X + Y.
1780   else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1781     bool XKnownNonNegative, XKnownNegative;
1782     bool YKnownNonNegative, YKnownNegative;
1783     ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q);
1784     ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Depth, Q);
1785 
1786     // If X and Y are both non-negative (as signed values) then their sum is not
1787     // zero unless both X and Y are zero.
1788     if (XKnownNonNegative && YKnownNonNegative)
1789       if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
1790         return true;
1791 
1792     // If X and Y are both negative (as signed values) then their sum is not
1793     // zero unless both X and Y equal INT_MIN.
1794     if (BitWidth && XKnownNegative && YKnownNegative) {
1795       APInt KnownZero(BitWidth, 0);
1796       APInt KnownOne(BitWidth, 0);
1797       APInt Mask = APInt::getSignedMaxValue(BitWidth);
1798       // The sign bit of X is set.  If some other bit is set then X is not equal
1799       // to INT_MIN.
1800       computeKnownBits(X, KnownZero, KnownOne, Depth, Q);
1801       if ((KnownOne & Mask) != 0)
1802         return true;
1803       // The sign bit of Y is set.  If some other bit is set then Y is not equal
1804       // to INT_MIN.
1805       computeKnownBits(Y, KnownZero, KnownOne, Depth, Q);
1806       if ((KnownOne & Mask) != 0)
1807         return true;
1808     }
1809 
1810     // The sum of a non-negative number and a power of two is not zero.
1811     if (XKnownNonNegative &&
1812         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
1813       return true;
1814     if (YKnownNonNegative &&
1815         isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
1816       return true;
1817   }
1818   // X * Y.
1819   else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1820     OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1821     // If X and Y are non-zero then so is X * Y as long as the multiplication
1822     // does not overflow.
1823     if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1824         isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
1825       return true;
1826   }
1827   // (C ? X : Y) != 0 if X != 0 and Y != 0.
1828   else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1829     if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
1830         isKnownNonZero(SI->getFalseValue(), Depth, Q))
1831       return true;
1832   }
1833   // PHI
1834   else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1835     // Try and detect a recurrence that monotonically increases from a
1836     // starting value, as these are common as induction variables.
1837     if (PN->getNumIncomingValues() == 2) {
1838       Value *Start = PN->getIncomingValue(0);
1839       Value *Induction = PN->getIncomingValue(1);
1840       if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
1841         std::swap(Start, Induction);
1842       if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
1843         if (!C->isZero() && !C->isNegative()) {
1844           ConstantInt *X;
1845           if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
1846                match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
1847               !X->isNegative())
1848             return true;
1849         }
1850       }
1851     }
1852     // Check if all incoming values are non-zero constant.
1853     bool AllNonZeroConstants = all_of(PN->operands(), [](Value *V) {
1854       return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZeroValue();
1855     });
1856     if (AllNonZeroConstants)
1857       return true;
1858   }
1859 
1860   if (!BitWidth) return false;
1861   APInt KnownZero(BitWidth, 0);
1862   APInt KnownOne(BitWidth, 0);
1863   computeKnownBits(V, KnownZero, KnownOne, Depth, Q);
1864   return KnownOne != 0;
1865 }
1866 
1867 /// Return true if V2 == V1 + X, where X is known non-zero.
isAddOfNonZero(Value * V1,Value * V2,const Query & Q)1868 static bool isAddOfNonZero(Value *V1, Value *V2, const Query &Q) {
1869   BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
1870   if (!BO || BO->getOpcode() != Instruction::Add)
1871     return false;
1872   Value *Op = nullptr;
1873   if (V2 == BO->getOperand(0))
1874     Op = BO->getOperand(1);
1875   else if (V2 == BO->getOperand(1))
1876     Op = BO->getOperand(0);
1877   else
1878     return false;
1879   return isKnownNonZero(Op, 0, Q);
1880 }
1881 
1882 /// Return true if it is known that V1 != V2.
isKnownNonEqual(Value * V1,Value * V2,const Query & Q)1883 static bool isKnownNonEqual(Value *V1, Value *V2, const Query &Q) {
1884   if (V1->getType()->isVectorTy() || V1 == V2)
1885     return false;
1886   if (V1->getType() != V2->getType())
1887     // We can't look through casts yet.
1888     return false;
1889   if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
1890     return true;
1891 
1892   if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
1893     // Are any known bits in V1 contradictory to known bits in V2? If V1
1894     // has a known zero where V2 has a known one, they must not be equal.
1895     auto BitWidth = Ty->getBitWidth();
1896     APInt KnownZero1(BitWidth, 0);
1897     APInt KnownOne1(BitWidth, 0);
1898     computeKnownBits(V1, KnownZero1, KnownOne1, 0, Q);
1899     APInt KnownZero2(BitWidth, 0);
1900     APInt KnownOne2(BitWidth, 0);
1901     computeKnownBits(V2, KnownZero2, KnownOne2, 0, Q);
1902 
1903     auto OppositeBits = (KnownZero1 & KnownOne2) | (KnownZero2 & KnownOne1);
1904     if (OppositeBits.getBoolValue())
1905       return true;
1906   }
1907   return false;
1908 }
1909 
1910 /// Return true if 'V & Mask' is known to be zero.  We use this predicate to
1911 /// simplify operations downstream. Mask is known to be zero for bits that V
1912 /// cannot have.
1913 ///
1914 /// This function is defined on values with integer type, values with pointer
1915 /// type, and vectors of integers.  In the case
1916 /// where V is a vector, the mask, known zero, and known one values are the
1917 /// same width as the vector element, and the bit is set only if it is true
1918 /// for all of the elements in the vector.
MaskedValueIsZero(Value * V,const APInt & Mask,unsigned Depth,const Query & Q)1919 bool MaskedValueIsZero(Value *V, const APInt &Mask, unsigned Depth,
1920                        const Query &Q) {
1921   APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1922   computeKnownBits(V, KnownZero, KnownOne, Depth, Q);
1923   return (KnownZero & Mask) == Mask;
1924 }
1925 
1926 /// For vector constants, loop over the elements and find the constant with the
1927 /// minimum number of sign bits. Return 0 if the value is not a vector constant
1928 /// or if any element was not analyzed; otherwise, return the count for the
1929 /// element with the minimum number of sign bits.
computeNumSignBitsVectorConstant(Value * V,unsigned TyBits)1930 static unsigned computeNumSignBitsVectorConstant(Value *V, unsigned TyBits) {
1931   auto *CV = dyn_cast<Constant>(V);
1932   if (!CV || !CV->getType()->isVectorTy())
1933     return 0;
1934 
1935   unsigned MinSignBits = TyBits;
1936   unsigned NumElts = CV->getType()->getVectorNumElements();
1937   for (unsigned i = 0; i != NumElts; ++i) {
1938     // If we find a non-ConstantInt, bail out.
1939     auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
1940     if (!Elt)
1941       return 0;
1942 
1943     // If the sign bit is 1, flip the bits, so we always count leading zeros.
1944     APInt EltVal = Elt->getValue();
1945     if (EltVal.isNegative())
1946       EltVal = ~EltVal;
1947     MinSignBits = std::min(MinSignBits, EltVal.countLeadingZeros());
1948   }
1949 
1950   return MinSignBits;
1951 }
1952 
1953 /// Return the number of times the sign bit of the register is replicated into
1954 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1955 /// (itself), but other cases can give us information. For example, immediately
1956 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1957 /// other, so we return 3. For vectors, return the number of sign bits for the
1958 /// vector element with the mininum number of known sign bits.
ComputeNumSignBits(Value * V,unsigned Depth,const Query & Q)1959 unsigned ComputeNumSignBits(Value *V, unsigned Depth, const Query &Q) {
1960   unsigned TyBits = Q.DL.getTypeSizeInBits(V->getType()->getScalarType());
1961   unsigned Tmp, Tmp2;
1962   unsigned FirstAnswer = 1;
1963 
1964   // Note that ConstantInt is handled by the general computeKnownBits case
1965   // below.
1966 
1967   if (Depth == 6)
1968     return 1;  // Limit search depth.
1969 
1970   Operator *U = dyn_cast<Operator>(V);
1971   switch (Operator::getOpcode(V)) {
1972   default: break;
1973   case Instruction::SExt:
1974     Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1975     return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
1976 
1977   case Instruction::SDiv: {
1978     const APInt *Denominator;
1979     // sdiv X, C -> adds log(C) sign bits.
1980     if (match(U->getOperand(1), m_APInt(Denominator))) {
1981 
1982       // Ignore non-positive denominator.
1983       if (!Denominator->isStrictlyPositive())
1984         break;
1985 
1986       // Calculate the incoming numerator bits.
1987       unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
1988 
1989       // Add floor(log(C)) bits to the numerator bits.
1990       return std::min(TyBits, NumBits + Denominator->logBase2());
1991     }
1992     break;
1993   }
1994 
1995   case Instruction::SRem: {
1996     const APInt *Denominator;
1997     // srem X, C -> we know that the result is within [-C+1,C) when C is a
1998     // positive constant.  This let us put a lower bound on the number of sign
1999     // bits.
2000     if (match(U->getOperand(1), m_APInt(Denominator))) {
2001 
2002       // Ignore non-positive denominator.
2003       if (!Denominator->isStrictlyPositive())
2004         break;
2005 
2006       // Calculate the incoming numerator bits. SRem by a positive constant
2007       // can't lower the number of sign bits.
2008       unsigned NumrBits =
2009           ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2010 
2011       // Calculate the leading sign bit constraints by examining the
2012       // denominator.  Given that the denominator is positive, there are two
2013       // cases:
2014       //
2015       //  1. the numerator is positive.  The result range is [0,C) and [0,C) u<
2016       //     (1 << ceilLogBase2(C)).
2017       //
2018       //  2. the numerator is negative.  Then the result range is (-C,0] and
2019       //     integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2020       //
2021       // Thus a lower bound on the number of sign bits is `TyBits -
2022       // ceilLogBase2(C)`.
2023 
2024       unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2025       return std::max(NumrBits, ResBits);
2026     }
2027     break;
2028   }
2029 
2030   case Instruction::AShr: {
2031     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2032     // ashr X, C   -> adds C sign bits.  Vectors too.
2033     const APInt *ShAmt;
2034     if (match(U->getOperand(1), m_APInt(ShAmt))) {
2035       Tmp += ShAmt->getZExtValue();
2036       if (Tmp > TyBits) Tmp = TyBits;
2037     }
2038     return Tmp;
2039   }
2040   case Instruction::Shl: {
2041     const APInt *ShAmt;
2042     if (match(U->getOperand(1), m_APInt(ShAmt))) {
2043       // shl destroys sign bits.
2044       Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2045       Tmp2 = ShAmt->getZExtValue();
2046       if (Tmp2 >= TyBits ||      // Bad shift.
2047           Tmp2 >= Tmp) break;    // Shifted all sign bits out.
2048       return Tmp - Tmp2;
2049     }
2050     break;
2051   }
2052   case Instruction::And:
2053   case Instruction::Or:
2054   case Instruction::Xor:    // NOT is handled here.
2055     // Logical binary ops preserve the number of sign bits at the worst.
2056     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2057     if (Tmp != 1) {
2058       Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2059       FirstAnswer = std::min(Tmp, Tmp2);
2060       // We computed what we know about the sign bits as our first
2061       // answer. Now proceed to the generic code that uses
2062       // computeKnownBits, and pick whichever answer is better.
2063     }
2064     break;
2065 
2066   case Instruction::Select:
2067     Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2068     if (Tmp == 1) return 1;  // Early out.
2069     Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2070     return std::min(Tmp, Tmp2);
2071 
2072   case Instruction::Add:
2073     // Add can have at most one carry bit.  Thus we know that the output
2074     // is, at worst, one more bit than the inputs.
2075     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2076     if (Tmp == 1) return 1;  // Early out.
2077 
2078     // Special case decrementing a value (ADD X, -1):
2079     if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2080       if (CRHS->isAllOnesValue()) {
2081         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2082         computeKnownBits(U->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
2083 
2084         // If the input is known to be 0 or 1, the output is 0/-1, which is all
2085         // sign bits set.
2086         if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2087           return TyBits;
2088 
2089         // If we are subtracting one from a positive number, there is no carry
2090         // out of the result.
2091         if (KnownZero.isNegative())
2092           return Tmp;
2093       }
2094 
2095     Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2096     if (Tmp2 == 1) return 1;
2097     return std::min(Tmp, Tmp2)-1;
2098 
2099   case Instruction::Sub:
2100     Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2101     if (Tmp2 == 1) return 1;
2102 
2103     // Handle NEG.
2104     if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2105       if (CLHS->isNullValue()) {
2106         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2107         computeKnownBits(U->getOperand(1), KnownZero, KnownOne, Depth + 1, Q);
2108         // If the input is known to be 0 or 1, the output is 0/-1, which is all
2109         // sign bits set.
2110         if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2111           return TyBits;
2112 
2113         // If the input is known to be positive (the sign bit is known clear),
2114         // the output of the NEG has the same number of sign bits as the input.
2115         if (KnownZero.isNegative())
2116           return Tmp2;
2117 
2118         // Otherwise, we treat this like a SUB.
2119       }
2120 
2121     // Sub can have at most one carry bit.  Thus we know that the output
2122     // is, at worst, one more bit than the inputs.
2123     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2124     if (Tmp == 1) return 1;  // Early out.
2125     return std::min(Tmp, Tmp2)-1;
2126 
2127   case Instruction::PHI: {
2128     PHINode *PN = cast<PHINode>(U);
2129     unsigned NumIncomingValues = PN->getNumIncomingValues();
2130     // Don't analyze large in-degree PHIs.
2131     if (NumIncomingValues > 4) break;
2132     // Unreachable blocks may have zero-operand PHI nodes.
2133     if (NumIncomingValues == 0) break;
2134 
2135     // Take the minimum of all incoming values.  This can't infinitely loop
2136     // because of our depth threshold.
2137     Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2138     for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2139       if (Tmp == 1) return Tmp;
2140       Tmp = std::min(
2141           Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2142     }
2143     return Tmp;
2144   }
2145 
2146   case Instruction::Trunc:
2147     // FIXME: it's tricky to do anything useful for this, but it is an important
2148     // case for targets like X86.
2149     break;
2150   }
2151 
2152   // Finally, if we can prove that the top bits of the result are 0's or 1's,
2153   // use this information.
2154 
2155   // If we can examine all elements of a vector constant successfully, we're
2156   // done (we can't do any better than that). If not, keep trying.
2157   if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2158     return VecSignBits;
2159 
2160   APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2161   computeKnownBits(V, KnownZero, KnownOne, Depth, Q);
2162 
2163   // If we know that the sign bit is either zero or one, determine the number of
2164   // identical bits in the top of the input value.
2165   if (KnownZero.isNegative())
2166     return std::max(FirstAnswer, KnownZero.countLeadingOnes());
2167 
2168   if (KnownOne.isNegative())
2169     return std::max(FirstAnswer, KnownOne.countLeadingOnes());
2170 
2171   // computeKnownBits gave us no extra information about the top bits.
2172   return FirstAnswer;
2173 }
2174 
2175 /// This function computes the integer multiple of Base that equals V.
2176 /// If successful, it returns true and returns the multiple in
2177 /// Multiple. If unsuccessful, it returns false. It looks
2178 /// through SExt instructions only if LookThroughSExt is true.
ComputeMultiple(Value * V,unsigned Base,Value * & Multiple,bool LookThroughSExt,unsigned Depth)2179 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2180                            bool LookThroughSExt, unsigned Depth) {
2181   const unsigned MaxDepth = 6;
2182 
2183   assert(V && "No Value?");
2184   assert(Depth <= MaxDepth && "Limit Search Depth");
2185   assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2186 
2187   Type *T = V->getType();
2188 
2189   ConstantInt *CI = dyn_cast<ConstantInt>(V);
2190 
2191   if (Base == 0)
2192     return false;
2193 
2194   if (Base == 1) {
2195     Multiple = V;
2196     return true;
2197   }
2198 
2199   ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2200   Constant *BaseVal = ConstantInt::get(T, Base);
2201   if (CO && CO == BaseVal) {
2202     // Multiple is 1.
2203     Multiple = ConstantInt::get(T, 1);
2204     return true;
2205   }
2206 
2207   if (CI && CI->getZExtValue() % Base == 0) {
2208     Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2209     return true;
2210   }
2211 
2212   if (Depth == MaxDepth) return false;  // Limit search depth.
2213 
2214   Operator *I = dyn_cast<Operator>(V);
2215   if (!I) return false;
2216 
2217   switch (I->getOpcode()) {
2218   default: break;
2219   case Instruction::SExt:
2220     if (!LookThroughSExt) return false;
2221     // otherwise fall through to ZExt
2222   case Instruction::ZExt:
2223     return ComputeMultiple(I->getOperand(0), Base, Multiple,
2224                            LookThroughSExt, Depth+1);
2225   case Instruction::Shl:
2226   case Instruction::Mul: {
2227     Value *Op0 = I->getOperand(0);
2228     Value *Op1 = I->getOperand(1);
2229 
2230     if (I->getOpcode() == Instruction::Shl) {
2231       ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2232       if (!Op1CI) return false;
2233       // Turn Op0 << Op1 into Op0 * 2^Op1
2234       APInt Op1Int = Op1CI->getValue();
2235       uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2236       APInt API(Op1Int.getBitWidth(), 0);
2237       API.setBit(BitToSet);
2238       Op1 = ConstantInt::get(V->getContext(), API);
2239     }
2240 
2241     Value *Mul0 = nullptr;
2242     if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2243       if (Constant *Op1C = dyn_cast<Constant>(Op1))
2244         if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2245           if (Op1C->getType()->getPrimitiveSizeInBits() <
2246               MulC->getType()->getPrimitiveSizeInBits())
2247             Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2248           if (Op1C->getType()->getPrimitiveSizeInBits() >
2249               MulC->getType()->getPrimitiveSizeInBits())
2250             MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2251 
2252           // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2253           Multiple = ConstantExpr::getMul(MulC, Op1C);
2254           return true;
2255         }
2256 
2257       if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2258         if (Mul0CI->getValue() == 1) {
2259           // V == Base * Op1, so return Op1
2260           Multiple = Op1;
2261           return true;
2262         }
2263     }
2264 
2265     Value *Mul1 = nullptr;
2266     if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2267       if (Constant *Op0C = dyn_cast<Constant>(Op0))
2268         if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2269           if (Op0C->getType()->getPrimitiveSizeInBits() <
2270               MulC->getType()->getPrimitiveSizeInBits())
2271             Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2272           if (Op0C->getType()->getPrimitiveSizeInBits() >
2273               MulC->getType()->getPrimitiveSizeInBits())
2274             MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2275 
2276           // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2277           Multiple = ConstantExpr::getMul(MulC, Op0C);
2278           return true;
2279         }
2280 
2281       if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2282         if (Mul1CI->getValue() == 1) {
2283           // V == Base * Op0, so return Op0
2284           Multiple = Op0;
2285           return true;
2286         }
2287     }
2288   }
2289   }
2290 
2291   // We could not determine if V is a multiple of Base.
2292   return false;
2293 }
2294 
getIntrinsicForCallSite(ImmutableCallSite ICS,const TargetLibraryInfo * TLI)2295 Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
2296                                             const TargetLibraryInfo *TLI) {
2297   const Function *F = ICS.getCalledFunction();
2298   if (!F)
2299     return Intrinsic::not_intrinsic;
2300 
2301   if (F->isIntrinsic())
2302     return F->getIntrinsicID();
2303 
2304   if (!TLI)
2305     return Intrinsic::not_intrinsic;
2306 
2307   LibFunc::Func Func;
2308   // We're going to make assumptions on the semantics of the functions, check
2309   // that the target knows that it's available in this environment and it does
2310   // not have local linkage.
2311   if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2312     return Intrinsic::not_intrinsic;
2313 
2314   if (!ICS.onlyReadsMemory())
2315     return Intrinsic::not_intrinsic;
2316 
2317   // Otherwise check if we have a call to a function that can be turned into a
2318   // vector intrinsic.
2319   switch (Func) {
2320   default:
2321     break;
2322   case LibFunc::sin:
2323   case LibFunc::sinf:
2324   case LibFunc::sinl:
2325     return Intrinsic::sin;
2326   case LibFunc::cos:
2327   case LibFunc::cosf:
2328   case LibFunc::cosl:
2329     return Intrinsic::cos;
2330   case LibFunc::exp:
2331   case LibFunc::expf:
2332   case LibFunc::expl:
2333     return Intrinsic::exp;
2334   case LibFunc::exp2:
2335   case LibFunc::exp2f:
2336   case LibFunc::exp2l:
2337     return Intrinsic::exp2;
2338   case LibFunc::log:
2339   case LibFunc::logf:
2340   case LibFunc::logl:
2341     return Intrinsic::log;
2342   case LibFunc::log10:
2343   case LibFunc::log10f:
2344   case LibFunc::log10l:
2345     return Intrinsic::log10;
2346   case LibFunc::log2:
2347   case LibFunc::log2f:
2348   case LibFunc::log2l:
2349     return Intrinsic::log2;
2350   case LibFunc::fabs:
2351   case LibFunc::fabsf:
2352   case LibFunc::fabsl:
2353     return Intrinsic::fabs;
2354   case LibFunc::fmin:
2355   case LibFunc::fminf:
2356   case LibFunc::fminl:
2357     return Intrinsic::minnum;
2358   case LibFunc::fmax:
2359   case LibFunc::fmaxf:
2360   case LibFunc::fmaxl:
2361     return Intrinsic::maxnum;
2362   case LibFunc::copysign:
2363   case LibFunc::copysignf:
2364   case LibFunc::copysignl:
2365     return Intrinsic::copysign;
2366   case LibFunc::floor:
2367   case LibFunc::floorf:
2368   case LibFunc::floorl:
2369     return Intrinsic::floor;
2370   case LibFunc::ceil:
2371   case LibFunc::ceilf:
2372   case LibFunc::ceill:
2373     return Intrinsic::ceil;
2374   case LibFunc::trunc:
2375   case LibFunc::truncf:
2376   case LibFunc::truncl:
2377     return Intrinsic::trunc;
2378   case LibFunc::rint:
2379   case LibFunc::rintf:
2380   case LibFunc::rintl:
2381     return Intrinsic::rint;
2382   case LibFunc::nearbyint:
2383   case LibFunc::nearbyintf:
2384   case LibFunc::nearbyintl:
2385     return Intrinsic::nearbyint;
2386   case LibFunc::round:
2387   case LibFunc::roundf:
2388   case LibFunc::roundl:
2389     return Intrinsic::round;
2390   case LibFunc::pow:
2391   case LibFunc::powf:
2392   case LibFunc::powl:
2393     return Intrinsic::pow;
2394   case LibFunc::sqrt:
2395   case LibFunc::sqrtf:
2396   case LibFunc::sqrtl:
2397     if (ICS->hasNoNaNs())
2398       return Intrinsic::sqrt;
2399     return Intrinsic::not_intrinsic;
2400   }
2401 
2402   return Intrinsic::not_intrinsic;
2403 }
2404 
2405 /// Return true if we can prove that the specified FP value is never equal to
2406 /// -0.0.
2407 ///
2408 /// NOTE: this function will need to be revisited when we support non-default
2409 /// rounding modes!
2410 ///
CannotBeNegativeZero(const Value * V,const TargetLibraryInfo * TLI,unsigned Depth)2411 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
2412                                 unsigned Depth) {
2413   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2414     return !CFP->getValueAPF().isNegZero();
2415 
2416   // FIXME: Magic number! At the least, this should be given a name because it's
2417   // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2418   // expose it as a parameter, so it can be used for testing / experimenting.
2419   if (Depth == 6)
2420     return false;  // Limit search depth.
2421 
2422   const Operator *I = dyn_cast<Operator>(V);
2423   if (!I) return false;
2424 
2425   // Check if the nsz fast-math flag is set
2426   if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2427     if (FPO->hasNoSignedZeros())
2428       return true;
2429 
2430   // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2431   if (I->getOpcode() == Instruction::FAdd)
2432     if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2433       if (CFP->isNullValue())
2434         return true;
2435 
2436   // sitofp and uitofp turn into +0.0 for zero.
2437   if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2438     return true;
2439 
2440   if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2441     Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2442     switch (IID) {
2443     default:
2444       break;
2445     // sqrt(-0.0) = -0.0, no other negative results are possible.
2446     case Intrinsic::sqrt:
2447       return CannotBeNegativeZero(CI->getArgOperand(0), TLI, Depth + 1);
2448     // fabs(x) != -0.0
2449     case Intrinsic::fabs:
2450       return true;
2451     }
2452   }
2453 
2454   return false;
2455 }
2456 
CannotBeOrderedLessThanZero(const Value * V,const TargetLibraryInfo * TLI,unsigned Depth)2457 bool llvm::CannotBeOrderedLessThanZero(const Value *V,
2458                                        const TargetLibraryInfo *TLI,
2459                                        unsigned Depth) {
2460   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2461     return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2462 
2463   // FIXME: Magic number! At the least, this should be given a name because it's
2464   // used similarly in CannotBeNegativeZero(). A better fix may be to
2465   // expose it as a parameter, so it can be used for testing / experimenting.
2466   if (Depth == 6)
2467     return false;  // Limit search depth.
2468 
2469   const Operator *I = dyn_cast<Operator>(V);
2470   if (!I) return false;
2471 
2472   switch (I->getOpcode()) {
2473   default: break;
2474   // Unsigned integers are always nonnegative.
2475   case Instruction::UIToFP:
2476     return true;
2477   case Instruction::FMul:
2478     // x*x is always non-negative or a NaN.
2479     if (I->getOperand(0) == I->getOperand(1))
2480       return true;
2481     // Fall through
2482   case Instruction::FAdd:
2483   case Instruction::FDiv:
2484   case Instruction::FRem:
2485     return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1) &&
2486            CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1);
2487   case Instruction::Select:
2488     return CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1) &&
2489            CannotBeOrderedLessThanZero(I->getOperand(2), TLI, Depth + 1);
2490   case Instruction::FPExt:
2491   case Instruction::FPTrunc:
2492     // Widening/narrowing never change sign.
2493     return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1);
2494   case Instruction::Call:
2495     Intrinsic::ID IID = getIntrinsicForCallSite(cast<CallInst>(I), TLI);
2496     switch (IID) {
2497     default:
2498       break;
2499     case Intrinsic::maxnum:
2500       return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1) ||
2501              CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1);
2502     case Intrinsic::minnum:
2503       return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1) &&
2504              CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1);
2505     case Intrinsic::exp:
2506     case Intrinsic::exp2:
2507     case Intrinsic::fabs:
2508     case Intrinsic::sqrt:
2509       return true;
2510     case Intrinsic::powi:
2511       if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2512         // powi(x,n) is non-negative if n is even.
2513         if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2514           return true;
2515       }
2516       return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1);
2517     case Intrinsic::fma:
2518     case Intrinsic::fmuladd:
2519       // x*x+y is non-negative if y is non-negative.
2520       return I->getOperand(0) == I->getOperand(1) &&
2521              CannotBeOrderedLessThanZero(I->getOperand(2), TLI, Depth + 1);
2522     }
2523     break;
2524   }
2525   return false;
2526 }
2527 
2528 /// If the specified value can be set by repeating the same byte in memory,
2529 /// return the i8 value that it is represented with.  This is
2530 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2531 /// i16 0xF0F0, double 0.0 etc.  If the value can't be handled with a repeated
2532 /// byte store (e.g. i16 0x1234), return null.
isBytewiseValue(Value * V)2533 Value *llvm::isBytewiseValue(Value *V) {
2534   // All byte-wide stores are splatable, even of arbitrary variables.
2535   if (V->getType()->isIntegerTy(8)) return V;
2536 
2537   // Handle 'null' ConstantArrayZero etc.
2538   if (Constant *C = dyn_cast<Constant>(V))
2539     if (C->isNullValue())
2540       return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2541 
2542   // Constant float and double values can be handled as integer values if the
2543   // corresponding integer value is "byteable".  An important case is 0.0.
2544   if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2545     if (CFP->getType()->isFloatTy())
2546       V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2547     if (CFP->getType()->isDoubleTy())
2548       V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2549     // Don't handle long double formats, which have strange constraints.
2550   }
2551 
2552   // We can handle constant integers that are multiple of 8 bits.
2553   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2554     if (CI->getBitWidth() % 8 == 0) {
2555       assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2556 
2557       if (!CI->getValue().isSplat(8))
2558         return nullptr;
2559       return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2560     }
2561   }
2562 
2563   // A ConstantDataArray/Vector is splatable if all its members are equal and
2564   // also splatable.
2565   if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2566     Value *Elt = CA->getElementAsConstant(0);
2567     Value *Val = isBytewiseValue(Elt);
2568     if (!Val)
2569       return nullptr;
2570 
2571     for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2572       if (CA->getElementAsConstant(I) != Elt)
2573         return nullptr;
2574 
2575     return Val;
2576   }
2577 
2578   // Conceptually, we could handle things like:
2579   //   %a = zext i8 %X to i16
2580   //   %b = shl i16 %a, 8
2581   //   %c = or i16 %a, %b
2582   // but until there is an example that actually needs this, it doesn't seem
2583   // worth worrying about.
2584   return nullptr;
2585 }
2586 
2587 
2588 // This is the recursive version of BuildSubAggregate. It takes a few different
2589 // arguments. Idxs is the index within the nested struct From that we are
2590 // looking at now (which is of type IndexedType). IdxSkip is the number of
2591 // indices from Idxs that should be left out when inserting into the resulting
2592 // struct. To is the result struct built so far, new insertvalue instructions
2593 // build on that.
BuildSubAggregate(Value * From,Value * To,Type * IndexedType,SmallVectorImpl<unsigned> & Idxs,unsigned IdxSkip,Instruction * InsertBefore)2594 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2595                                 SmallVectorImpl<unsigned> &Idxs,
2596                                 unsigned IdxSkip,
2597                                 Instruction *InsertBefore) {
2598   llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2599   if (STy) {
2600     // Save the original To argument so we can modify it
2601     Value *OrigTo = To;
2602     // General case, the type indexed by Idxs is a struct
2603     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2604       // Process each struct element recursively
2605       Idxs.push_back(i);
2606       Value *PrevTo = To;
2607       To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2608                              InsertBefore);
2609       Idxs.pop_back();
2610       if (!To) {
2611         // Couldn't find any inserted value for this index? Cleanup
2612         while (PrevTo != OrigTo) {
2613           InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2614           PrevTo = Del->getAggregateOperand();
2615           Del->eraseFromParent();
2616         }
2617         // Stop processing elements
2618         break;
2619       }
2620     }
2621     // If we successfully found a value for each of our subaggregates
2622     if (To)
2623       return To;
2624   }
2625   // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2626   // the struct's elements had a value that was inserted directly. In the latter
2627   // case, perhaps we can't determine each of the subelements individually, but
2628   // we might be able to find the complete struct somewhere.
2629 
2630   // Find the value that is at that particular spot
2631   Value *V = FindInsertedValue(From, Idxs);
2632 
2633   if (!V)
2634     return nullptr;
2635 
2636   // Insert the value in the new (sub) aggregrate
2637   return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2638                                        "tmp", InsertBefore);
2639 }
2640 
2641 // This helper takes a nested struct and extracts a part of it (which is again a
2642 // struct) into a new value. For example, given the struct:
2643 // { a, { b, { c, d }, e } }
2644 // and the indices "1, 1" this returns
2645 // { c, d }.
2646 //
2647 // It does this by inserting an insertvalue for each element in the resulting
2648 // struct, as opposed to just inserting a single struct. This will only work if
2649 // each of the elements of the substruct are known (ie, inserted into From by an
2650 // insertvalue instruction somewhere).
2651 //
2652 // All inserted insertvalue instructions are inserted before InsertBefore
BuildSubAggregate(Value * From,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)2653 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2654                                 Instruction *InsertBefore) {
2655   assert(InsertBefore && "Must have someplace to insert!");
2656   Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2657                                                              idx_range);
2658   Value *To = UndefValue::get(IndexedType);
2659   SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2660   unsigned IdxSkip = Idxs.size();
2661 
2662   return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2663 }
2664 
2665 /// Given an aggregrate and an sequence of indices, see if
2666 /// the scalar value indexed is already around as a register, for example if it
2667 /// were inserted directly into the aggregrate.
2668 ///
2669 /// If InsertBefore is not null, this function will duplicate (modified)
2670 /// insertvalues when a part of a nested struct is extracted.
FindInsertedValue(Value * V,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)2671 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2672                                Instruction *InsertBefore) {
2673   // Nothing to index? Just return V then (this is useful at the end of our
2674   // recursion).
2675   if (idx_range.empty())
2676     return V;
2677   // We have indices, so V should have an indexable type.
2678   assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2679          "Not looking at a struct or array?");
2680   assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2681          "Invalid indices for type?");
2682 
2683   if (Constant *C = dyn_cast<Constant>(V)) {
2684     C = C->getAggregateElement(idx_range[0]);
2685     if (!C) return nullptr;
2686     return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2687   }
2688 
2689   if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2690     // Loop the indices for the insertvalue instruction in parallel with the
2691     // requested indices
2692     const unsigned *req_idx = idx_range.begin();
2693     for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2694          i != e; ++i, ++req_idx) {
2695       if (req_idx == idx_range.end()) {
2696         // We can't handle this without inserting insertvalues
2697         if (!InsertBefore)
2698           return nullptr;
2699 
2700         // The requested index identifies a part of a nested aggregate. Handle
2701         // this specially. For example,
2702         // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2703         // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2704         // %C = extractvalue {i32, { i32, i32 } } %B, 1
2705         // This can be changed into
2706         // %A = insertvalue {i32, i32 } undef, i32 10, 0
2707         // %C = insertvalue {i32, i32 } %A, i32 11, 1
2708         // which allows the unused 0,0 element from the nested struct to be
2709         // removed.
2710         return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2711                                  InsertBefore);
2712       }
2713 
2714       // This insert value inserts something else than what we are looking for.
2715       // See if the (aggregate) value inserted into has the value we are
2716       // looking for, then.
2717       if (*req_idx != *i)
2718         return FindInsertedValue(I->getAggregateOperand(), idx_range,
2719                                  InsertBefore);
2720     }
2721     // If we end up here, the indices of the insertvalue match with those
2722     // requested (though possibly only partially). Now we recursively look at
2723     // the inserted value, passing any remaining indices.
2724     return FindInsertedValue(I->getInsertedValueOperand(),
2725                              makeArrayRef(req_idx, idx_range.end()),
2726                              InsertBefore);
2727   }
2728 
2729   if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2730     // If we're extracting a value from an aggregate that was extracted from
2731     // something else, we can extract from that something else directly instead.
2732     // However, we will need to chain I's indices with the requested indices.
2733 
2734     // Calculate the number of indices required
2735     unsigned size = I->getNumIndices() + idx_range.size();
2736     // Allocate some space to put the new indices in
2737     SmallVector<unsigned, 5> Idxs;
2738     Idxs.reserve(size);
2739     // Add indices from the extract value instruction
2740     Idxs.append(I->idx_begin(), I->idx_end());
2741 
2742     // Add requested indices
2743     Idxs.append(idx_range.begin(), idx_range.end());
2744 
2745     assert(Idxs.size() == size
2746            && "Number of indices added not correct?");
2747 
2748     return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2749   }
2750   // Otherwise, we don't know (such as, extracting from a function return value
2751   // or load instruction)
2752   return nullptr;
2753 }
2754 
2755 /// Analyze the specified pointer to see if it can be expressed as a base
2756 /// pointer plus a constant offset. Return the base and offset to the caller.
GetPointerBaseWithConstantOffset(Value * Ptr,int64_t & Offset,const DataLayout & DL)2757 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2758                                               const DataLayout &DL) {
2759   unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2760   APInt ByteOffset(BitWidth, 0);
2761 
2762   // We walk up the defs but use a visited set to handle unreachable code. In
2763   // that case, we stop after accumulating the cycle once (not that it
2764   // matters).
2765   SmallPtrSet<Value *, 16> Visited;
2766   while (Visited.insert(Ptr).second) {
2767     if (Ptr->getType()->isVectorTy())
2768       break;
2769 
2770     if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2771       APInt GEPOffset(BitWidth, 0);
2772       if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2773         break;
2774 
2775       ByteOffset += GEPOffset;
2776 
2777       Ptr = GEP->getPointerOperand();
2778     } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2779                Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2780       Ptr = cast<Operator>(Ptr)->getOperand(0);
2781     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2782       if (GA->isInterposable())
2783         break;
2784       Ptr = GA->getAliasee();
2785     } else {
2786       break;
2787     }
2788   }
2789   Offset = ByteOffset.getSExtValue();
2790   return Ptr;
2791 }
2792 
isGEPBasedOnPointerToString(const GEPOperator * GEP)2793 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP) {
2794   // Make sure the GEP has exactly three arguments.
2795   if (GEP->getNumOperands() != 3)
2796     return false;
2797 
2798   // Make sure the index-ee is a pointer to array of i8.
2799   ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
2800   if (!AT || !AT->getElementType()->isIntegerTy(8))
2801     return false;
2802 
2803   // Check to make sure that the first operand of the GEP is an integer and
2804   // has value 0 so that we are sure we're indexing into the initializer.
2805   const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2806   if (!FirstIdx || !FirstIdx->isZero())
2807     return false;
2808 
2809   return true;
2810 }
2811 
2812 /// This function computes the length of a null-terminated C string pointed to
2813 /// by V. If successful, it returns true and returns the string in Str.
2814 /// If unsuccessful, it returns false.
getConstantStringInfo(const Value * V,StringRef & Str,uint64_t Offset,bool TrimAtNul)2815 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2816                                  uint64_t Offset, bool TrimAtNul) {
2817   assert(V);
2818 
2819   // Look through bitcast instructions and geps.
2820   V = V->stripPointerCasts();
2821 
2822   // If the value is a GEP instruction or constant expression, treat it as an
2823   // offset.
2824   if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2825     // The GEP operator should be based on a pointer to string constant, and is
2826     // indexing into the string constant.
2827     if (!isGEPBasedOnPointerToString(GEP))
2828       return false;
2829 
2830     // If the second index isn't a ConstantInt, then this is a variable index
2831     // into the array.  If this occurs, we can't say anything meaningful about
2832     // the string.
2833     uint64_t StartIdx = 0;
2834     if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2835       StartIdx = CI->getZExtValue();
2836     else
2837       return false;
2838     return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2839                                  TrimAtNul);
2840   }
2841 
2842   // The GEP instruction, constant or instruction, must reference a global
2843   // variable that is a constant and is initialized. The referenced constant
2844   // initializer is the array that we'll use for optimization.
2845   const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2846   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2847     return false;
2848 
2849   // Handle the all-zeros case.
2850   if (GV->getInitializer()->isNullValue()) {
2851     // This is a degenerate case. The initializer is constant zero so the
2852     // length of the string must be zero.
2853     Str = "";
2854     return true;
2855   }
2856 
2857   // This must be a ConstantDataArray.
2858   const auto *Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
2859   if (!Array || !Array->isString())
2860     return false;
2861 
2862   // Get the number of elements in the array.
2863   uint64_t NumElts = Array->getType()->getArrayNumElements();
2864 
2865   // Start out with the entire array in the StringRef.
2866   Str = Array->getAsString();
2867 
2868   if (Offset > NumElts)
2869     return false;
2870 
2871   // Skip over 'offset' bytes.
2872   Str = Str.substr(Offset);
2873 
2874   if (TrimAtNul) {
2875     // Trim off the \0 and anything after it.  If the array is not nul
2876     // terminated, we just return the whole end of string.  The client may know
2877     // some other way that the string is length-bound.
2878     Str = Str.substr(0, Str.find('\0'));
2879   }
2880   return true;
2881 }
2882 
2883 // These next two are very similar to the above, but also look through PHI
2884 // nodes.
2885 // TODO: See if we can integrate these two together.
2886 
2887 /// If we can compute the length of the string pointed to by
2888 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLengthH(Value * V,SmallPtrSetImpl<PHINode * > & PHIs)2889 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2890   // Look through noop bitcast instructions.
2891   V = V->stripPointerCasts();
2892 
2893   // If this is a PHI node, there are two cases: either we have already seen it
2894   // or we haven't.
2895   if (PHINode *PN = dyn_cast<PHINode>(V)) {
2896     if (!PHIs.insert(PN).second)
2897       return ~0ULL;  // already in the set.
2898 
2899     // If it was new, see if all the input strings are the same length.
2900     uint64_t LenSoFar = ~0ULL;
2901     for (Value *IncValue : PN->incoming_values()) {
2902       uint64_t Len = GetStringLengthH(IncValue, PHIs);
2903       if (Len == 0) return 0; // Unknown length -> unknown.
2904 
2905       if (Len == ~0ULL) continue;
2906 
2907       if (Len != LenSoFar && LenSoFar != ~0ULL)
2908         return 0;    // Disagree -> unknown.
2909       LenSoFar = Len;
2910     }
2911 
2912     // Success, all agree.
2913     return LenSoFar;
2914   }
2915 
2916   // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2917   if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2918     uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2919     if (Len1 == 0) return 0;
2920     uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2921     if (Len2 == 0) return 0;
2922     if (Len1 == ~0ULL) return Len2;
2923     if (Len2 == ~0ULL) return Len1;
2924     if (Len1 != Len2) return 0;
2925     return Len1;
2926   }
2927 
2928   // Otherwise, see if we can read the string.
2929   StringRef StrData;
2930   if (!getConstantStringInfo(V, StrData))
2931     return 0;
2932 
2933   return StrData.size()+1;
2934 }
2935 
2936 /// If we can compute the length of the string pointed to by
2937 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLength(Value * V)2938 uint64_t llvm::GetStringLength(Value *V) {
2939   if (!V->getType()->isPointerTy()) return 0;
2940 
2941   SmallPtrSet<PHINode*, 32> PHIs;
2942   uint64_t Len = GetStringLengthH(V, PHIs);
2943   // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2944   // an empty string as a length.
2945   return Len == ~0ULL ? 1 : Len;
2946 }
2947 
2948 /// \brief \p PN defines a loop-variant pointer to an object.  Check if the
2949 /// previous iteration of the loop was referring to the same object as \p PN.
isSameUnderlyingObjectInLoop(PHINode * PN,LoopInfo * LI)2950 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2951   // Find the loop-defined value.
2952   Loop *L = LI->getLoopFor(PN->getParent());
2953   if (PN->getNumIncomingValues() != 2)
2954     return true;
2955 
2956   // Find the value from previous iteration.
2957   auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2958   if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2959     PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2960   if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2961     return true;
2962 
2963   // If a new pointer is loaded in the loop, the pointer references a different
2964   // object in every iteration.  E.g.:
2965   //    for (i)
2966   //       int *p = a[i];
2967   //       ...
2968   if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2969     if (!L->isLoopInvariant(Load->getPointerOperand()))
2970       return false;
2971   return true;
2972 }
2973 
GetUnderlyingObject(Value * V,const DataLayout & DL,unsigned MaxLookup)2974 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2975                                  unsigned MaxLookup) {
2976   if (!V->getType()->isPointerTy())
2977     return V;
2978   for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2979     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2980       V = GEP->getPointerOperand();
2981     } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2982                Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2983       V = cast<Operator>(V)->getOperand(0);
2984     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2985       if (GA->isInterposable())
2986         return V;
2987       V = GA->getAliasee();
2988     } else {
2989       if (auto CS = CallSite(V))
2990         if (Value *RV = CS.getReturnedArgOperand()) {
2991           V = RV;
2992           continue;
2993         }
2994 
2995       // See if InstructionSimplify knows any relevant tricks.
2996       if (Instruction *I = dyn_cast<Instruction>(V))
2997         // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2998         if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
2999           V = Simplified;
3000           continue;
3001         }
3002 
3003       return V;
3004     }
3005     assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3006   }
3007   return V;
3008 }
3009 
GetUnderlyingObjects(Value * V,SmallVectorImpl<Value * > & Objects,const DataLayout & DL,LoopInfo * LI,unsigned MaxLookup)3010 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3011                                 const DataLayout &DL, LoopInfo *LI,
3012                                 unsigned MaxLookup) {
3013   SmallPtrSet<Value *, 4> Visited;
3014   SmallVector<Value *, 4> Worklist;
3015   Worklist.push_back(V);
3016   do {
3017     Value *P = Worklist.pop_back_val();
3018     P = GetUnderlyingObject(P, DL, MaxLookup);
3019 
3020     if (!Visited.insert(P).second)
3021       continue;
3022 
3023     if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3024       Worklist.push_back(SI->getTrueValue());
3025       Worklist.push_back(SI->getFalseValue());
3026       continue;
3027     }
3028 
3029     if (PHINode *PN = dyn_cast<PHINode>(P)) {
3030       // If this PHI changes the underlying object in every iteration of the
3031       // loop, don't look through it.  Consider:
3032       //   int **A;
3033       //   for (i) {
3034       //     Prev = Curr;     // Prev = PHI (Prev_0, Curr)
3035       //     Curr = A[i];
3036       //     *Prev, *Curr;
3037       //
3038       // Prev is tracking Curr one iteration behind so they refer to different
3039       // underlying objects.
3040       if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3041           isSameUnderlyingObjectInLoop(PN, LI))
3042         for (Value *IncValue : PN->incoming_values())
3043           Worklist.push_back(IncValue);
3044       continue;
3045     }
3046 
3047     Objects.push_back(P);
3048   } while (!Worklist.empty());
3049 }
3050 
3051 /// Return true if the only users of this pointer are lifetime markers.
onlyUsedByLifetimeMarkers(const Value * V)3052 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3053   for (const User *U : V->users()) {
3054     const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3055     if (!II) return false;
3056 
3057     if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3058         II->getIntrinsicID() != Intrinsic::lifetime_end)
3059       return false;
3060   }
3061   return true;
3062 }
3063 
isSafeToSpeculativelyExecute(const Value * V,const Instruction * CtxI,const DominatorTree * DT)3064 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3065                                         const Instruction *CtxI,
3066                                         const DominatorTree *DT) {
3067   const Operator *Inst = dyn_cast<Operator>(V);
3068   if (!Inst)
3069     return false;
3070 
3071   for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3072     if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3073       if (C->canTrap())
3074         return false;
3075 
3076   switch (Inst->getOpcode()) {
3077   default:
3078     return true;
3079   case Instruction::UDiv:
3080   case Instruction::URem: {
3081     // x / y is undefined if y == 0.
3082     const APInt *V;
3083     if (match(Inst->getOperand(1), m_APInt(V)))
3084       return *V != 0;
3085     return false;
3086   }
3087   case Instruction::SDiv:
3088   case Instruction::SRem: {
3089     // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3090     const APInt *Numerator, *Denominator;
3091     if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3092       return false;
3093     // We cannot hoist this division if the denominator is 0.
3094     if (*Denominator == 0)
3095       return false;
3096     // It's safe to hoist if the denominator is not 0 or -1.
3097     if (*Denominator != -1)
3098       return true;
3099     // At this point we know that the denominator is -1.  It is safe to hoist as
3100     // long we know that the numerator is not INT_MIN.
3101     if (match(Inst->getOperand(0), m_APInt(Numerator)))
3102       return !Numerator->isMinSignedValue();
3103     // The numerator *might* be MinSignedValue.
3104     return false;
3105   }
3106   case Instruction::Load: {
3107     const LoadInst *LI = cast<LoadInst>(Inst);
3108     if (!LI->isUnordered() ||
3109         // Speculative load may create a race that did not exist in the source.
3110         LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3111         // Speculative load may load data from dirty regions.
3112         LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress))
3113       return false;
3114     const DataLayout &DL = LI->getModule()->getDataLayout();
3115     return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
3116                                               LI->getAlignment(), DL, CtxI, DT);
3117   }
3118   case Instruction::Call: {
3119     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3120       switch (II->getIntrinsicID()) {
3121       // These synthetic intrinsics have no side-effects and just mark
3122       // information about their operands.
3123       // FIXME: There are other no-op synthetic instructions that potentially
3124       // should be considered at least *safe* to speculate...
3125       case Intrinsic::dbg_declare:
3126       case Intrinsic::dbg_value:
3127         return true;
3128 
3129       case Intrinsic::bswap:
3130       case Intrinsic::ctlz:
3131       case Intrinsic::ctpop:
3132       case Intrinsic::cttz:
3133       case Intrinsic::objectsize:
3134       case Intrinsic::sadd_with_overflow:
3135       case Intrinsic::smul_with_overflow:
3136       case Intrinsic::ssub_with_overflow:
3137       case Intrinsic::uadd_with_overflow:
3138       case Intrinsic::umul_with_overflow:
3139       case Intrinsic::usub_with_overflow:
3140         return true;
3141       // These intrinsics are defined to have the same behavior as libm
3142       // functions except for setting errno.
3143       case Intrinsic::sqrt:
3144       case Intrinsic::fma:
3145       case Intrinsic::fmuladd:
3146         return true;
3147       // These intrinsics are defined to have the same behavior as libm
3148       // functions, and the corresponding libm functions never set errno.
3149       case Intrinsic::trunc:
3150       case Intrinsic::copysign:
3151       case Intrinsic::fabs:
3152       case Intrinsic::minnum:
3153       case Intrinsic::maxnum:
3154         return true;
3155       // These intrinsics are defined to have the same behavior as libm
3156       // functions, which never overflow when operating on the IEEE754 types
3157       // that we support, and never set errno otherwise.
3158       case Intrinsic::ceil:
3159       case Intrinsic::floor:
3160       case Intrinsic::nearbyint:
3161       case Intrinsic::rint:
3162       case Intrinsic::round:
3163         return true;
3164       // TODO: are convert_{from,to}_fp16 safe?
3165       // TODO: can we list target-specific intrinsics here?
3166       default: break;
3167       }
3168     }
3169     return false; // The called function could have undefined behavior or
3170                   // side-effects, even if marked readnone nounwind.
3171   }
3172   case Instruction::VAArg:
3173   case Instruction::Alloca:
3174   case Instruction::Invoke:
3175   case Instruction::PHI:
3176   case Instruction::Store:
3177   case Instruction::Ret:
3178   case Instruction::Br:
3179   case Instruction::IndirectBr:
3180   case Instruction::Switch:
3181   case Instruction::Unreachable:
3182   case Instruction::Fence:
3183   case Instruction::AtomicRMW:
3184   case Instruction::AtomicCmpXchg:
3185   case Instruction::LandingPad:
3186   case Instruction::Resume:
3187   case Instruction::CatchSwitch:
3188   case Instruction::CatchPad:
3189   case Instruction::CatchRet:
3190   case Instruction::CleanupPad:
3191   case Instruction::CleanupRet:
3192     return false; // Misc instructions which have effects
3193   }
3194 }
3195 
mayBeMemoryDependent(const Instruction & I)3196 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3197   return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3198 }
3199 
3200 /// Return true if we know that the specified value is never null.
isKnownNonNull(const Value * V)3201 bool llvm::isKnownNonNull(const Value *V) {
3202   assert(V->getType()->isPointerTy() && "V must be pointer type");
3203 
3204   // Alloca never returns null, malloc might.
3205   if (isa<AllocaInst>(V)) return true;
3206 
3207   // A byval, inalloca, or nonnull argument is never null.
3208   if (const Argument *A = dyn_cast<Argument>(V))
3209     return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3210 
3211   // A global variable in address space 0 is non null unless extern weak.
3212   // Other address spaces may have null as a valid address for a global,
3213   // so we can't assume anything.
3214   if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3215     return !GV->hasExternalWeakLinkage() &&
3216            GV->getType()->getAddressSpace() == 0;
3217 
3218   // A Load tagged with nonnull metadata is never null.
3219   if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3220     return LI->getMetadata(LLVMContext::MD_nonnull);
3221 
3222   if (auto CS = ImmutableCallSite(V))
3223     if (CS.isReturnNonNull())
3224       return true;
3225 
3226   return false;
3227 }
3228 
isKnownNonNullFromDominatingCondition(const Value * V,const Instruction * CtxI,const DominatorTree * DT)3229 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3230                                                   const Instruction *CtxI,
3231                                                   const DominatorTree *DT) {
3232   assert(V->getType()->isPointerTy() && "V must be pointer type");
3233 
3234   unsigned NumUsesExplored = 0;
3235   for (auto *U : V->users()) {
3236     // Avoid massive lists
3237     if (NumUsesExplored >= DomConditionsMaxUses)
3238       break;
3239     NumUsesExplored++;
3240     // Consider only compare instructions uniquely controlling a branch
3241     CmpInst::Predicate Pred;
3242     if (!match(const_cast<User *>(U),
3243                m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
3244         (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
3245       continue;
3246 
3247     for (auto *CmpU : U->users()) {
3248       if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) {
3249         assert(BI->isConditional() && "uses a comparison!");
3250 
3251         BasicBlock *NonNullSuccessor =
3252             BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
3253         BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3254         if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3255           return true;
3256       } else if (Pred == ICmpInst::ICMP_NE &&
3257                  match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) &&
3258                  DT->dominates(cast<Instruction>(CmpU), CtxI)) {
3259         return true;
3260       }
3261     }
3262   }
3263 
3264   return false;
3265 }
3266 
isKnownNonNullAt(const Value * V,const Instruction * CtxI,const DominatorTree * DT)3267 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3268                             const DominatorTree *DT) {
3269   if (isKnownNonNull(V))
3270     return true;
3271 
3272   return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3273 }
3274 
computeOverflowForUnsignedMul(Value * LHS,Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3275 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3276                                                    const DataLayout &DL,
3277                                                    AssumptionCache *AC,
3278                                                    const Instruction *CxtI,
3279                                                    const DominatorTree *DT) {
3280   // Multiplying n * m significant bits yields a result of n + m significant
3281   // bits. If the total number of significant bits does not exceed the
3282   // result bit width (minus 1), there is no overflow.
3283   // This means if we have enough leading zero bits in the operands
3284   // we can guarantee that the result does not overflow.
3285   // Ref: "Hacker's Delight" by Henry Warren
3286   unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3287   APInt LHSKnownZero(BitWidth, 0);
3288   APInt LHSKnownOne(BitWidth, 0);
3289   APInt RHSKnownZero(BitWidth, 0);
3290   APInt RHSKnownOne(BitWidth, 0);
3291   computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3292                    DT);
3293   computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3294                    DT);
3295   // Note that underestimating the number of zero bits gives a more
3296   // conservative answer.
3297   unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3298                       RHSKnownZero.countLeadingOnes();
3299   // First handle the easy case: if we have enough zero bits there's
3300   // definitely no overflow.
3301   if (ZeroBits >= BitWidth)
3302     return OverflowResult::NeverOverflows;
3303 
3304   // Get the largest possible values for each operand.
3305   APInt LHSMax = ~LHSKnownZero;
3306   APInt RHSMax = ~RHSKnownZero;
3307 
3308   // We know the multiply operation doesn't overflow if the maximum values for
3309   // each operand will not overflow after we multiply them together.
3310   bool MaxOverflow;
3311   LHSMax.umul_ov(RHSMax, MaxOverflow);
3312   if (!MaxOverflow)
3313     return OverflowResult::NeverOverflows;
3314 
3315   // We know it always overflows if multiplying the smallest possible values for
3316   // the operands also results in overflow.
3317   bool MinOverflow;
3318   LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3319   if (MinOverflow)
3320     return OverflowResult::AlwaysOverflows;
3321 
3322   return OverflowResult::MayOverflow;
3323 }
3324 
computeOverflowForUnsignedAdd(Value * LHS,Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3325 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3326                                                    const DataLayout &DL,
3327                                                    AssumptionCache *AC,
3328                                                    const Instruction *CxtI,
3329                                                    const DominatorTree *DT) {
3330   bool LHSKnownNonNegative, LHSKnownNegative;
3331   ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3332                  AC, CxtI, DT);
3333   if (LHSKnownNonNegative || LHSKnownNegative) {
3334     bool RHSKnownNonNegative, RHSKnownNegative;
3335     ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3336                    AC, CxtI, DT);
3337 
3338     if (LHSKnownNegative && RHSKnownNegative) {
3339       // The sign bit is set in both cases: this MUST overflow.
3340       // Create a simple add instruction, and insert it into the struct.
3341       return OverflowResult::AlwaysOverflows;
3342     }
3343 
3344     if (LHSKnownNonNegative && RHSKnownNonNegative) {
3345       // The sign bit is clear in both cases: this CANNOT overflow.
3346       // Create a simple add instruction, and insert it into the struct.
3347       return OverflowResult::NeverOverflows;
3348     }
3349   }
3350 
3351   return OverflowResult::MayOverflow;
3352 }
3353 
computeOverflowForSignedAdd(Value * LHS,Value * RHS,AddOperator * Add,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3354 static OverflowResult computeOverflowForSignedAdd(
3355     Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3356     AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3357   if (Add && Add->hasNoSignedWrap()) {
3358     return OverflowResult::NeverOverflows;
3359   }
3360 
3361   bool LHSKnownNonNegative, LHSKnownNegative;
3362   bool RHSKnownNonNegative, RHSKnownNegative;
3363   ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3364                  AC, CxtI, DT);
3365   ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3366                  AC, CxtI, DT);
3367 
3368   if ((LHSKnownNonNegative && RHSKnownNegative) ||
3369       (LHSKnownNegative && RHSKnownNonNegative)) {
3370     // The sign bits are opposite: this CANNOT overflow.
3371     return OverflowResult::NeverOverflows;
3372   }
3373 
3374   // The remaining code needs Add to be available. Early returns if not so.
3375   if (!Add)
3376     return OverflowResult::MayOverflow;
3377 
3378   // If the sign of Add is the same as at least one of the operands, this add
3379   // CANNOT overflow. This is particularly useful when the sum is
3380   // @llvm.assume'ed non-negative rather than proved so from analyzing its
3381   // operands.
3382   bool LHSOrRHSKnownNonNegative =
3383       (LHSKnownNonNegative || RHSKnownNonNegative);
3384   bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3385   if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3386     bool AddKnownNonNegative, AddKnownNegative;
3387     ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3388                    /*Depth=*/0, AC, CxtI, DT);
3389     if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3390         (AddKnownNegative && LHSOrRHSKnownNegative)) {
3391       return OverflowResult::NeverOverflows;
3392     }
3393   }
3394 
3395   return OverflowResult::MayOverflow;
3396 }
3397 
isOverflowIntrinsicNoWrap(IntrinsicInst * II,DominatorTree & DT)3398 bool llvm::isOverflowIntrinsicNoWrap(IntrinsicInst *II, DominatorTree &DT) {
3399 #ifndef NDEBUG
3400   auto IID = II->getIntrinsicID();
3401   assert((IID == Intrinsic::sadd_with_overflow ||
3402           IID == Intrinsic::uadd_with_overflow ||
3403           IID == Intrinsic::ssub_with_overflow ||
3404           IID == Intrinsic::usub_with_overflow ||
3405           IID == Intrinsic::smul_with_overflow ||
3406           IID == Intrinsic::umul_with_overflow) &&
3407          "Not an overflow intrinsic!");
3408 #endif
3409 
3410   SmallVector<BranchInst *, 2> GuardingBranches;
3411   SmallVector<ExtractValueInst *, 2> Results;
3412 
3413   for (User *U : II->users()) {
3414     if (auto *EVI = dyn_cast<ExtractValueInst>(U)) {
3415       assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
3416 
3417       if (EVI->getIndices()[0] == 0)
3418         Results.push_back(EVI);
3419       else {
3420         assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
3421 
3422         for (auto *U : EVI->users())
3423           if (auto *B = dyn_cast<BranchInst>(U)) {
3424             assert(B->isConditional() && "How else is it using an i1?");
3425             GuardingBranches.push_back(B);
3426           }
3427       }
3428     } else {
3429       // We are using the aggregate directly in a way we don't want to analyze
3430       // here (storing it to a global, say).
3431       return false;
3432     }
3433   }
3434 
3435   auto AllUsesGuardedByBranch = [&](BranchInst *BI) {
3436     BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
3437     if (!NoWrapEdge.isSingleEdge())
3438       return false;
3439 
3440     // Check if all users of the add are provably no-wrap.
3441     for (auto *Result : Results) {
3442       // If the extractvalue itself is not executed on overflow, the we don't
3443       // need to check each use separately, since domination is transitive.
3444       if (DT.dominates(NoWrapEdge, Result->getParent()))
3445         continue;
3446 
3447       for (auto &RU : Result->uses())
3448         if (!DT.dominates(NoWrapEdge, RU))
3449           return false;
3450     }
3451 
3452     return true;
3453   };
3454 
3455   return any_of(GuardingBranches, AllUsesGuardedByBranch);
3456 }
3457 
3458 
computeOverflowForSignedAdd(AddOperator * Add,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3459 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3460                                                  const DataLayout &DL,
3461                                                  AssumptionCache *AC,
3462                                                  const Instruction *CxtI,
3463                                                  const DominatorTree *DT) {
3464   return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3465                                        Add, DL, AC, CxtI, DT);
3466 }
3467 
computeOverflowForSignedAdd(Value * LHS,Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3468 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3469                                                  const DataLayout &DL,
3470                                                  AssumptionCache *AC,
3471                                                  const Instruction *CxtI,
3472                                                  const DominatorTree *DT) {
3473   return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3474 }
3475 
isGuaranteedToTransferExecutionToSuccessor(const Instruction * I)3476 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3477   // A memory operation returns normally if it isn't volatile. A volatile
3478   // operation is allowed to trap.
3479   //
3480   // An atomic operation isn't guaranteed to return in a reasonable amount of
3481   // time because it's possible for another thread to interfere with it for an
3482   // arbitrary length of time, but programs aren't allowed to rely on that.
3483   if (const LoadInst *LI = dyn_cast<LoadInst>(I))
3484     return !LI->isVolatile();
3485   if (const StoreInst *SI = dyn_cast<StoreInst>(I))
3486     return !SI->isVolatile();
3487   if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
3488     return !CXI->isVolatile();
3489   if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
3490     return !RMWI->isVolatile();
3491   if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
3492     return !MII->isVolatile();
3493 
3494   // If there is no successor, then execution can't transfer to it.
3495   if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
3496     return !CRI->unwindsToCaller();
3497   if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
3498     return !CatchSwitch->unwindsToCaller();
3499   if (isa<ResumeInst>(I))
3500     return false;
3501   if (isa<ReturnInst>(I))
3502     return false;
3503 
3504   // Calls can throw, or contain an infinite loop, or kill the process.
3505   if (CallSite CS = CallSite(const_cast<Instruction*>(I))) {
3506     // Calls which don't write to arbitrary memory are safe.
3507     // FIXME: Ignoring infinite loops without any side-effects is too aggressive,
3508     // but it's consistent with other passes. See http://llvm.org/PR965 .
3509     // FIXME: This isn't aggressive enough; a call which only writes to a
3510     // global is guaranteed to return.
3511     return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
3512            match(I, m_Intrinsic<Intrinsic::assume>());
3513   }
3514 
3515   // Other instructions return normally.
3516   return true;
3517 }
3518 
isGuaranteedToExecuteForEveryIteration(const Instruction * I,const Loop * L)3519 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3520                                                   const Loop *L) {
3521   // The loop header is guaranteed to be executed for every iteration.
3522   //
3523   // FIXME: Relax this constraint to cover all basic blocks that are
3524   // guaranteed to be executed at every iteration.
3525   if (I->getParent() != L->getHeader()) return false;
3526 
3527   for (const Instruction &LI : *L->getHeader()) {
3528     if (&LI == I) return true;
3529     if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3530   }
3531   llvm_unreachable("Instruction not contained in its own parent basic block.");
3532 }
3533 
propagatesFullPoison(const Instruction * I)3534 bool llvm::propagatesFullPoison(const Instruction *I) {
3535   switch (I->getOpcode()) {
3536     case Instruction::Add:
3537     case Instruction::Sub:
3538     case Instruction::Xor:
3539     case Instruction::Trunc:
3540     case Instruction::BitCast:
3541     case Instruction::AddrSpaceCast:
3542       // These operations all propagate poison unconditionally. Note that poison
3543       // is not any particular value, so xor or subtraction of poison with
3544       // itself still yields poison, not zero.
3545       return true;
3546 
3547     case Instruction::AShr:
3548     case Instruction::SExt:
3549       // For these operations, one bit of the input is replicated across
3550       // multiple output bits. A replicated poison bit is still poison.
3551       return true;
3552 
3553     case Instruction::Shl: {
3554       // Left shift *by* a poison value is poison. The number of
3555       // positions to shift is unsigned, so no negative values are
3556       // possible there. Left shift by zero places preserves poison. So
3557       // it only remains to consider left shift of poison by a positive
3558       // number of places.
3559       //
3560       // A left shift by a positive number of places leaves the lowest order bit
3561       // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3562       // make the poison operand violate that flag, yielding a fresh full-poison
3563       // value.
3564       auto *OBO = cast<OverflowingBinaryOperator>(I);
3565       return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3566     }
3567 
3568     case Instruction::Mul: {
3569       // A multiplication by zero yields a non-poison zero result, so we need to
3570       // rule out zero as an operand. Conservatively, multiplication by a
3571       // non-zero constant is not multiplication by zero.
3572       //
3573       // Multiplication by a non-zero constant can leave some bits
3574       // non-poisoned. For example, a multiplication by 2 leaves the lowest
3575       // order bit unpoisoned. So we need to consider that.
3576       //
3577       // Multiplication by 1 preserves poison. If the multiplication has a
3578       // no-wrap flag, then we can make the poison operand violate that flag
3579       // when multiplied by any integer other than 0 and 1.
3580       auto *OBO = cast<OverflowingBinaryOperator>(I);
3581       if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3582         for (Value *V : OBO->operands()) {
3583           if (auto *CI = dyn_cast<ConstantInt>(V)) {
3584             // A ConstantInt cannot yield poison, so we can assume that it is
3585             // the other operand that is poison.
3586             return !CI->isZero();
3587           }
3588         }
3589       }
3590       return false;
3591     }
3592 
3593     case Instruction::ICmp:
3594       // Comparing poison with any value yields poison.  This is why, for
3595       // instance, x s< (x +nsw 1) can be folded to true.
3596       return true;
3597 
3598     case Instruction::GetElementPtr:
3599       // A GEP implicitly represents a sequence of additions, subtractions,
3600       // truncations, sign extensions and multiplications. The multiplications
3601       // are by the non-zero sizes of some set of types, so we do not have to be
3602       // concerned with multiplication by zero. If the GEP is in-bounds, then
3603       // these operations are implicitly no-signed-wrap so poison is propagated
3604       // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3605       return cast<GEPOperator>(I)->isInBounds();
3606 
3607     default:
3608       return false;
3609   }
3610 }
3611 
getGuaranteedNonFullPoisonOp(const Instruction * I)3612 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3613   switch (I->getOpcode()) {
3614     case Instruction::Store:
3615       return cast<StoreInst>(I)->getPointerOperand();
3616 
3617     case Instruction::Load:
3618       return cast<LoadInst>(I)->getPointerOperand();
3619 
3620     case Instruction::AtomicCmpXchg:
3621       return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3622 
3623     case Instruction::AtomicRMW:
3624       return cast<AtomicRMWInst>(I)->getPointerOperand();
3625 
3626     case Instruction::UDiv:
3627     case Instruction::SDiv:
3628     case Instruction::URem:
3629     case Instruction::SRem:
3630       return I->getOperand(1);
3631 
3632     default:
3633       return nullptr;
3634   }
3635 }
3636 
isKnownNotFullPoison(const Instruction * PoisonI)3637 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3638   // We currently only look for uses of poison values within the same basic
3639   // block, as that makes it easier to guarantee that the uses will be
3640   // executed given that PoisonI is executed.
3641   //
3642   // FIXME: Expand this to consider uses beyond the same basic block. To do
3643   // this, look out for the distinction between post-dominance and strong
3644   // post-dominance.
3645   const BasicBlock *BB = PoisonI->getParent();
3646 
3647   // Set of instructions that we have proved will yield poison if PoisonI
3648   // does.
3649   SmallSet<const Value *, 16> YieldsPoison;
3650   SmallSet<const BasicBlock *, 4> Visited;
3651   YieldsPoison.insert(PoisonI);
3652   Visited.insert(PoisonI->getParent());
3653 
3654   BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
3655 
3656   unsigned Iter = 0;
3657   while (Iter++ < MaxDepth) {
3658     for (auto &I : make_range(Begin, End)) {
3659       if (&I != PoisonI) {
3660         const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
3661         if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
3662           return true;
3663         if (!isGuaranteedToTransferExecutionToSuccessor(&I))
3664           return false;
3665       }
3666 
3667       // Mark poison that propagates from I through uses of I.
3668       if (YieldsPoison.count(&I)) {
3669         for (const User *User : I.users()) {
3670           const Instruction *UserI = cast<Instruction>(User);
3671           if (propagatesFullPoison(UserI))
3672             YieldsPoison.insert(User);
3673         }
3674       }
3675     }
3676 
3677     if (auto *NextBB = BB->getSingleSuccessor()) {
3678       if (Visited.insert(NextBB).second) {
3679         BB = NextBB;
3680         Begin = BB->getFirstNonPHI()->getIterator();
3681         End = BB->end();
3682         continue;
3683       }
3684     }
3685 
3686     break;
3687   };
3688   return false;
3689 }
3690 
isKnownNonNaN(Value * V,FastMathFlags FMF)3691 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3692   if (FMF.noNaNs())
3693     return true;
3694 
3695   if (auto *C = dyn_cast<ConstantFP>(V))
3696     return !C->isNaN();
3697   return false;
3698 }
3699 
isKnownNonZero(Value * V)3700 static bool isKnownNonZero(Value *V) {
3701   if (auto *C = dyn_cast<ConstantFP>(V))
3702     return !C->isZero();
3703   return false;
3704 }
3705 
matchSelectPattern(CmpInst::Predicate Pred,FastMathFlags FMF,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS)3706 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3707                                               FastMathFlags FMF,
3708                                               Value *CmpLHS, Value *CmpRHS,
3709                                               Value *TrueVal, Value *FalseVal,
3710                                               Value *&LHS, Value *&RHS) {
3711   LHS = CmpLHS;
3712   RHS = CmpRHS;
3713 
3714   // If the predicate is an "or-equal"  (FP) predicate, then signed zeroes may
3715   // return inconsistent results between implementations.
3716   //   (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3717   //   minNum(0.0, -0.0)          // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3718   // Therefore we behave conservatively and only proceed if at least one of the
3719   // operands is known to not be zero, or if we don't care about signed zeroes.
3720   switch (Pred) {
3721   default: break;
3722   case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3723   case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3724     if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3725         !isKnownNonZero(CmpRHS))
3726       return {SPF_UNKNOWN, SPNB_NA, false};
3727   }
3728 
3729   SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3730   bool Ordered = false;
3731 
3732   // When given one NaN and one non-NaN input:
3733   //   - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3734   //   - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3735   //     ordered comparison fails), which could be NaN or non-NaN.
3736   // so here we discover exactly what NaN behavior is required/accepted.
3737   if (CmpInst::isFPPredicate(Pred)) {
3738     bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3739     bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3740 
3741     if (LHSSafe && RHSSafe) {
3742       // Both operands are known non-NaN.
3743       NaNBehavior = SPNB_RETURNS_ANY;
3744     } else if (CmpInst::isOrdered(Pred)) {
3745       // An ordered comparison will return false when given a NaN, so it
3746       // returns the RHS.
3747       Ordered = true;
3748       if (LHSSafe)
3749         // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3750         NaNBehavior = SPNB_RETURNS_NAN;
3751       else if (RHSSafe)
3752         NaNBehavior = SPNB_RETURNS_OTHER;
3753       else
3754         // Completely unsafe.
3755         return {SPF_UNKNOWN, SPNB_NA, false};
3756     } else {
3757       Ordered = false;
3758       // An unordered comparison will return true when given a NaN, so it
3759       // returns the LHS.
3760       if (LHSSafe)
3761         // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3762         NaNBehavior = SPNB_RETURNS_OTHER;
3763       else if (RHSSafe)
3764         NaNBehavior = SPNB_RETURNS_NAN;
3765       else
3766         // Completely unsafe.
3767         return {SPF_UNKNOWN, SPNB_NA, false};
3768     }
3769   }
3770 
3771   if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3772     std::swap(CmpLHS, CmpRHS);
3773     Pred = CmpInst::getSwappedPredicate(Pred);
3774     if (NaNBehavior == SPNB_RETURNS_NAN)
3775       NaNBehavior = SPNB_RETURNS_OTHER;
3776     else if (NaNBehavior == SPNB_RETURNS_OTHER)
3777       NaNBehavior = SPNB_RETURNS_NAN;
3778     Ordered = !Ordered;
3779   }
3780 
3781   // ([if]cmp X, Y) ? X : Y
3782   if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3783     switch (Pred) {
3784     default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3785     case ICmpInst::ICMP_UGT:
3786     case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3787     case ICmpInst::ICMP_SGT:
3788     case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3789     case ICmpInst::ICMP_ULT:
3790     case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3791     case ICmpInst::ICMP_SLT:
3792     case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3793     case FCmpInst::FCMP_UGT:
3794     case FCmpInst::FCMP_UGE:
3795     case FCmpInst::FCMP_OGT:
3796     case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3797     case FCmpInst::FCMP_ULT:
3798     case FCmpInst::FCMP_ULE:
3799     case FCmpInst::FCMP_OLT:
3800     case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3801     }
3802   }
3803 
3804   if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3805     if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3806         (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3807 
3808       // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3809       // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3810       if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3811         return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3812       }
3813 
3814       // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3815       // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3816       if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3817         return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3818       }
3819     }
3820 
3821     // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3822     if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3823       if (Pred == ICmpInst::ICMP_SGT && C1->getType() == C2->getType() &&
3824           ~C1->getValue() == C2->getValue() &&
3825           (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3826            match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3827         LHS = TrueVal;
3828         RHS = FalseVal;
3829         return {SPF_SMIN, SPNB_NA, false};
3830       }
3831     }
3832   }
3833 
3834   // TODO: (X > 4) ? X : 5   -->  (X >= 5) ? X : 5  -->  MAX(X, 5)
3835 
3836   return {SPF_UNKNOWN, SPNB_NA, false};
3837 }
3838 
lookThroughCast(CmpInst * CmpI,Value * V1,Value * V2,Instruction::CastOps * CastOp)3839 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
3840                               Instruction::CastOps *CastOp) {
3841   CastInst *CI = dyn_cast<CastInst>(V1);
3842   Constant *C = dyn_cast<Constant>(V2);
3843   if (!CI)
3844     return nullptr;
3845   *CastOp = CI->getOpcode();
3846 
3847   if (auto *CI2 = dyn_cast<CastInst>(V2)) {
3848     // If V1 and V2 are both the same cast from the same type, we can look
3849     // through V1.
3850     if (CI2->getOpcode() == CI->getOpcode() &&
3851         CI2->getSrcTy() == CI->getSrcTy())
3852       return CI2->getOperand(0);
3853     return nullptr;
3854   } else if (!C) {
3855     return nullptr;
3856   }
3857 
3858   Constant *CastedTo = nullptr;
3859 
3860   if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3861     CastedTo = ConstantExpr::getTrunc(C, CI->getSrcTy());
3862 
3863   if (isa<SExtInst>(CI) && CmpI->isSigned())
3864     CastedTo = ConstantExpr::getTrunc(C, CI->getSrcTy(), true);
3865 
3866   if (isa<TruncInst>(CI))
3867     CastedTo = ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3868 
3869   if (isa<FPTruncInst>(CI))
3870     CastedTo = ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
3871 
3872   if (isa<FPExtInst>(CI))
3873     CastedTo = ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
3874 
3875   if (isa<FPToUIInst>(CI))
3876     CastedTo = ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
3877 
3878   if (isa<FPToSIInst>(CI))
3879     CastedTo = ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
3880 
3881   if (isa<UIToFPInst>(CI))
3882     CastedTo = ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
3883 
3884   if (isa<SIToFPInst>(CI))
3885     CastedTo = ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
3886 
3887   if (!CastedTo)
3888     return nullptr;
3889 
3890   Constant *CastedBack =
3891       ConstantExpr::getCast(CI->getOpcode(), CastedTo, C->getType(), true);
3892   // Make sure the cast doesn't lose any information.
3893   if (CastedBack != C)
3894     return nullptr;
3895 
3896   return CastedTo;
3897 }
3898 
matchSelectPattern(Value * V,Value * & LHS,Value * & RHS,Instruction::CastOps * CastOp)3899 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
3900                                              Instruction::CastOps *CastOp) {
3901   SelectInst *SI = dyn_cast<SelectInst>(V);
3902   if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
3903 
3904   CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
3905   if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
3906 
3907   CmpInst::Predicate Pred = CmpI->getPredicate();
3908   Value *CmpLHS = CmpI->getOperand(0);
3909   Value *CmpRHS = CmpI->getOperand(1);
3910   Value *TrueVal = SI->getTrueValue();
3911   Value *FalseVal = SI->getFalseValue();
3912   FastMathFlags FMF;
3913   if (isa<FPMathOperator>(CmpI))
3914     FMF = CmpI->getFastMathFlags();
3915 
3916   // Bail out early.
3917   if (CmpI->isEquality())
3918     return {SPF_UNKNOWN, SPNB_NA, false};
3919 
3920   // Deal with type mismatches.
3921   if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
3922     if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
3923       return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3924                                   cast<CastInst>(TrueVal)->getOperand(0), C,
3925                                   LHS, RHS);
3926     if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
3927       return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3928                                   C, cast<CastInst>(FalseVal)->getOperand(0),
3929                                   LHS, RHS);
3930   }
3931   return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
3932                               LHS, RHS);
3933 }
3934 
getConstantRangeFromMetadata(MDNode & Ranges)3935 ConstantRange llvm::getConstantRangeFromMetadata(MDNode &Ranges) {
3936   const unsigned NumRanges = Ranges.getNumOperands() / 2;
3937   assert(NumRanges >= 1 && "Must have at least one range!");
3938   assert(Ranges.getNumOperands() % 2 == 0 && "Must be a sequence of pairs");
3939 
3940   auto *FirstLow = mdconst::extract<ConstantInt>(Ranges.getOperand(0));
3941   auto *FirstHigh = mdconst::extract<ConstantInt>(Ranges.getOperand(1));
3942 
3943   ConstantRange CR(FirstLow->getValue(), FirstHigh->getValue());
3944 
3945   for (unsigned i = 1; i < NumRanges; ++i) {
3946     auto *Low = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
3947     auto *High = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
3948 
3949     // Note: unionWith will potentially create a range that contains values not
3950     // contained in any of the original N ranges.
3951     CR = CR.unionWith(ConstantRange(Low->getValue(), High->getValue()));
3952   }
3953 
3954   return CR;
3955 }
3956 
3957 /// Return true if "icmp Pred LHS RHS" is always true.
isTruePredicate(CmpInst::Predicate Pred,Value * LHS,Value * RHS,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3958 static bool isTruePredicate(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
3959                             const DataLayout &DL, unsigned Depth,
3960                             AssumptionCache *AC, const Instruction *CxtI,
3961                             const DominatorTree *DT) {
3962   assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
3963   if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
3964     return true;
3965 
3966   switch (Pred) {
3967   default:
3968     return false;
3969 
3970   case CmpInst::ICMP_SLE: {
3971     const APInt *C;
3972 
3973     // LHS s<= LHS +_{nsw} C   if C >= 0
3974     if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
3975       return !C->isNegative();
3976     return false;
3977   }
3978 
3979   case CmpInst::ICMP_ULE: {
3980     const APInt *C;
3981 
3982     // LHS u<= LHS +_{nuw} C   for any C
3983     if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
3984       return true;
3985 
3986     // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
3987     auto MatchNUWAddsToSameValue = [&](Value *A, Value *B, Value *&X,
3988                                        const APInt *&CA, const APInt *&CB) {
3989       if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
3990           match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
3991         return true;
3992 
3993       // If X & C == 0 then (X | C) == X +_{nuw} C
3994       if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
3995           match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
3996         unsigned BitWidth = CA->getBitWidth();
3997         APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3998         computeKnownBits(X, KnownZero, KnownOne, DL, Depth + 1, AC, CxtI, DT);
3999 
4000         if ((KnownZero & *CA) == *CA && (KnownZero & *CB) == *CB)
4001           return true;
4002       }
4003 
4004       return false;
4005     };
4006 
4007     Value *X;
4008     const APInt *CLHS, *CRHS;
4009     if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
4010       return CLHS->ule(*CRHS);
4011 
4012     return false;
4013   }
4014   }
4015 }
4016 
4017 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
4018 /// ALHS ARHS" is true.  Otherwise, return None.
4019 static Optional<bool>
isImpliedCondOperands(CmpInst::Predicate Pred,Value * ALHS,Value * ARHS,Value * BLHS,Value * BRHS,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)4020 isImpliedCondOperands(CmpInst::Predicate Pred, Value *ALHS, Value *ARHS,
4021                       Value *BLHS, Value *BRHS, const DataLayout &DL,
4022                       unsigned Depth, AssumptionCache *AC,
4023                       const Instruction *CxtI, const DominatorTree *DT) {
4024   switch (Pred) {
4025   default:
4026     return None;
4027 
4028   case CmpInst::ICMP_SLT:
4029   case CmpInst::ICMP_SLE:
4030     if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth, AC, CxtI,
4031                         DT) &&
4032         isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4033       return true;
4034     return None;
4035 
4036   case CmpInst::ICMP_ULT:
4037   case CmpInst::ICMP_ULE:
4038     if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth, AC, CxtI,
4039                         DT) &&
4040         isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4041       return true;
4042     return None;
4043   }
4044 }
4045 
4046 /// Return true if the operands of the two compares match.  IsSwappedOps is true
4047 /// when the operands match, but are swapped.
isMatchingOps(Value * ALHS,Value * ARHS,Value * BLHS,Value * BRHS,bool & IsSwappedOps)4048 static bool isMatchingOps(Value *ALHS, Value *ARHS, Value *BLHS, Value *BRHS,
4049                           bool &IsSwappedOps) {
4050 
4051   bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
4052   IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
4053   return IsMatchingOps || IsSwappedOps;
4054 }
4055 
4056 /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is
4057 /// true.  Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS
4058 /// BRHS" is false.  Otherwise, return None if we can't infer anything.
isImpliedCondMatchingOperands(CmpInst::Predicate APred,Value * ALHS,Value * ARHS,CmpInst::Predicate BPred,Value * BLHS,Value * BRHS,bool IsSwappedOps)4059 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
4060                                                     Value *ALHS, Value *ARHS,
4061                                                     CmpInst::Predicate BPred,
4062                                                     Value *BLHS, Value *BRHS,
4063                                                     bool IsSwappedOps) {
4064   // Canonicalize the operands so they're matching.
4065   if (IsSwappedOps) {
4066     std::swap(BLHS, BRHS);
4067     BPred = ICmpInst::getSwappedPredicate(BPred);
4068   }
4069   if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
4070     return true;
4071   if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
4072     return false;
4073 
4074   return None;
4075 }
4076 
4077 /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is
4078 /// true.  Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS
4079 /// C2" is false.  Otherwise, return None if we can't infer anything.
4080 static Optional<bool>
isImpliedCondMatchingImmOperands(CmpInst::Predicate APred,Value * ALHS,ConstantInt * C1,CmpInst::Predicate BPred,Value * BLHS,ConstantInt * C2)4081 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, Value *ALHS,
4082                                  ConstantInt *C1, CmpInst::Predicate BPred,
4083                                  Value *BLHS, ConstantInt *C2) {
4084   assert(ALHS == BLHS && "LHS operands must match.");
4085   ConstantRange DomCR =
4086       ConstantRange::makeExactICmpRegion(APred, C1->getValue());
4087   ConstantRange CR =
4088       ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
4089   ConstantRange Intersection = DomCR.intersectWith(CR);
4090   ConstantRange Difference = DomCR.difference(CR);
4091   if (Intersection.isEmptySet())
4092     return false;
4093   if (Difference.isEmptySet())
4094     return true;
4095   return None;
4096 }
4097 
isImpliedCondition(Value * LHS,Value * RHS,const DataLayout & DL,bool InvertAPred,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)4098 Optional<bool> llvm::isImpliedCondition(Value *LHS, Value *RHS,
4099                                         const DataLayout &DL, bool InvertAPred,
4100                                         unsigned Depth, AssumptionCache *AC,
4101                                         const Instruction *CxtI,
4102                                         const DominatorTree *DT) {
4103   // A mismatch occurs when we compare a scalar cmp to a vector cmp, for example.
4104   if (LHS->getType() != RHS->getType())
4105     return None;
4106 
4107   Type *OpTy = LHS->getType();
4108   assert(OpTy->getScalarType()->isIntegerTy(1));
4109 
4110   // LHS ==> RHS by definition
4111   if (!InvertAPred && LHS == RHS)
4112     return true;
4113 
4114   if (OpTy->isVectorTy())
4115     // TODO: extending the code below to handle vectors
4116     return None;
4117   assert(OpTy->isIntegerTy(1) && "implied by above");
4118 
4119   ICmpInst::Predicate APred, BPred;
4120   Value *ALHS, *ARHS;
4121   Value *BLHS, *BRHS;
4122 
4123   if (!match(LHS, m_ICmp(APred, m_Value(ALHS), m_Value(ARHS))) ||
4124       !match(RHS, m_ICmp(BPred, m_Value(BLHS), m_Value(BRHS))))
4125     return None;
4126 
4127   if (InvertAPred)
4128     APred = CmpInst::getInversePredicate(APred);
4129 
4130   // Can we infer anything when the two compares have matching operands?
4131   bool IsSwappedOps;
4132   if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) {
4133     if (Optional<bool> Implication = isImpliedCondMatchingOperands(
4134             APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps))
4135       return Implication;
4136     // No amount of additional analysis will infer the second condition, so
4137     // early exit.
4138     return None;
4139   }
4140 
4141   // Can we infer anything when the LHS operands match and the RHS operands are
4142   // constants (not necessarily matching)?
4143   if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
4144     if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
4145             APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS,
4146             cast<ConstantInt>(BRHS)))
4147       return Implication;
4148     // No amount of additional analysis will infer the second condition, so
4149     // early exit.
4150     return None;
4151   }
4152 
4153   if (APred == BPred)
4154     return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth, AC,
4155                                  CxtI, DT);
4156 
4157   return None;
4158 }
4159