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/SmallPtrSet.h"
17 #include "llvm/Analysis/AssumptionCache.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/MemoryBuiltins.h"
20 #include "llvm/IR/CallSite.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/MathExtras.h"
36 #include <cstring>
37 using namespace llvm;
38 using namespace llvm::PatternMatch;
39 
40 const unsigned MaxDepth = 6;
41 
42 /// Enable an experimental feature to leverage information about dominating
43 /// conditions to compute known bits.  The individual options below control how
44 /// hard we search.  The defaults are choosen to be fairly aggressive.  If you
45 /// run into compile time problems when testing, scale them back and report
46 /// your findings.
47 static cl::opt<bool> EnableDomConditions("value-tracking-dom-conditions",
48                                          cl::Hidden, cl::init(false));
49 
50 // This is expensive, so we only do it for the top level query value.
51 // (TODO: evaluate cost vs profit, consider higher thresholds)
52 static cl::opt<unsigned> DomConditionsMaxDepth("dom-conditions-max-depth",
53                                                cl::Hidden, cl::init(1));
54 
55 /// How many dominating blocks should be scanned looking for dominating
56 /// conditions?
57 static cl::opt<unsigned> DomConditionsMaxDomBlocks("dom-conditions-dom-blocks",
58                                                    cl::Hidden,
59                                                    cl::init(20000));
60 
61 // Controls the number of uses of the value searched for possible
62 // dominating comparisons.
63 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
64                                               cl::Hidden, cl::init(2000));
65 
66 // If true, don't consider only compares whose only use is a branch.
67 static cl::opt<bool> DomConditionsSingleCmpUse("dom-conditions-single-cmp-use",
68                                                cl::Hidden, cl::init(false));
69 
70 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
71 /// 0). For vector types, returns the element type's bitwidth.
getBitWidth(Type * Ty,const DataLayout & DL)72 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
73   if (unsigned BitWidth = Ty->getScalarSizeInBits())
74     return BitWidth;
75 
76   return DL.getPointerTypeSizeInBits(Ty);
77 }
78 
79 // Many of these functions have internal versions that take an assumption
80 // exclusion set. This is because of the potential for mutual recursion to
81 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
82 // classic case of this is assume(x = y), which will attempt to determine
83 // bits in x from bits in y, which will attempt to determine bits in y from
84 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
85 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
86 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
87 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
88 
89 namespace {
90 // Simplifying using an assume can only be done in a particular control-flow
91 // context (the context instruction provides that context). If an assume and
92 // the context instruction are not in the same block then the DT helps in
93 // figuring out if we can use it.
94 struct Query {
95   ExclInvsSet ExclInvs;
96   AssumptionCache *AC;
97   const Instruction *CxtI;
98   const DominatorTree *DT;
99 
Query__anon17f20b950111::Query100   Query(AssumptionCache *AC = nullptr, const Instruction *CxtI = nullptr,
101         const DominatorTree *DT = nullptr)
102       : AC(AC), CxtI(CxtI), DT(DT) {}
103 
Query__anon17f20b950111::Query104   Query(const Query &Q, const Value *NewExcl)
105       : ExclInvs(Q.ExclInvs), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT) {
106     ExclInvs.insert(NewExcl);
107   }
108 };
109 } // end anonymous namespace
110 
111 // Given the provided Value and, potentially, a context instruction, return
112 // the preferred context instruction (if any).
safeCxtI(const Value * V,const Instruction * CxtI)113 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
114   // If we've been provided with a context instruction, then use that (provided
115   // it has been inserted).
116   if (CxtI && CxtI->getParent())
117     return CxtI;
118 
119   // If the value is really an already-inserted instruction, then use that.
120   CxtI = dyn_cast<Instruction>(V);
121   if (CxtI && CxtI->getParent())
122     return CxtI;
123 
124   return nullptr;
125 }
126 
127 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
128                              const DataLayout &DL, unsigned Depth,
129                              const Query &Q);
130 
computeKnownBits(Value * V,APInt & KnownZero,APInt & KnownOne,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)131 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
132                             const DataLayout &DL, unsigned Depth,
133                             AssumptionCache *AC, const Instruction *CxtI,
134                             const DominatorTree *DT) {
135   ::computeKnownBits(V, KnownZero, KnownOne, DL, Depth,
136                      Query(AC, safeCxtI(V, CxtI), DT));
137 }
138 
139 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
140                            const DataLayout &DL, unsigned Depth,
141                            const Query &Q);
142 
ComputeSignBit(Value * V,bool & KnownZero,bool & KnownOne,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)143 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
144                           const DataLayout &DL, unsigned Depth,
145                           AssumptionCache *AC, const Instruction *CxtI,
146                           const DominatorTree *DT) {
147   ::ComputeSignBit(V, KnownZero, KnownOne, DL, Depth,
148                    Query(AC, safeCxtI(V, CxtI), DT));
149 }
150 
151 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
152                                    const Query &Q, const DataLayout &DL);
153 
isKnownToBeAPowerOfTwo(Value * V,const DataLayout & DL,bool OrZero,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)154 bool llvm::isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, bool OrZero,
155                                   unsigned Depth, AssumptionCache *AC,
156                                   const Instruction *CxtI,
157                                   const DominatorTree *DT) {
158   return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
159                                   Query(AC, safeCxtI(V, CxtI), DT), DL);
160 }
161 
162 static bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
163                            const Query &Q);
164 
isKnownNonZero(Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)165 bool llvm::isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
166                           AssumptionCache *AC, const Instruction *CxtI,
167                           const DominatorTree *DT) {
168   return ::isKnownNonZero(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
169 }
170 
171 static bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
172                               unsigned Depth, const Query &Q);
173 
MaskedValueIsZero(Value * V,const APInt & Mask,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)174 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
175                              unsigned Depth, AssumptionCache *AC,
176                              const Instruction *CxtI, const DominatorTree *DT) {
177   return ::MaskedValueIsZero(V, Mask, DL, Depth,
178                              Query(AC, safeCxtI(V, CxtI), DT));
179 }
180 
181 static unsigned ComputeNumSignBits(Value *V, const DataLayout &DL,
182                                    unsigned Depth, const Query &Q);
183 
ComputeNumSignBits(Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)184 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL,
185                                   unsigned Depth, AssumptionCache *AC,
186                                   const Instruction *CxtI,
187                                   const DominatorTree *DT) {
188   return ::ComputeNumSignBits(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
189 }
190 
computeKnownBitsAddSub(bool Add,Value * Op0,Value * Op1,bool NSW,APInt & KnownZero,APInt & KnownOne,APInt & KnownZero2,APInt & KnownOne2,const DataLayout & DL,unsigned Depth,const Query & Q)191 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
192                                    APInt &KnownZero, APInt &KnownOne,
193                                    APInt &KnownZero2, APInt &KnownOne2,
194                                    const DataLayout &DL, unsigned Depth,
195                                    const Query &Q) {
196   if (!Add) {
197     if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
198       // We know that the top bits of C-X are clear if X contains less bits
199       // than C (i.e. no wrap-around can happen).  For example, 20-X is
200       // positive if we can prove that X is >= 0 and < 16.
201       if (!CLHS->getValue().isNegative()) {
202         unsigned BitWidth = KnownZero.getBitWidth();
203         unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
204         // NLZ can't be BitWidth with no sign bit
205         APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
206         computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
207 
208         // If all of the MaskV bits are known to be zero, then we know the
209         // output top bits are zero, because we now know that the output is
210         // from [0-C].
211         if ((KnownZero2 & MaskV) == MaskV) {
212           unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
213           // Top bits known zero.
214           KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
215         }
216       }
217     }
218   }
219 
220   unsigned BitWidth = KnownZero.getBitWidth();
221 
222   // If an initial sequence of bits in the result is not needed, the
223   // corresponding bits in the operands are not needed.
224   APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
225   computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, DL, Depth + 1, Q);
226   computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
227 
228   // Carry in a 1 for a subtract, rather than a 0.
229   APInt CarryIn(BitWidth, 0);
230   if (!Add) {
231     // Sum = LHS + ~RHS + 1
232     std::swap(KnownZero2, KnownOne2);
233     CarryIn.setBit(0);
234   }
235 
236   APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
237   APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
238 
239   // Compute known bits of the carry.
240   APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
241   APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
242 
243   // Compute set of known bits (where all three relevant bits are known).
244   APInt LHSKnown = LHSKnownZero | LHSKnownOne;
245   APInt RHSKnown = KnownZero2 | KnownOne2;
246   APInt CarryKnown = CarryKnownZero | CarryKnownOne;
247   APInt Known = LHSKnown & RHSKnown & CarryKnown;
248 
249   assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
250          "known bits of sum differ");
251 
252   // Compute known bits of the result.
253   KnownZero = ~PossibleSumOne & Known;
254   KnownOne = PossibleSumOne & Known;
255 
256   // Are we still trying to solve for the sign bit?
257   if (!Known.isNegative()) {
258     if (NSW) {
259       // Adding two non-negative numbers, or subtracting a negative number from
260       // a non-negative one, can't wrap into negative.
261       if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
262         KnownZero |= APInt::getSignBit(BitWidth);
263       // Adding two negative numbers, or subtracting a non-negative number from
264       // a negative one, can't wrap into non-negative.
265       else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
266         KnownOne |= APInt::getSignBit(BitWidth);
267     }
268   }
269 }
270 
computeKnownBitsMul(Value * Op0,Value * Op1,bool NSW,APInt & KnownZero,APInt & KnownOne,APInt & KnownZero2,APInt & KnownOne2,const DataLayout & DL,unsigned Depth,const Query & Q)271 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
272                                 APInt &KnownZero, APInt &KnownOne,
273                                 APInt &KnownZero2, APInt &KnownOne2,
274                                 const DataLayout &DL, unsigned Depth,
275                                 const Query &Q) {
276   unsigned BitWidth = KnownZero.getBitWidth();
277   computeKnownBits(Op1, KnownZero, KnownOne, DL, Depth + 1, Q);
278   computeKnownBits(Op0, KnownZero2, KnownOne2, DL, Depth + 1, Q);
279 
280   bool isKnownNegative = false;
281   bool isKnownNonNegative = false;
282   // If the multiplication is known not to overflow, compute the sign bit.
283   if (NSW) {
284     if (Op0 == Op1) {
285       // The product of a number with itself is non-negative.
286       isKnownNonNegative = true;
287     } else {
288       bool isKnownNonNegativeOp1 = KnownZero.isNegative();
289       bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
290       bool isKnownNegativeOp1 = KnownOne.isNegative();
291       bool isKnownNegativeOp0 = KnownOne2.isNegative();
292       // The product of two numbers with the same sign is non-negative.
293       isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
294         (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
295       // The product of a negative number and a non-negative number is either
296       // negative or zero.
297       if (!isKnownNonNegative)
298         isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
299                            isKnownNonZero(Op0, DL, Depth, Q)) ||
300                           (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
301                            isKnownNonZero(Op1, DL, Depth, Q));
302     }
303   }
304 
305   // If low bits are zero in either operand, output low known-0 bits.
306   // Also compute a conserative estimate for high known-0 bits.
307   // More trickiness is possible, but this is sufficient for the
308   // interesting case of alignment computation.
309   KnownOne.clearAllBits();
310   unsigned TrailZ = KnownZero.countTrailingOnes() +
311                     KnownZero2.countTrailingOnes();
312   unsigned LeadZ =  std::max(KnownZero.countLeadingOnes() +
313                              KnownZero2.countLeadingOnes(),
314                              BitWidth) - BitWidth;
315 
316   TrailZ = std::min(TrailZ, BitWidth);
317   LeadZ = std::min(LeadZ, BitWidth);
318   KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
319               APInt::getHighBitsSet(BitWidth, LeadZ);
320 
321   // Only make use of no-wrap flags if we failed to compute the sign bit
322   // directly.  This matters if the multiplication always overflows, in
323   // which case we prefer to follow the result of the direct computation,
324   // though as the program is invoking undefined behaviour we can choose
325   // whatever we like here.
326   if (isKnownNonNegative && !KnownOne.isNegative())
327     KnownZero.setBit(BitWidth - 1);
328   else if (isKnownNegative && !KnownZero.isNegative())
329     KnownOne.setBit(BitWidth - 1);
330 }
331 
computeKnownBitsFromRangeMetadata(const MDNode & Ranges,APInt & KnownZero)332 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
333                                              APInt &KnownZero) {
334   unsigned BitWidth = KnownZero.getBitWidth();
335   unsigned NumRanges = Ranges.getNumOperands() / 2;
336   assert(NumRanges >= 1);
337 
338   // Use the high end of the ranges to find leading zeros.
339   unsigned MinLeadingZeros = BitWidth;
340   for (unsigned i = 0; i < NumRanges; ++i) {
341     ConstantInt *Lower =
342         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
343     ConstantInt *Upper =
344         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
345     ConstantRange Range(Lower->getValue(), Upper->getValue());
346     if (Range.isWrappedSet())
347       MinLeadingZeros = 0; // -1 has no zeros
348     unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
349     MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
350   }
351 
352   KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
353 }
354 
isEphemeralValueOf(Instruction * I,const Value * E)355 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
356   SmallVector<const Value *, 16> WorkSet(1, I);
357   SmallPtrSet<const Value *, 32> Visited;
358   SmallPtrSet<const Value *, 16> EphValues;
359 
360   while (!WorkSet.empty()) {
361     const Value *V = WorkSet.pop_back_val();
362     if (!Visited.insert(V).second)
363       continue;
364 
365     // If all uses of this value are ephemeral, then so is this value.
366     bool FoundNEUse = false;
367     for (const User *I : V->users())
368       if (!EphValues.count(I)) {
369         FoundNEUse = true;
370         break;
371       }
372 
373     if (!FoundNEUse) {
374       if (V == E)
375         return true;
376 
377       EphValues.insert(V);
378       if (const User *U = dyn_cast<User>(V))
379         for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
380              J != JE; ++J) {
381           if (isSafeToSpeculativelyExecute(*J))
382             WorkSet.push_back(*J);
383         }
384     }
385   }
386 
387   return false;
388 }
389 
390 // Is this an intrinsic that cannot be speculated but also cannot trap?
isAssumeLikeIntrinsic(const Instruction * I)391 static bool isAssumeLikeIntrinsic(const Instruction *I) {
392   if (const CallInst *CI = dyn_cast<CallInst>(I))
393     if (Function *F = CI->getCalledFunction())
394       switch (F->getIntrinsicID()) {
395       default: break;
396       // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
397       case Intrinsic::assume:
398       case Intrinsic::dbg_declare:
399       case Intrinsic::dbg_value:
400       case Intrinsic::invariant_start:
401       case Intrinsic::invariant_end:
402       case Intrinsic::lifetime_start:
403       case Intrinsic::lifetime_end:
404       case Intrinsic::objectsize:
405       case Intrinsic::ptr_annotation:
406       case Intrinsic::var_annotation:
407         return true;
408       }
409 
410   return false;
411 }
412 
isValidAssumeForContext(Value * V,const Query & Q)413 static bool isValidAssumeForContext(Value *V, const Query &Q) {
414   Instruction *Inv = cast<Instruction>(V);
415 
416   // There are two restrictions on the use of an assume:
417   //  1. The assume must dominate the context (or the control flow must
418   //     reach the assume whenever it reaches the context).
419   //  2. The context must not be in the assume's set of ephemeral values
420   //     (otherwise we will use the assume to prove that the condition
421   //     feeding the assume is trivially true, thus causing the removal of
422   //     the assume).
423 
424   if (Q.DT) {
425     if (Q.DT->dominates(Inv, Q.CxtI)) {
426       return true;
427     } else if (Inv->getParent() == Q.CxtI->getParent()) {
428       // The context comes first, but they're both in the same block. Make sure
429       // there is nothing in between that might interrupt the control flow.
430       for (BasicBlock::const_iterator I =
431              std::next(BasicBlock::const_iterator(Q.CxtI)),
432                                       IE(Inv); I != IE; ++I)
433         if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I))
434           return false;
435 
436       return !isEphemeralValueOf(Inv, Q.CxtI);
437     }
438 
439     return false;
440   }
441 
442   // When we don't have a DT, we do a limited search...
443   if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
444     return true;
445   } else if (Inv->getParent() == Q.CxtI->getParent()) {
446     // Search forward from the assume until we reach the context (or the end
447     // of the block); the common case is that the assume will come first.
448     for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
449          IE = Inv->getParent()->end(); I != IE; ++I)
450       if (I == Q.CxtI)
451         return true;
452 
453     // The context must come first...
454     for (BasicBlock::const_iterator I =
455            std::next(BasicBlock::const_iterator(Q.CxtI)),
456                                     IE(Inv); I != IE; ++I)
457       if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I))
458         return false;
459 
460     return !isEphemeralValueOf(Inv, Q.CxtI);
461   }
462 
463   return false;
464 }
465 
isValidAssumeForContext(const Instruction * I,const Instruction * CxtI,const DominatorTree * DT)466 bool llvm::isValidAssumeForContext(const Instruction *I,
467                                    const Instruction *CxtI,
468                                    const DominatorTree *DT) {
469   return ::isValidAssumeForContext(const_cast<Instruction *>(I),
470                                    Query(nullptr, CxtI, DT));
471 }
472 
473 template<typename LHS, typename RHS>
474 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
475                         CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
m_c_ICmp(ICmpInst::Predicate & Pred,const LHS & L,const RHS & R)476 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
477   return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
478 }
479 
480 template<typename LHS, typename RHS>
481 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
482                         BinaryOp_match<RHS, LHS, Instruction::And>>
m_c_And(const LHS & L,const RHS & R)483 m_c_And(const LHS &L, const RHS &R) {
484   return m_CombineOr(m_And(L, R), m_And(R, L));
485 }
486 
487 template<typename LHS, typename RHS>
488 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
489                         BinaryOp_match<RHS, LHS, Instruction::Or>>
m_c_Or(const LHS & L,const RHS & R)490 m_c_Or(const LHS &L, const RHS &R) {
491   return m_CombineOr(m_Or(L, R), m_Or(R, L));
492 }
493 
494 template<typename LHS, typename RHS>
495 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
496                         BinaryOp_match<RHS, LHS, Instruction::Xor>>
m_c_Xor(const LHS & L,const RHS & R)497 m_c_Xor(const LHS &L, const RHS &R) {
498   return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
499 }
500 
501 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
502 /// true (at the context instruction.)  This is mostly a utility function for
503 /// the prototype dominating conditions reasoning below.
computeKnownBitsFromTrueCondition(Value * V,ICmpInst * Cmp,APInt & KnownZero,APInt & KnownOne,const DataLayout & DL,unsigned Depth,const Query & Q)504 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
505                                               APInt &KnownZero,
506                                               APInt &KnownOne,
507                                               const DataLayout &DL,
508                                               unsigned Depth, const Query &Q) {
509   Value *LHS = Cmp->getOperand(0);
510   Value *RHS = Cmp->getOperand(1);
511   // TODO: We could potentially be more aggressive here.  This would be worth
512   // evaluating.  If we can, explore commoning this code with the assume
513   // handling logic.
514   if (LHS != V && RHS != V)
515     return;
516 
517   const unsigned BitWidth = KnownZero.getBitWidth();
518 
519   switch (Cmp->getPredicate()) {
520   default:
521     // We know nothing from this condition
522     break;
523   // TODO: implement unsigned bound from below (known one bits)
524   // TODO: common condition check implementations with assumes
525   // TODO: implement other patterns from assume (e.g. V & B == A)
526   case ICmpInst::ICMP_SGT:
527     if (LHS == V) {
528       APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
529       computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
530       if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
531         // We know that the sign bit is zero.
532         KnownZero |= APInt::getSignBit(BitWidth);
533       }
534     }
535     break;
536   case ICmpInst::ICMP_EQ:
537     if (LHS == V)
538       computeKnownBits(RHS, KnownZero, KnownOne, DL, Depth + 1, Q);
539     else if (RHS == V)
540       computeKnownBits(LHS, KnownZero, KnownOne, DL, Depth + 1, Q);
541     else
542       llvm_unreachable("missing use?");
543     break;
544   case ICmpInst::ICMP_ULE:
545     if (LHS == V) {
546       APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
547       computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
548       // The known zero bits carry over
549       unsigned SignBits = KnownZeroTemp.countLeadingOnes();
550       KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
551     }
552     break;
553   case ICmpInst::ICMP_ULT:
554     if (LHS == V) {
555       APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
556       computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
557       // Whatever high bits in rhs are zero are known to be zero (if rhs is a
558       // power of 2, then one more).
559       unsigned SignBits = KnownZeroTemp.countLeadingOnes();
560       if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
561         SignBits++;
562       KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
563     }
564     break;
565   };
566 }
567 
568 /// Compute known bits in 'V' from conditions which are known to be true along
569 /// all paths leading to the context instruction.  In particular, look for
570 /// cases where one branch of an interesting condition dominates the context
571 /// instruction.  This does not do general dataflow.
572 /// NOTE: This code is EXPERIMENTAL and currently off by default.
computeKnownBitsFromDominatingCondition(Value * V,APInt & KnownZero,APInt & KnownOne,const DataLayout & DL,unsigned Depth,const Query & Q)573 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
574                                                     APInt &KnownOne,
575                                                     const DataLayout &DL,
576                                                     unsigned Depth,
577                                                     const Query &Q) {
578   // Need both the dominator tree and the query location to do anything useful
579   if (!Q.DT || !Q.CxtI)
580     return;
581   Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
582 
583   // Avoid useless work
584   if (auto VI = dyn_cast<Instruction>(V))
585     if (VI->getParent() == Cxt->getParent())
586       return;
587 
588   // Note: We currently implement two options.  It's not clear which of these
589   // will survive long term, we need data for that.
590   // Option 1 - Try walking the dominator tree looking for conditions which
591   // might apply.  This works well for local conditions (loop guards, etc..),
592   // but not as well for things far from the context instruction (presuming a
593   // low max blocks explored).  If we can set an high enough limit, this would
594   // be all we need.
595   // Option 2 - We restrict out search to those conditions which are uses of
596   // the value we're interested in.  This is independent of dom structure,
597   // but is slightly less powerful without looking through lots of use chains.
598   // It does handle conditions far from the context instruction (e.g. early
599   // function exits on entry) really well though.
600 
601   // Option 1 - Search the dom tree
602   unsigned NumBlocksExplored = 0;
603   BasicBlock *Current = Cxt->getParent();
604   while (true) {
605     // Stop searching if we've gone too far up the chain
606     if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
607       break;
608     NumBlocksExplored++;
609 
610     if (!Q.DT->getNode(Current)->getIDom())
611       break;
612     Current = Q.DT->getNode(Current)->getIDom()->getBlock();
613     if (!Current)
614       // found function entry
615       break;
616 
617     BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
618     if (!BI || BI->isUnconditional())
619       continue;
620     ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
621     if (!Cmp)
622       continue;
623 
624     // We're looking for conditions that are guaranteed to hold at the context
625     // instruction.  Finding a condition where one path dominates the context
626     // isn't enough because both the true and false cases could merge before
627     // the context instruction we're actually interested in.  Instead, we need
628     // to ensure that the taken *edge* dominates the context instruction.
629     BasicBlock *BB0 = BI->getSuccessor(0);
630     BasicBlockEdge Edge(BI->getParent(), BB0);
631     if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
632       continue;
633 
634     computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
635                                       Q);
636   }
637 
638   // Option 2 - Search the other uses of V
639   unsigned NumUsesExplored = 0;
640   for (auto U : V->users()) {
641     // Avoid massive lists
642     if (NumUsesExplored >= DomConditionsMaxUses)
643       break;
644     NumUsesExplored++;
645     // Consider only compare instructions uniquely controlling a branch
646     ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
647     if (!Cmp)
648       continue;
649 
650     if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
651       continue;
652 
653     for (auto *CmpU : Cmp->users()) {
654       BranchInst *BI = dyn_cast<BranchInst>(CmpU);
655       if (!BI || BI->isUnconditional())
656         continue;
657       // We're looking for conditions that are guaranteed to hold at the
658       // context instruction.  Finding a condition where one path dominates
659       // the context isn't enough because both the true and false cases could
660       // merge before the context instruction we're actually interested in.
661       // Instead, we need to ensure that the taken *edge* dominates the context
662       // instruction.
663       BasicBlock *BB0 = BI->getSuccessor(0);
664       BasicBlockEdge Edge(BI->getParent(), BB0);
665       if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
666         continue;
667 
668       computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
669                                         Q);
670     }
671   }
672 }
673 
computeKnownBitsFromAssume(Value * V,APInt & KnownZero,APInt & KnownOne,const DataLayout & DL,unsigned Depth,const Query & Q)674 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
675                                        APInt &KnownOne, const DataLayout &DL,
676                                        unsigned Depth, const Query &Q) {
677   // Use of assumptions is context-sensitive. If we don't have a context, we
678   // cannot use them!
679   if (!Q.AC || !Q.CxtI)
680     return;
681 
682   unsigned BitWidth = KnownZero.getBitWidth();
683 
684   for (auto &AssumeVH : Q.AC->assumptions()) {
685     if (!AssumeVH)
686       continue;
687     CallInst *I = cast<CallInst>(AssumeVH);
688     assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
689            "Got assumption for the wrong function!");
690     if (Q.ExclInvs.count(I))
691       continue;
692 
693     // Warning: This loop can end up being somewhat performance sensetive.
694     // We're running this loop for once for each value queried resulting in a
695     // runtime of ~O(#assumes * #values).
696 
697     assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
698            "must be an assume intrinsic");
699 
700     Value *Arg = I->getArgOperand(0);
701 
702     if (Arg == V && isValidAssumeForContext(I, Q)) {
703       assert(BitWidth == 1 && "assume operand is not i1?");
704       KnownZero.clearAllBits();
705       KnownOne.setAllBits();
706       return;
707     }
708 
709     // The remaining tests are all recursive, so bail out if we hit the limit.
710     if (Depth == MaxDepth)
711       continue;
712 
713     Value *A, *B;
714     auto m_V = m_CombineOr(m_Specific(V),
715                            m_CombineOr(m_PtrToInt(m_Specific(V)),
716                            m_BitCast(m_Specific(V))));
717 
718     CmpInst::Predicate Pred;
719     ConstantInt *C;
720     // assume(v = a)
721     if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
722         Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
723       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
724       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
725       KnownZero |= RHSKnownZero;
726       KnownOne  |= RHSKnownOne;
727     // assume(v & b = a)
728     } else if (match(Arg,
729                      m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
730                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
731       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
732       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
733       APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
734       computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
735 
736       // For those bits in the mask that are known to be one, we can propagate
737       // known bits from the RHS to V.
738       KnownZero |= RHSKnownZero & MaskKnownOne;
739       KnownOne  |= RHSKnownOne  & MaskKnownOne;
740     // assume(~(v & b) = a)
741     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
742                                    m_Value(A))) &&
743                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
744       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
745       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
746       APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
747       computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
748 
749       // For those bits in the mask that are known to be one, we can propagate
750       // inverted known bits from the RHS to V.
751       KnownZero |= RHSKnownOne  & MaskKnownOne;
752       KnownOne  |= RHSKnownZero & MaskKnownOne;
753     // assume(v | b = a)
754     } else if (match(Arg,
755                      m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
756                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
757       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
758       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
759       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
760       computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
761 
762       // For those bits in B that are known to be zero, we can propagate known
763       // bits from the RHS to V.
764       KnownZero |= RHSKnownZero & BKnownZero;
765       KnownOne  |= RHSKnownOne  & BKnownZero;
766     // assume(~(v | b) = a)
767     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
768                                    m_Value(A))) &&
769                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
770       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
771       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
772       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
773       computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
774 
775       // For those bits in B that are known to be zero, we can propagate
776       // inverted known bits from the RHS to V.
777       KnownZero |= RHSKnownOne  & BKnownZero;
778       KnownOne  |= RHSKnownZero & BKnownZero;
779     // assume(v ^ b = a)
780     } else if (match(Arg,
781                      m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
782                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
783       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
784       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
785       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
786       computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
787 
788       // For those bits in B that are known to be zero, we can propagate known
789       // bits from the RHS to V. For those bits in B that are known to be one,
790       // we can propagate inverted known bits from the RHS to V.
791       KnownZero |= RHSKnownZero & BKnownZero;
792       KnownOne  |= RHSKnownOne  & BKnownZero;
793       KnownZero |= RHSKnownOne  & BKnownOne;
794       KnownOne  |= RHSKnownZero & BKnownOne;
795     // assume(~(v ^ b) = a)
796     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
797                                    m_Value(A))) &&
798                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
799       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
800       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
801       APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
802       computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
803 
804       // For those bits in B that are known to be zero, we can propagate
805       // inverted known bits from the RHS to V. For those bits in B that are
806       // known to be one, we can propagate known bits from the RHS to V.
807       KnownZero |= RHSKnownOne  & BKnownZero;
808       KnownOne  |= RHSKnownZero & BKnownZero;
809       KnownZero |= RHSKnownZero & BKnownOne;
810       KnownOne  |= RHSKnownOne  & BKnownOne;
811     // assume(v << c = a)
812     } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
813                                    m_Value(A))) &&
814                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
815       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
816       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
817       // For those bits in RHS that are known, we can propagate them to known
818       // bits in V shifted to the right by C.
819       KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
820       KnownOne  |= RHSKnownOne.lshr(C->getZExtValue());
821     // assume(~(v << c) = a)
822     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
823                                    m_Value(A))) &&
824                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
825       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
826       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
827       // For those bits in RHS that are known, we can propagate them inverted
828       // to known bits in V shifted to the right by C.
829       KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
830       KnownOne  |= RHSKnownZero.lshr(C->getZExtValue());
831     // assume(v >> c = a)
832     } else if (match(Arg,
833                      m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
834                                                 m_AShr(m_V, m_ConstantInt(C))),
835                               m_Value(A))) &&
836                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
837       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
838       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
839       // For those bits in RHS that are known, we can propagate them to known
840       // bits in V shifted to the right by C.
841       KnownZero |= RHSKnownZero << C->getZExtValue();
842       KnownOne  |= RHSKnownOne  << C->getZExtValue();
843     // assume(~(v >> c) = a)
844     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
845                                              m_LShr(m_V, m_ConstantInt(C)),
846                                              m_AShr(m_V, m_ConstantInt(C)))),
847                                    m_Value(A))) &&
848                Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
849       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
850       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
851       // For those bits in RHS that are known, we can propagate them inverted
852       // to known bits in V shifted to the right by C.
853       KnownZero |= RHSKnownOne  << C->getZExtValue();
854       KnownOne  |= RHSKnownZero << C->getZExtValue();
855     // assume(v >=_s c) where c is non-negative
856     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
857                Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
858       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
859       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
860 
861       if (RHSKnownZero.isNegative()) {
862         // We know that the sign bit is zero.
863         KnownZero |= APInt::getSignBit(BitWidth);
864       }
865     // assume(v >_s c) where c is at least -1.
866     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
867                Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
868       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
869       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
870 
871       if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
872         // We know that the sign bit is zero.
873         KnownZero |= APInt::getSignBit(BitWidth);
874       }
875     // assume(v <=_s c) where c is negative
876     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
877                Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
878       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
879       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
880 
881       if (RHSKnownOne.isNegative()) {
882         // We know that the sign bit is one.
883         KnownOne |= APInt::getSignBit(BitWidth);
884       }
885     // assume(v <_s c) where c is non-positive
886     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
887                Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
888       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
889       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
890 
891       if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
892         // We know that the sign bit is one.
893         KnownOne |= APInt::getSignBit(BitWidth);
894       }
895     // assume(v <=_u c)
896     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
897                Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
898       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
899       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
900 
901       // Whatever high bits in c are zero are known to be zero.
902       KnownZero |=
903         APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
904     // assume(v <_u c)
905     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
906                Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
907       APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
908       computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
909 
910       // Whatever high bits in c are zero are known to be zero (if c is a power
911       // of 2, then one more).
912       if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
913         KnownZero |=
914           APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
915       else
916         KnownZero |=
917           APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
918     }
919   }
920 }
921 
922 /// Determine which bits of V are known to be either zero or one and return
923 /// them in the KnownZero/KnownOne bit sets.
924 ///
925 /// NOTE: we cannot consider 'undef' to be "IsZero" here.  The problem is that
926 /// we cannot optimize based on the assumption that it is zero without changing
927 /// it to be an explicit zero.  If we don't change it to zero, other code could
928 /// optimized based on the contradictory assumption that it is non-zero.
929 /// Because instcombine aggressively folds operations with undef args anyway,
930 /// this won't lose us code quality.
931 ///
932 /// This function is defined on values with integer type, values with pointer
933 /// type, and vectors of integers.  In the case
934 /// where V is a vector, known zero, and known one values are the
935 /// same width as the vector element, and the bit is set only if it is true
936 /// for all of the elements in the vector.
computeKnownBits(Value * V,APInt & KnownZero,APInt & KnownOne,const DataLayout & DL,unsigned Depth,const Query & Q)937 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
938                       const DataLayout &DL, unsigned Depth, const Query &Q) {
939   assert(V && "No Value?");
940   assert(Depth <= MaxDepth && "Limit Search Depth");
941   unsigned BitWidth = KnownZero.getBitWidth();
942 
943   assert((V->getType()->isIntOrIntVectorTy() ||
944           V->getType()->getScalarType()->isPointerTy()) &&
945          "Not integer or pointer type!");
946   assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
947          (!V->getType()->isIntOrIntVectorTy() ||
948           V->getType()->getScalarSizeInBits() == BitWidth) &&
949          KnownZero.getBitWidth() == BitWidth &&
950          KnownOne.getBitWidth() == BitWidth &&
951          "V, KnownOne and KnownZero should have same BitWidth");
952 
953   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
954     // We know all of the bits for a constant!
955     KnownOne = CI->getValue();
956     KnownZero = ~KnownOne;
957     return;
958   }
959   // Null and aggregate-zero are all-zeros.
960   if (isa<ConstantPointerNull>(V) ||
961       isa<ConstantAggregateZero>(V)) {
962     KnownOne.clearAllBits();
963     KnownZero = APInt::getAllOnesValue(BitWidth);
964     return;
965   }
966   // Handle a constant vector by taking the intersection of the known bits of
967   // each element.  There is no real need to handle ConstantVector here, because
968   // we don't handle undef in any particularly useful way.
969   if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
970     // We know that CDS must be a vector of integers. Take the intersection of
971     // each element.
972     KnownZero.setAllBits(); KnownOne.setAllBits();
973     APInt Elt(KnownZero.getBitWidth(), 0);
974     for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
975       Elt = CDS->getElementAsInteger(i);
976       KnownZero &= ~Elt;
977       KnownOne &= Elt;
978     }
979     return;
980   }
981 
982   // The address of an aligned GlobalValue has trailing zeros.
983   if (auto *GO = dyn_cast<GlobalObject>(V)) {
984     unsigned Align = GO->getAlignment();
985     if (Align == 0) {
986       if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
987         Type *ObjectType = GVar->getType()->getElementType();
988         if (ObjectType->isSized()) {
989           // If the object is defined in the current Module, we'll be giving
990           // it the preferred alignment. Otherwise, we have to assume that it
991           // may only have the minimum ABI alignment.
992           if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
993             Align = DL.getPreferredAlignment(GVar);
994           else
995             Align = DL.getABITypeAlignment(ObjectType);
996         }
997       }
998     }
999     if (Align > 0)
1000       KnownZero = APInt::getLowBitsSet(BitWidth,
1001                                        countTrailingZeros(Align));
1002     else
1003       KnownZero.clearAllBits();
1004     KnownOne.clearAllBits();
1005     return;
1006   }
1007 
1008   if (Argument *A = dyn_cast<Argument>(V)) {
1009     unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1010 
1011     if (!Align && A->hasStructRetAttr()) {
1012       // An sret parameter has at least the ABI alignment of the return type.
1013       Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1014       if (EltTy->isSized())
1015         Align = DL.getABITypeAlignment(EltTy);
1016     }
1017 
1018     if (Align)
1019       KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1020     else
1021       KnownZero.clearAllBits();
1022     KnownOne.clearAllBits();
1023 
1024     // Don't give up yet... there might be an assumption that provides more
1025     // information...
1026     computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1027 
1028     // Or a dominating condition for that matter
1029     if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1030       computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL,
1031                                               Depth, Q);
1032     return;
1033   }
1034 
1035   // Start out not knowing anything.
1036   KnownZero.clearAllBits(); KnownOne.clearAllBits();
1037 
1038   // Limit search depth.
1039   // All recursive calls that increase depth must come after this.
1040   if (Depth == MaxDepth)
1041     return;
1042 
1043   // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1044   // the bits of its aliasee.
1045   if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1046     if (!GA->mayBeOverridden())
1047       computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1048     return;
1049   }
1050 
1051   // Check whether a nearby assume intrinsic can determine some known bits.
1052   computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1053 
1054   // Check whether there's a dominating condition which implies something about
1055   // this value at the given context.
1056   if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1057     computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1058                                             Q);
1059 
1060   Operator *I = dyn_cast<Operator>(V);
1061   if (!I) return;
1062 
1063   APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
1064   switch (I->getOpcode()) {
1065   default: break;
1066   case Instruction::Load:
1067     if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
1068       computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1069     break;
1070   case Instruction::And: {
1071     // If either the LHS or the RHS are Zero, the result is zero.
1072     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1073     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1074 
1075     // Output known-1 bits are only known if set in both the LHS & RHS.
1076     KnownOne &= KnownOne2;
1077     // Output known-0 are known to be clear if zero in either the LHS | RHS.
1078     KnownZero |= KnownZero2;
1079     break;
1080   }
1081   case Instruction::Or: {
1082     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1083     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1084 
1085     // Output known-0 bits are only known if clear in both the LHS & RHS.
1086     KnownZero &= KnownZero2;
1087     // Output known-1 are known to be set if set in either the LHS | RHS.
1088     KnownOne |= KnownOne2;
1089     break;
1090   }
1091   case Instruction::Xor: {
1092     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1093     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1094 
1095     // Output known-0 bits are known if clear or set in both the LHS & RHS.
1096     APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1097     // Output known-1 are known to be set if set in only one of the LHS, RHS.
1098     KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1099     KnownZero = KnownZeroOut;
1100     break;
1101   }
1102   case Instruction::Mul: {
1103     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1104     computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1105                         KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1106     break;
1107   }
1108   case Instruction::UDiv: {
1109     // For the purposes of computing leading zeros we can conservatively
1110     // treat a udiv as a logical right shift by the power of 2 known to
1111     // be less than the denominator.
1112     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1113     unsigned LeadZ = KnownZero2.countLeadingOnes();
1114 
1115     KnownOne2.clearAllBits();
1116     KnownZero2.clearAllBits();
1117     computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1118     unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1119     if (RHSUnknownLeadingOnes != BitWidth)
1120       LeadZ = std::min(BitWidth,
1121                        LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1122 
1123     KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1124     break;
1125   }
1126   case Instruction::Select:
1127     computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1128     computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1129 
1130     // Only known if known in both the LHS and RHS.
1131     KnownOne &= KnownOne2;
1132     KnownZero &= KnownZero2;
1133     break;
1134   case Instruction::FPTrunc:
1135   case Instruction::FPExt:
1136   case Instruction::FPToUI:
1137   case Instruction::FPToSI:
1138   case Instruction::SIToFP:
1139   case Instruction::UIToFP:
1140     break; // Can't work with floating point.
1141   case Instruction::PtrToInt:
1142   case Instruction::IntToPtr:
1143   case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1144     // FALL THROUGH and handle them the same as zext/trunc.
1145   case Instruction::ZExt:
1146   case Instruction::Trunc: {
1147     Type *SrcTy = I->getOperand(0)->getType();
1148 
1149     unsigned SrcBitWidth;
1150     // Note that we handle pointer operands here because of inttoptr/ptrtoint
1151     // which fall through here.
1152     SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1153 
1154     assert(SrcBitWidth && "SrcBitWidth can't be zero");
1155     KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1156     KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1157     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1158     KnownZero = KnownZero.zextOrTrunc(BitWidth);
1159     KnownOne = KnownOne.zextOrTrunc(BitWidth);
1160     // Any top bits are known to be zero.
1161     if (BitWidth > SrcBitWidth)
1162       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1163     break;
1164   }
1165   case Instruction::BitCast: {
1166     Type *SrcTy = I->getOperand(0)->getType();
1167     if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
1168         // TODO: For now, not handling conversions like:
1169         // (bitcast i64 %x to <2 x i32>)
1170         !I->getType()->isVectorTy()) {
1171       computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1172       break;
1173     }
1174     break;
1175   }
1176   case Instruction::SExt: {
1177     // Compute the bits in the result that are not present in the input.
1178     unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1179 
1180     KnownZero = KnownZero.trunc(SrcBitWidth);
1181     KnownOne = KnownOne.trunc(SrcBitWidth);
1182     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1183     KnownZero = KnownZero.zext(BitWidth);
1184     KnownOne = KnownOne.zext(BitWidth);
1185 
1186     // If the sign bit of the input is known set or clear, then we know the
1187     // top bits of the result.
1188     if (KnownZero[SrcBitWidth-1])             // Input sign bit known zero
1189       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1190     else if (KnownOne[SrcBitWidth-1])           // Input sign bit known set
1191       KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1192     break;
1193   }
1194   case Instruction::Shl:
1195     // (shl X, C1) & C2 == 0   iff   (X & C2 >>u C1) == 0
1196     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1197       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1198       computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1199       KnownZero <<= ShiftAmt;
1200       KnownOne  <<= ShiftAmt;
1201       KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1202     }
1203     break;
1204   case Instruction::LShr:
1205     // (ushr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
1206     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1207       // Compute the new bits that are at the top now.
1208       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1209 
1210       // Unsigned shift right.
1211       computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1212       KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1213       KnownOne  = APIntOps::lshr(KnownOne, ShiftAmt);
1214       // high bits known zero.
1215       KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1216     }
1217     break;
1218   case Instruction::AShr:
1219     // (ashr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
1220     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1221       // Compute the new bits that are at the top now.
1222       uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1223 
1224       // Signed shift right.
1225       computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1226       KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1227       KnownOne  = APIntOps::lshr(KnownOne, ShiftAmt);
1228 
1229       APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1230       if (KnownZero[BitWidth-ShiftAmt-1])    // New bits are known zero.
1231         KnownZero |= HighBits;
1232       else if (KnownOne[BitWidth-ShiftAmt-1])  // New bits are known one.
1233         KnownOne |= HighBits;
1234     }
1235     break;
1236   case Instruction::Sub: {
1237     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1238     computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1239                            KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1240                            Depth, Q);
1241     break;
1242   }
1243   case Instruction::Add: {
1244     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1245     computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1246                            KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1247                            Depth, Q);
1248     break;
1249   }
1250   case Instruction::SRem:
1251     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1252       APInt RA = Rem->getValue().abs();
1253       if (RA.isPowerOf2()) {
1254         APInt LowBits = RA - 1;
1255         computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1256                          Q);
1257 
1258         // The low bits of the first operand are unchanged by the srem.
1259         KnownZero = KnownZero2 & LowBits;
1260         KnownOne = KnownOne2 & LowBits;
1261 
1262         // If the first operand is non-negative or has all low bits zero, then
1263         // the upper bits are all zero.
1264         if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1265           KnownZero |= ~LowBits;
1266 
1267         // If the first operand is negative and not all low bits are zero, then
1268         // the upper bits are all one.
1269         if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1270           KnownOne |= ~LowBits;
1271 
1272         assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1273       }
1274     }
1275 
1276     // The sign bit is the LHS's sign bit, except when the result of the
1277     // remainder is zero.
1278     if (KnownZero.isNonNegative()) {
1279       APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1280       computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1281                        Depth + 1, Q);
1282       // If it's known zero, our sign bit is also zero.
1283       if (LHSKnownZero.isNegative())
1284         KnownZero.setBit(BitWidth - 1);
1285     }
1286 
1287     break;
1288   case Instruction::URem: {
1289     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1290       APInt RA = Rem->getValue();
1291       if (RA.isPowerOf2()) {
1292         APInt LowBits = (RA - 1);
1293         computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1294                          Q);
1295         KnownZero |= ~LowBits;
1296         KnownOne &= LowBits;
1297         break;
1298       }
1299     }
1300 
1301     // Since the result is less than or equal to either operand, any leading
1302     // zero bits in either operand must also exist in the result.
1303     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1304     computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1305 
1306     unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1307                                 KnownZero2.countLeadingOnes());
1308     KnownOne.clearAllBits();
1309     KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1310     break;
1311   }
1312 
1313   case Instruction::Alloca: {
1314     AllocaInst *AI = cast<AllocaInst>(V);
1315     unsigned Align = AI->getAlignment();
1316     if (Align == 0)
1317       Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1318 
1319     if (Align > 0)
1320       KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1321     break;
1322   }
1323   case Instruction::GetElementPtr: {
1324     // Analyze all of the subscripts of this getelementptr instruction
1325     // to determine if we can prove known low zero bits.
1326     APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1327     computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1328                      Depth + 1, Q);
1329     unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1330 
1331     gep_type_iterator GTI = gep_type_begin(I);
1332     for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1333       Value *Index = I->getOperand(i);
1334       if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1335         // Handle struct member offset arithmetic.
1336 
1337         // Handle case when index is vector zeroinitializer
1338         Constant *CIndex = cast<Constant>(Index);
1339         if (CIndex->isZeroValue())
1340           continue;
1341 
1342         if (CIndex->getType()->isVectorTy())
1343           Index = CIndex->getSplatValue();
1344 
1345         unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1346         const StructLayout *SL = DL.getStructLayout(STy);
1347         uint64_t Offset = SL->getElementOffset(Idx);
1348         TrailZ = std::min<unsigned>(TrailZ,
1349                                     countTrailingZeros(Offset));
1350       } else {
1351         // Handle array index arithmetic.
1352         Type *IndexedTy = GTI.getIndexedType();
1353         if (!IndexedTy->isSized()) {
1354           TrailZ = 0;
1355           break;
1356         }
1357         unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1358         uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1359         LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1360         computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1361                          Q);
1362         TrailZ = std::min(TrailZ,
1363                           unsigned(countTrailingZeros(TypeSize) +
1364                                    LocalKnownZero.countTrailingOnes()));
1365       }
1366     }
1367 
1368     KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1369     break;
1370   }
1371   case Instruction::PHI: {
1372     PHINode *P = cast<PHINode>(I);
1373     // Handle the case of a simple two-predecessor recurrence PHI.
1374     // There's a lot more that could theoretically be done here, but
1375     // this is sufficient to catch some interesting cases.
1376     if (P->getNumIncomingValues() == 2) {
1377       for (unsigned i = 0; i != 2; ++i) {
1378         Value *L = P->getIncomingValue(i);
1379         Value *R = P->getIncomingValue(!i);
1380         Operator *LU = dyn_cast<Operator>(L);
1381         if (!LU)
1382           continue;
1383         unsigned Opcode = LU->getOpcode();
1384         // Check for operations that have the property that if
1385         // both their operands have low zero bits, the result
1386         // will have low zero bits.
1387         if (Opcode == Instruction::Add ||
1388             Opcode == Instruction::Sub ||
1389             Opcode == Instruction::And ||
1390             Opcode == Instruction::Or ||
1391             Opcode == Instruction::Mul) {
1392           Value *LL = LU->getOperand(0);
1393           Value *LR = LU->getOperand(1);
1394           // Find a recurrence.
1395           if (LL == I)
1396             L = LR;
1397           else if (LR == I)
1398             L = LL;
1399           else
1400             break;
1401           // Ok, we have a PHI of the form L op= R. Check for low
1402           // zero bits.
1403           computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1404 
1405           // We need to take the minimum number of known bits
1406           APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1407           computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1408 
1409           KnownZero = APInt::getLowBitsSet(BitWidth,
1410                                            std::min(KnownZero2.countTrailingOnes(),
1411                                                     KnownZero3.countTrailingOnes()));
1412           break;
1413         }
1414       }
1415     }
1416 
1417     // Unreachable blocks may have zero-operand PHI nodes.
1418     if (P->getNumIncomingValues() == 0)
1419       break;
1420 
1421     // Otherwise take the unions of the known bit sets of the operands,
1422     // taking conservative care to avoid excessive recursion.
1423     if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1424       // Skip if every incoming value references to ourself.
1425       if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1426         break;
1427 
1428       KnownZero = APInt::getAllOnesValue(BitWidth);
1429       KnownOne = APInt::getAllOnesValue(BitWidth);
1430       for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1431         // Skip direct self references.
1432         if (P->getIncomingValue(i) == P) continue;
1433 
1434         KnownZero2 = APInt(BitWidth, 0);
1435         KnownOne2 = APInt(BitWidth, 0);
1436         // Recurse, but cap the recursion to one level, because we don't
1437         // want to waste time spinning around in loops.
1438         computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, DL,
1439                          MaxDepth - 1, Q);
1440         KnownZero &= KnownZero2;
1441         KnownOne &= KnownOne2;
1442         // If all bits have been ruled out, there's no need to check
1443         // more operands.
1444         if (!KnownZero && !KnownOne)
1445           break;
1446       }
1447     }
1448     break;
1449   }
1450   case Instruction::Call:
1451   case Instruction::Invoke:
1452     if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1453       computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1454     // If a range metadata is attached to this IntrinsicInst, intersect the
1455     // explicit range specified by the metadata and the implicit range of
1456     // the intrinsic.
1457     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1458       switch (II->getIntrinsicID()) {
1459       default: break;
1460       case Intrinsic::ctlz:
1461       case Intrinsic::cttz: {
1462         unsigned LowBits = Log2_32(BitWidth)+1;
1463         // If this call is undefined for 0, the result will be less than 2^n.
1464         if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1465           LowBits -= 1;
1466         KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1467         break;
1468       }
1469       case Intrinsic::ctpop: {
1470         unsigned LowBits = Log2_32(BitWidth)+1;
1471         KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1472         break;
1473       }
1474       case Intrinsic::x86_sse42_crc32_64_64:
1475         KnownZero |= APInt::getHighBitsSet(64, 32);
1476         break;
1477       }
1478     }
1479     break;
1480   case Instruction::ExtractValue:
1481     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1482       ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1483       if (EVI->getNumIndices() != 1) break;
1484       if (EVI->getIndices()[0] == 0) {
1485         switch (II->getIntrinsicID()) {
1486         default: break;
1487         case Intrinsic::uadd_with_overflow:
1488         case Intrinsic::sadd_with_overflow:
1489           computeKnownBitsAddSub(true, II->getArgOperand(0),
1490                                  II->getArgOperand(1), false, KnownZero,
1491                                  KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1492           break;
1493         case Intrinsic::usub_with_overflow:
1494         case Intrinsic::ssub_with_overflow:
1495           computeKnownBitsAddSub(false, II->getArgOperand(0),
1496                                  II->getArgOperand(1), false, KnownZero,
1497                                  KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1498           break;
1499         case Intrinsic::umul_with_overflow:
1500         case Intrinsic::smul_with_overflow:
1501           computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1502                               KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1503                               Depth, Q);
1504           break;
1505         }
1506       }
1507     }
1508   }
1509 
1510   assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1511 }
1512 
1513 /// Determine whether the sign bit is known to be zero or one.
1514 /// Convenience wrapper around computeKnownBits.
ComputeSignBit(Value * V,bool & KnownZero,bool & KnownOne,const DataLayout & DL,unsigned Depth,const Query & Q)1515 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1516                     const DataLayout &DL, unsigned Depth, const Query &Q) {
1517   unsigned BitWidth = getBitWidth(V->getType(), DL);
1518   if (!BitWidth) {
1519     KnownZero = false;
1520     KnownOne = false;
1521     return;
1522   }
1523   APInt ZeroBits(BitWidth, 0);
1524   APInt OneBits(BitWidth, 0);
1525   computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1526   KnownOne = OneBits[BitWidth - 1];
1527   KnownZero = ZeroBits[BitWidth - 1];
1528 }
1529 
1530 /// Return true if the given value is known to have exactly one
1531 /// bit set when defined. For vectors return true if every element is known to
1532 /// be a power of two when defined. Supports values with integer or pointer
1533 /// types and vectors of integers.
isKnownToBeAPowerOfTwo(Value * V,bool OrZero,unsigned Depth,const Query & Q,const DataLayout & DL)1534 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1535                             const Query &Q, const DataLayout &DL) {
1536   if (Constant *C = dyn_cast<Constant>(V)) {
1537     if (C->isNullValue())
1538       return OrZero;
1539     if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1540       return CI->getValue().isPowerOf2();
1541     // TODO: Handle vector constants.
1542   }
1543 
1544   // 1 << X is clearly a power of two if the one is not shifted off the end.  If
1545   // it is shifted off the end then the result is undefined.
1546   if (match(V, m_Shl(m_One(), m_Value())))
1547     return true;
1548 
1549   // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1550   // bottom.  If it is shifted off the bottom then the result is undefined.
1551   if (match(V, m_LShr(m_SignBit(), m_Value())))
1552     return true;
1553 
1554   // The remaining tests are all recursive, so bail out if we hit the limit.
1555   if (Depth++ == MaxDepth)
1556     return false;
1557 
1558   Value *X = nullptr, *Y = nullptr;
1559   // A shift of a power of two is a power of two or zero.
1560   if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1561                  match(V, m_Shr(m_Value(X), m_Value()))))
1562     return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1563 
1564   if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1565     return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1566 
1567   if (SelectInst *SI = dyn_cast<SelectInst>(V))
1568     return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1569            isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1570 
1571   if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1572     // A power of two and'd with anything is a power of two or zero.
1573     if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1574         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1575       return true;
1576     // X & (-X) is always a power of two or zero.
1577     if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1578       return true;
1579     return false;
1580   }
1581 
1582   // Adding a power-of-two or zero to the same power-of-two or zero yields
1583   // either the original power-of-two, a larger power-of-two or zero.
1584   if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1585     OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1586     if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1587       if (match(X, m_And(m_Specific(Y), m_Value())) ||
1588           match(X, m_And(m_Value(), m_Specific(Y))))
1589         if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1590           return true;
1591       if (match(Y, m_And(m_Specific(X), m_Value())) ||
1592           match(Y, m_And(m_Value(), m_Specific(X))))
1593         if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1594           return true;
1595 
1596       unsigned BitWidth = V->getType()->getScalarSizeInBits();
1597       APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1598       computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1599 
1600       APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1601       computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1602       // If i8 V is a power of two or zero:
1603       //  ZeroBits: 1 1 1 0 1 1 1 1
1604       // ~ZeroBits: 0 0 0 1 0 0 0 0
1605       if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1606         // If OrZero isn't set, we cannot give back a zero result.
1607         // Make sure either the LHS or RHS has a bit set.
1608         if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1609           return true;
1610     }
1611   }
1612 
1613   // An exact divide or right shift can only shift off zero bits, so the result
1614   // is a power of two only if the first operand is a power of two and not
1615   // copying a sign bit (sdiv int_min, 2).
1616   if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1617       match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1618     return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1619                                   Depth, Q, DL);
1620   }
1621 
1622   return false;
1623 }
1624 
1625 /// \brief Test whether a GEP's result is known to be non-null.
1626 ///
1627 /// Uses properties inherent in a GEP to try to determine whether it is known
1628 /// to be non-null.
1629 ///
1630 /// Currently this routine does not support vector GEPs.
isGEPKnownNonNull(GEPOperator * GEP,const DataLayout & DL,unsigned Depth,const Query & Q)1631 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1632                               unsigned Depth, const Query &Q) {
1633   if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1634     return false;
1635 
1636   // FIXME: Support vector-GEPs.
1637   assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1638 
1639   // If the base pointer is non-null, we cannot walk to a null address with an
1640   // inbounds GEP in address space zero.
1641   if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1642     return true;
1643 
1644   // Walk the GEP operands and see if any operand introduces a non-zero offset.
1645   // If so, then the GEP cannot produce a null pointer, as doing so would
1646   // inherently violate the inbounds contract within address space zero.
1647   for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1648        GTI != GTE; ++GTI) {
1649     // Struct types are easy -- they must always be indexed by a constant.
1650     if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1651       ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1652       unsigned ElementIdx = OpC->getZExtValue();
1653       const StructLayout *SL = DL.getStructLayout(STy);
1654       uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1655       if (ElementOffset > 0)
1656         return true;
1657       continue;
1658     }
1659 
1660     // If we have a zero-sized type, the index doesn't matter. Keep looping.
1661     if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1662       continue;
1663 
1664     // Fast path the constant operand case both for efficiency and so we don't
1665     // increment Depth when just zipping down an all-constant GEP.
1666     if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1667       if (!OpC->isZero())
1668         return true;
1669       continue;
1670     }
1671 
1672     // We post-increment Depth here because while isKnownNonZero increments it
1673     // as well, when we pop back up that increment won't persist. We don't want
1674     // to recurse 10k times just because we have 10k GEP operands. We don't
1675     // bail completely out because we want to handle constant GEPs regardless
1676     // of depth.
1677     if (Depth++ >= MaxDepth)
1678       continue;
1679 
1680     if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1681       return true;
1682   }
1683 
1684   return false;
1685 }
1686 
1687 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1688 /// ensure that the value it's attached to is never Value?  'RangeType' is
1689 /// is the type of the value described by the range.
rangeMetadataExcludesValue(MDNode * Ranges,const APInt & Value)1690 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1691                                        const APInt& Value) {
1692   const unsigned NumRanges = Ranges->getNumOperands() / 2;
1693   assert(NumRanges >= 1);
1694   for (unsigned i = 0; i < NumRanges; ++i) {
1695     ConstantInt *Lower =
1696         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1697     ConstantInt *Upper =
1698         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1699     ConstantRange Range(Lower->getValue(), Upper->getValue());
1700     if (Range.contains(Value))
1701       return false;
1702   }
1703   return true;
1704 }
1705 
1706 /// Return true if the given value is known to be non-zero when defined.
1707 /// For vectors return true if every element is known to be non-zero when
1708 /// defined. Supports values with integer or pointer type and vectors of
1709 /// integers.
isKnownNonZero(Value * V,const DataLayout & DL,unsigned Depth,const Query & Q)1710 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1711                     const Query &Q) {
1712   if (Constant *C = dyn_cast<Constant>(V)) {
1713     if (C->isNullValue())
1714       return false;
1715     if (isa<ConstantInt>(C))
1716       // Must be non-zero due to null test above.
1717       return true;
1718     // TODO: Handle vectors
1719     return false;
1720   }
1721 
1722   if (Instruction* I = dyn_cast<Instruction>(V)) {
1723     if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1724       // If the possible ranges don't contain zero, then the value is
1725       // definitely non-zero.
1726       if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1727         const APInt ZeroValue(Ty->getBitWidth(), 0);
1728         if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1729           return true;
1730       }
1731     }
1732   }
1733 
1734   // The remaining tests are all recursive, so bail out if we hit the limit.
1735   if (Depth++ >= MaxDepth)
1736     return false;
1737 
1738   // Check for pointer simplifications.
1739   if (V->getType()->isPointerTy()) {
1740     if (isKnownNonNull(V))
1741       return true;
1742     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1743       if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1744         return true;
1745   }
1746 
1747   unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1748 
1749   // X | Y != 0 if X != 0 or Y != 0.
1750   Value *X = nullptr, *Y = nullptr;
1751   if (match(V, m_Or(m_Value(X), m_Value(Y))))
1752     return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1753 
1754   // ext X != 0 if X != 0.
1755   if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1756     return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1757 
1758   // shl X, Y != 0 if X is odd.  Note that the value of the shift is undefined
1759   // if the lowest bit is shifted off the end.
1760   if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1761     // shl nuw can't remove any non-zero bits.
1762     OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1763     if (BO->hasNoUnsignedWrap())
1764       return isKnownNonZero(X, DL, Depth, Q);
1765 
1766     APInt KnownZero(BitWidth, 0);
1767     APInt KnownOne(BitWidth, 0);
1768     computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1769     if (KnownOne[0])
1770       return true;
1771   }
1772   // shr X, Y != 0 if X is negative.  Note that the value of the shift is not
1773   // defined if the sign bit is shifted off the end.
1774   else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1775     // shr exact can only shift out zero bits.
1776     PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1777     if (BO->isExact())
1778       return isKnownNonZero(X, DL, Depth, Q);
1779 
1780     bool XKnownNonNegative, XKnownNegative;
1781     ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1782     if (XKnownNegative)
1783       return true;
1784   }
1785   // div exact can only produce a zero if the dividend is zero.
1786   else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1787     return isKnownNonZero(X, DL, Depth, Q);
1788   }
1789   // X + Y.
1790   else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1791     bool XKnownNonNegative, XKnownNegative;
1792     bool YKnownNonNegative, YKnownNegative;
1793     ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1794     ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1795 
1796     // If X and Y are both non-negative (as signed values) then their sum is not
1797     // zero unless both X and Y are zero.
1798     if (XKnownNonNegative && YKnownNonNegative)
1799       if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1800         return true;
1801 
1802     // If X and Y are both negative (as signed values) then their sum is not
1803     // zero unless both X and Y equal INT_MIN.
1804     if (BitWidth && XKnownNegative && YKnownNegative) {
1805       APInt KnownZero(BitWidth, 0);
1806       APInt KnownOne(BitWidth, 0);
1807       APInt Mask = APInt::getSignedMaxValue(BitWidth);
1808       // The sign bit of X is set.  If some other bit is set then X is not equal
1809       // to INT_MIN.
1810       computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1811       if ((KnownOne & Mask) != 0)
1812         return true;
1813       // The sign bit of Y is set.  If some other bit is set then Y is not equal
1814       // to INT_MIN.
1815       computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1816       if ((KnownOne & Mask) != 0)
1817         return true;
1818     }
1819 
1820     // The sum of a non-negative number and a power of two is not zero.
1821     if (XKnownNonNegative &&
1822         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1823       return true;
1824     if (YKnownNonNegative &&
1825         isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1826       return true;
1827   }
1828   // X * Y.
1829   else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1830     OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1831     // If X and Y are non-zero then so is X * Y as long as the multiplication
1832     // does not overflow.
1833     if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1834         isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
1835       return true;
1836   }
1837   // (C ? X : Y) != 0 if X != 0 and Y != 0.
1838   else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1839     if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
1840         isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
1841       return true;
1842   }
1843 
1844   if (!BitWidth) return false;
1845   APInt KnownZero(BitWidth, 0);
1846   APInt KnownOne(BitWidth, 0);
1847   computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1848   return KnownOne != 0;
1849 }
1850 
1851 /// Return true if 'V & Mask' is known to be zero.  We use this predicate to
1852 /// simplify operations downstream. Mask is known to be zero for bits that V
1853 /// cannot have.
1854 ///
1855 /// This function is defined on values with integer type, values with pointer
1856 /// type, and vectors of integers.  In the case
1857 /// where V is a vector, the mask, known zero, and known one values are the
1858 /// same width as the vector element, and the bit is set only if it is true
1859 /// for all of the elements in the vector.
MaskedValueIsZero(Value * V,const APInt & Mask,const DataLayout & DL,unsigned Depth,const Query & Q)1860 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
1861                        unsigned Depth, const Query &Q) {
1862   APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1863   computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1864   return (KnownZero & Mask) == Mask;
1865 }
1866 
1867 
1868 
1869 /// Return the number of times the sign bit of the register is replicated into
1870 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1871 /// (itself), but other cases can give us information. For example, immediately
1872 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1873 /// other, so we return 3.
1874 ///
1875 /// 'Op' must have a scalar integer type.
1876 ///
ComputeNumSignBits(Value * V,const DataLayout & DL,unsigned Depth,const Query & Q)1877 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
1878                             const Query &Q) {
1879   unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
1880   unsigned Tmp, Tmp2;
1881   unsigned FirstAnswer = 1;
1882 
1883   // Note that ConstantInt is handled by the general computeKnownBits case
1884   // below.
1885 
1886   if (Depth == 6)
1887     return 1;  // Limit search depth.
1888 
1889   Operator *U = dyn_cast<Operator>(V);
1890   switch (Operator::getOpcode(V)) {
1891   default: break;
1892   case Instruction::SExt:
1893     Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1894     return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
1895 
1896   case Instruction::SDiv: {
1897     const APInt *Denominator;
1898     // sdiv X, C -> adds log(C) sign bits.
1899     if (match(U->getOperand(1), m_APInt(Denominator))) {
1900 
1901       // Ignore non-positive denominator.
1902       if (!Denominator->isStrictlyPositive())
1903         break;
1904 
1905       // Calculate the incoming numerator bits.
1906       unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1907 
1908       // Add floor(log(C)) bits to the numerator bits.
1909       return std::min(TyBits, NumBits + Denominator->logBase2());
1910     }
1911     break;
1912   }
1913 
1914   case Instruction::SRem: {
1915     const APInt *Denominator;
1916     // srem X, C -> we know that the result is within [-C+1,C) when C is a
1917     // positive constant.  This let us put a lower bound on the number of sign
1918     // bits.
1919     if (match(U->getOperand(1), m_APInt(Denominator))) {
1920 
1921       // Ignore non-positive denominator.
1922       if (!Denominator->isStrictlyPositive())
1923         break;
1924 
1925       // Calculate the incoming numerator bits. SRem by a positive constant
1926       // can't lower the number of sign bits.
1927       unsigned NumrBits =
1928           ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1929 
1930       // Calculate the leading sign bit constraints by examining the
1931       // denominator.  Given that the denominator is positive, there are two
1932       // cases:
1933       //
1934       //  1. the numerator is positive.  The result range is [0,C) and [0,C) u<
1935       //     (1 << ceilLogBase2(C)).
1936       //
1937       //  2. the numerator is negative.  Then the result range is (-C,0] and
1938       //     integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
1939       //
1940       // Thus a lower bound on the number of sign bits is `TyBits -
1941       // ceilLogBase2(C)`.
1942 
1943       unsigned ResBits = TyBits - Denominator->ceilLogBase2();
1944       return std::max(NumrBits, ResBits);
1945     }
1946     break;
1947   }
1948 
1949   case Instruction::AShr: {
1950     Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1951     // ashr X, C   -> adds C sign bits.  Vectors too.
1952     const APInt *ShAmt;
1953     if (match(U->getOperand(1), m_APInt(ShAmt))) {
1954       Tmp += ShAmt->getZExtValue();
1955       if (Tmp > TyBits) Tmp = TyBits;
1956     }
1957     return Tmp;
1958   }
1959   case Instruction::Shl: {
1960     const APInt *ShAmt;
1961     if (match(U->getOperand(1), m_APInt(ShAmt))) {
1962       // shl destroys sign bits.
1963       Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1964       Tmp2 = ShAmt->getZExtValue();
1965       if (Tmp2 >= TyBits ||      // Bad shift.
1966           Tmp2 >= Tmp) break;    // Shifted all sign bits out.
1967       return Tmp - Tmp2;
1968     }
1969     break;
1970   }
1971   case Instruction::And:
1972   case Instruction::Or:
1973   case Instruction::Xor:    // NOT is handled here.
1974     // Logical binary ops preserve the number of sign bits at the worst.
1975     Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1976     if (Tmp != 1) {
1977       Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
1978       FirstAnswer = std::min(Tmp, Tmp2);
1979       // We computed what we know about the sign bits as our first
1980       // answer. Now proceed to the generic code that uses
1981       // computeKnownBits, and pick whichever answer is better.
1982     }
1983     break;
1984 
1985   case Instruction::Select:
1986     Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
1987     if (Tmp == 1) return 1;  // Early out.
1988     Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
1989     return std::min(Tmp, Tmp2);
1990 
1991   case Instruction::Add:
1992     // Add can have at most one carry bit.  Thus we know that the output
1993     // is, at worst, one more bit than the inputs.
1994     Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1995     if (Tmp == 1) return 1;  // Early out.
1996 
1997     // Special case decrementing a value (ADD X, -1):
1998     if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
1999       if (CRHS->isAllOnesValue()) {
2000         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2001         computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2002                          Q);
2003 
2004         // If the input is known to be 0 or 1, the output is 0/-1, which is all
2005         // sign bits set.
2006         if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2007           return TyBits;
2008 
2009         // If we are subtracting one from a positive number, there is no carry
2010         // out of the result.
2011         if (KnownZero.isNegative())
2012           return Tmp;
2013       }
2014 
2015     Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2016     if (Tmp2 == 1) return 1;
2017     return std::min(Tmp, Tmp2)-1;
2018 
2019   case Instruction::Sub:
2020     Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2021     if (Tmp2 == 1) return 1;
2022 
2023     // Handle NEG.
2024     if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2025       if (CLHS->isNullValue()) {
2026         APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2027         computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2028                          Q);
2029         // If the input is known to be 0 or 1, the output is 0/-1, which is all
2030         // sign bits set.
2031         if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2032           return TyBits;
2033 
2034         // If the input is known to be positive (the sign bit is known clear),
2035         // the output of the NEG has the same number of sign bits as the input.
2036         if (KnownZero.isNegative())
2037           return Tmp2;
2038 
2039         // Otherwise, we treat this like a SUB.
2040       }
2041 
2042     // Sub can have at most one carry bit.  Thus we know that the output
2043     // is, at worst, one more bit than the inputs.
2044     Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2045     if (Tmp == 1) return 1;  // Early out.
2046     return std::min(Tmp, Tmp2)-1;
2047 
2048   case Instruction::PHI: {
2049     PHINode *PN = cast<PHINode>(U);
2050     unsigned NumIncomingValues = PN->getNumIncomingValues();
2051     // Don't analyze large in-degree PHIs.
2052     if (NumIncomingValues > 4) break;
2053     // Unreachable blocks may have zero-operand PHI nodes.
2054     if (NumIncomingValues == 0) break;
2055 
2056     // Take the minimum of all incoming values.  This can't infinitely loop
2057     // because of our depth threshold.
2058     Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2059     for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2060       if (Tmp == 1) return Tmp;
2061       Tmp = std::min(
2062           Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2063     }
2064     return Tmp;
2065   }
2066 
2067   case Instruction::Trunc:
2068     // FIXME: it's tricky to do anything useful for this, but it is an important
2069     // case for targets like X86.
2070     break;
2071   }
2072 
2073   // Finally, if we can prove that the top bits of the result are 0's or 1's,
2074   // use this information.
2075   APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2076   APInt Mask;
2077   computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2078 
2079   if (KnownZero.isNegative()) {        // sign bit is 0
2080     Mask = KnownZero;
2081   } else if (KnownOne.isNegative()) {  // sign bit is 1;
2082     Mask = KnownOne;
2083   } else {
2084     // Nothing known.
2085     return FirstAnswer;
2086   }
2087 
2088   // Okay, we know that the sign bit in Mask is set.  Use CLZ to determine
2089   // the number of identical bits in the top of the input value.
2090   Mask = ~Mask;
2091   Mask <<= Mask.getBitWidth()-TyBits;
2092   // Return # leading zeros.  We use 'min' here in case Val was zero before
2093   // shifting.  We don't want to return '64' as for an i32 "0".
2094   return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2095 }
2096 
2097 /// This function computes the integer multiple of Base that equals V.
2098 /// If successful, it returns true and returns the multiple in
2099 /// Multiple. If unsuccessful, it returns false. It looks
2100 /// through SExt instructions only if LookThroughSExt is true.
ComputeMultiple(Value * V,unsigned Base,Value * & Multiple,bool LookThroughSExt,unsigned Depth)2101 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2102                            bool LookThroughSExt, unsigned Depth) {
2103   const unsigned MaxDepth = 6;
2104 
2105   assert(V && "No Value?");
2106   assert(Depth <= MaxDepth && "Limit Search Depth");
2107   assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2108 
2109   Type *T = V->getType();
2110 
2111   ConstantInt *CI = dyn_cast<ConstantInt>(V);
2112 
2113   if (Base == 0)
2114     return false;
2115 
2116   if (Base == 1) {
2117     Multiple = V;
2118     return true;
2119   }
2120 
2121   ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2122   Constant *BaseVal = ConstantInt::get(T, Base);
2123   if (CO && CO == BaseVal) {
2124     // Multiple is 1.
2125     Multiple = ConstantInt::get(T, 1);
2126     return true;
2127   }
2128 
2129   if (CI && CI->getZExtValue() % Base == 0) {
2130     Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2131     return true;
2132   }
2133 
2134   if (Depth == MaxDepth) return false;  // Limit search depth.
2135 
2136   Operator *I = dyn_cast<Operator>(V);
2137   if (!I) return false;
2138 
2139   switch (I->getOpcode()) {
2140   default: break;
2141   case Instruction::SExt:
2142     if (!LookThroughSExt) return false;
2143     // otherwise fall through to ZExt
2144   case Instruction::ZExt:
2145     return ComputeMultiple(I->getOperand(0), Base, Multiple,
2146                            LookThroughSExt, Depth+1);
2147   case Instruction::Shl:
2148   case Instruction::Mul: {
2149     Value *Op0 = I->getOperand(0);
2150     Value *Op1 = I->getOperand(1);
2151 
2152     if (I->getOpcode() == Instruction::Shl) {
2153       ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2154       if (!Op1CI) return false;
2155       // Turn Op0 << Op1 into Op0 * 2^Op1
2156       APInt Op1Int = Op1CI->getValue();
2157       uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2158       APInt API(Op1Int.getBitWidth(), 0);
2159       API.setBit(BitToSet);
2160       Op1 = ConstantInt::get(V->getContext(), API);
2161     }
2162 
2163     Value *Mul0 = nullptr;
2164     if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2165       if (Constant *Op1C = dyn_cast<Constant>(Op1))
2166         if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2167           if (Op1C->getType()->getPrimitiveSizeInBits() <
2168               MulC->getType()->getPrimitiveSizeInBits())
2169             Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2170           if (Op1C->getType()->getPrimitiveSizeInBits() >
2171               MulC->getType()->getPrimitiveSizeInBits())
2172             MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2173 
2174           // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2175           Multiple = ConstantExpr::getMul(MulC, Op1C);
2176           return true;
2177         }
2178 
2179       if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2180         if (Mul0CI->getValue() == 1) {
2181           // V == Base * Op1, so return Op1
2182           Multiple = Op1;
2183           return true;
2184         }
2185     }
2186 
2187     Value *Mul1 = nullptr;
2188     if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2189       if (Constant *Op0C = dyn_cast<Constant>(Op0))
2190         if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2191           if (Op0C->getType()->getPrimitiveSizeInBits() <
2192               MulC->getType()->getPrimitiveSizeInBits())
2193             Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2194           if (Op0C->getType()->getPrimitiveSizeInBits() >
2195               MulC->getType()->getPrimitiveSizeInBits())
2196             MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2197 
2198           // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2199           Multiple = ConstantExpr::getMul(MulC, Op0C);
2200           return true;
2201         }
2202 
2203       if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2204         if (Mul1CI->getValue() == 1) {
2205           // V == Base * Op0, so return Op0
2206           Multiple = Op0;
2207           return true;
2208         }
2209     }
2210   }
2211   }
2212 
2213   // We could not determine if V is a multiple of Base.
2214   return false;
2215 }
2216 
2217 /// Return true if we can prove that the specified FP value is never equal to
2218 /// -0.0.
2219 ///
2220 /// NOTE: this function will need to be revisited when we support non-default
2221 /// rounding modes!
2222 ///
CannotBeNegativeZero(const Value * V,unsigned Depth)2223 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2224   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2225     return !CFP->getValueAPF().isNegZero();
2226 
2227   // FIXME: Magic number! At the least, this should be given a name because it's
2228   // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2229   // expose it as a parameter, so it can be used for testing / experimenting.
2230   if (Depth == 6)
2231     return false;  // Limit search depth.
2232 
2233   const Operator *I = dyn_cast<Operator>(V);
2234   if (!I) return false;
2235 
2236   // Check if the nsz fast-math flag is set
2237   if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2238     if (FPO->hasNoSignedZeros())
2239       return true;
2240 
2241   // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2242   if (I->getOpcode() == Instruction::FAdd)
2243     if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2244       if (CFP->isNullValue())
2245         return true;
2246 
2247   // sitofp and uitofp turn into +0.0 for zero.
2248   if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2249     return true;
2250 
2251   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2252     // sqrt(-0.0) = -0.0, no other negative results are possible.
2253     if (II->getIntrinsicID() == Intrinsic::sqrt)
2254       return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2255 
2256   if (const CallInst *CI = dyn_cast<CallInst>(I))
2257     if (const Function *F = CI->getCalledFunction()) {
2258       if (F->isDeclaration()) {
2259         // abs(x) != -0.0
2260         if (F->getName() == "abs") return true;
2261         // fabs[lf](x) != -0.0
2262         if (F->getName() == "fabs") return true;
2263         if (F->getName() == "fabsf") return true;
2264         if (F->getName() == "fabsl") return true;
2265         if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2266             F->getName() == "sqrtl")
2267           return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2268       }
2269     }
2270 
2271   return false;
2272 }
2273 
CannotBeOrderedLessThanZero(const Value * V,unsigned Depth)2274 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2275   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2276     return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2277 
2278   // FIXME: Magic number! At the least, this should be given a name because it's
2279   // used similarly in CannotBeNegativeZero(). A better fix may be to
2280   // expose it as a parameter, so it can be used for testing / experimenting.
2281   if (Depth == 6)
2282     return false;  // Limit search depth.
2283 
2284   const Operator *I = dyn_cast<Operator>(V);
2285   if (!I) return false;
2286 
2287   switch (I->getOpcode()) {
2288   default: break;
2289   case Instruction::FMul:
2290     // x*x is always non-negative or a NaN.
2291     if (I->getOperand(0) == I->getOperand(1))
2292       return true;
2293     // Fall through
2294   case Instruction::FAdd:
2295   case Instruction::FDiv:
2296   case Instruction::FRem:
2297     return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2298            CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2299   case Instruction::FPExt:
2300   case Instruction::FPTrunc:
2301     // Widening/narrowing never change sign.
2302     return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2303   case Instruction::Call:
2304     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2305       switch (II->getIntrinsicID()) {
2306       default: break;
2307       case Intrinsic::exp:
2308       case Intrinsic::exp2:
2309       case Intrinsic::fabs:
2310       case Intrinsic::sqrt:
2311         return true;
2312       case Intrinsic::powi:
2313         if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2314           // powi(x,n) is non-negative if n is even.
2315           if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2316             return true;
2317         }
2318         return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2319       case Intrinsic::fma:
2320       case Intrinsic::fmuladd:
2321         // x*x+y is non-negative if y is non-negative.
2322         return I->getOperand(0) == I->getOperand(1) &&
2323                CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2324       }
2325     break;
2326   }
2327   return false;
2328 }
2329 
2330 /// If the specified value can be set by repeating the same byte in memory,
2331 /// return the i8 value that it is represented with.  This is
2332 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2333 /// i16 0xF0F0, double 0.0 etc.  If the value can't be handled with a repeated
2334 /// byte store (e.g. i16 0x1234), return null.
isBytewiseValue(Value * V)2335 Value *llvm::isBytewiseValue(Value *V) {
2336   // All byte-wide stores are splatable, even of arbitrary variables.
2337   if (V->getType()->isIntegerTy(8)) return V;
2338 
2339   // Handle 'null' ConstantArrayZero etc.
2340   if (Constant *C = dyn_cast<Constant>(V))
2341     if (C->isNullValue())
2342       return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2343 
2344   // Constant float and double values can be handled as integer values if the
2345   // corresponding integer value is "byteable".  An important case is 0.0.
2346   if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2347     if (CFP->getType()->isFloatTy())
2348       V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2349     if (CFP->getType()->isDoubleTy())
2350       V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2351     // Don't handle long double formats, which have strange constraints.
2352   }
2353 
2354   // We can handle constant integers that are multiple of 8 bits.
2355   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2356     if (CI->getBitWidth() % 8 == 0) {
2357       assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2358 
2359       if (!CI->getValue().isSplat(8))
2360         return nullptr;
2361       return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2362     }
2363   }
2364 
2365   // A ConstantDataArray/Vector is splatable if all its members are equal and
2366   // also splatable.
2367   if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2368     Value *Elt = CA->getElementAsConstant(0);
2369     Value *Val = isBytewiseValue(Elt);
2370     if (!Val)
2371       return nullptr;
2372 
2373     for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2374       if (CA->getElementAsConstant(I) != Elt)
2375         return nullptr;
2376 
2377     return Val;
2378   }
2379 
2380   // Conceptually, we could handle things like:
2381   //   %a = zext i8 %X to i16
2382   //   %b = shl i16 %a, 8
2383   //   %c = or i16 %a, %b
2384   // but until there is an example that actually needs this, it doesn't seem
2385   // worth worrying about.
2386   return nullptr;
2387 }
2388 
2389 
2390 // This is the recursive version of BuildSubAggregate. It takes a few different
2391 // arguments. Idxs is the index within the nested struct From that we are
2392 // looking at now (which is of type IndexedType). IdxSkip is the number of
2393 // indices from Idxs that should be left out when inserting into the resulting
2394 // struct. To is the result struct built so far, new insertvalue instructions
2395 // build on that.
BuildSubAggregate(Value * From,Value * To,Type * IndexedType,SmallVectorImpl<unsigned> & Idxs,unsigned IdxSkip,Instruction * InsertBefore)2396 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2397                                 SmallVectorImpl<unsigned> &Idxs,
2398                                 unsigned IdxSkip,
2399                                 Instruction *InsertBefore) {
2400   llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2401   if (STy) {
2402     // Save the original To argument so we can modify it
2403     Value *OrigTo = To;
2404     // General case, the type indexed by Idxs is a struct
2405     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2406       // Process each struct element recursively
2407       Idxs.push_back(i);
2408       Value *PrevTo = To;
2409       To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2410                              InsertBefore);
2411       Idxs.pop_back();
2412       if (!To) {
2413         // Couldn't find any inserted value for this index? Cleanup
2414         while (PrevTo != OrigTo) {
2415           InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2416           PrevTo = Del->getAggregateOperand();
2417           Del->eraseFromParent();
2418         }
2419         // Stop processing elements
2420         break;
2421       }
2422     }
2423     // If we successfully found a value for each of our subaggregates
2424     if (To)
2425       return To;
2426   }
2427   // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2428   // the struct's elements had a value that was inserted directly. In the latter
2429   // case, perhaps we can't determine each of the subelements individually, but
2430   // we might be able to find the complete struct somewhere.
2431 
2432   // Find the value that is at that particular spot
2433   Value *V = FindInsertedValue(From, Idxs);
2434 
2435   if (!V)
2436     return nullptr;
2437 
2438   // Insert the value in the new (sub) aggregrate
2439   return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2440                                        "tmp", InsertBefore);
2441 }
2442 
2443 // This helper takes a nested struct and extracts a part of it (which is again a
2444 // struct) into a new value. For example, given the struct:
2445 // { a, { b, { c, d }, e } }
2446 // and the indices "1, 1" this returns
2447 // { c, d }.
2448 //
2449 // It does this by inserting an insertvalue for each element in the resulting
2450 // struct, as opposed to just inserting a single struct. This will only work if
2451 // each of the elements of the substruct are known (ie, inserted into From by an
2452 // insertvalue instruction somewhere).
2453 //
2454 // All inserted insertvalue instructions are inserted before InsertBefore
BuildSubAggregate(Value * From,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)2455 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2456                                 Instruction *InsertBefore) {
2457   assert(InsertBefore && "Must have someplace to insert!");
2458   Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2459                                                              idx_range);
2460   Value *To = UndefValue::get(IndexedType);
2461   SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2462   unsigned IdxSkip = Idxs.size();
2463 
2464   return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2465 }
2466 
2467 /// Given an aggregrate and an sequence of indices, see if
2468 /// the scalar value indexed is already around as a register, for example if it
2469 /// were inserted directly into the aggregrate.
2470 ///
2471 /// If InsertBefore is not null, this function will duplicate (modified)
2472 /// insertvalues when a part of a nested struct is extracted.
FindInsertedValue(Value * V,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)2473 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2474                                Instruction *InsertBefore) {
2475   // Nothing to index? Just return V then (this is useful at the end of our
2476   // recursion).
2477   if (idx_range.empty())
2478     return V;
2479   // We have indices, so V should have an indexable type.
2480   assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2481          "Not looking at a struct or array?");
2482   assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2483          "Invalid indices for type?");
2484 
2485   if (Constant *C = dyn_cast<Constant>(V)) {
2486     C = C->getAggregateElement(idx_range[0]);
2487     if (!C) return nullptr;
2488     return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2489   }
2490 
2491   if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2492     // Loop the indices for the insertvalue instruction in parallel with the
2493     // requested indices
2494     const unsigned *req_idx = idx_range.begin();
2495     for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2496          i != e; ++i, ++req_idx) {
2497       if (req_idx == idx_range.end()) {
2498         // We can't handle this without inserting insertvalues
2499         if (!InsertBefore)
2500           return nullptr;
2501 
2502         // The requested index identifies a part of a nested aggregate. Handle
2503         // this specially. For example,
2504         // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2505         // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2506         // %C = extractvalue {i32, { i32, i32 } } %B, 1
2507         // This can be changed into
2508         // %A = insertvalue {i32, i32 } undef, i32 10, 0
2509         // %C = insertvalue {i32, i32 } %A, i32 11, 1
2510         // which allows the unused 0,0 element from the nested struct to be
2511         // removed.
2512         return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2513                                  InsertBefore);
2514       }
2515 
2516       // This insert value inserts something else than what we are looking for.
2517       // See if the (aggregrate) value inserted into has the value we are
2518       // looking for, then.
2519       if (*req_idx != *i)
2520         return FindInsertedValue(I->getAggregateOperand(), idx_range,
2521                                  InsertBefore);
2522     }
2523     // If we end up here, the indices of the insertvalue match with those
2524     // requested (though possibly only partially). Now we recursively look at
2525     // the inserted value, passing any remaining indices.
2526     return FindInsertedValue(I->getInsertedValueOperand(),
2527                              makeArrayRef(req_idx, idx_range.end()),
2528                              InsertBefore);
2529   }
2530 
2531   if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2532     // If we're extracting a value from an aggregrate that was extracted from
2533     // something else, we can extract from that something else directly instead.
2534     // However, we will need to chain I's indices with the requested indices.
2535 
2536     // Calculate the number of indices required
2537     unsigned size = I->getNumIndices() + idx_range.size();
2538     // Allocate some space to put the new indices in
2539     SmallVector<unsigned, 5> Idxs;
2540     Idxs.reserve(size);
2541     // Add indices from the extract value instruction
2542     Idxs.append(I->idx_begin(), I->idx_end());
2543 
2544     // Add requested indices
2545     Idxs.append(idx_range.begin(), idx_range.end());
2546 
2547     assert(Idxs.size() == size
2548            && "Number of indices added not correct?");
2549 
2550     return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2551   }
2552   // Otherwise, we don't know (such as, extracting from a function return value
2553   // or load instruction)
2554   return nullptr;
2555 }
2556 
2557 /// Analyze the specified pointer to see if it can be expressed as a base
2558 /// pointer plus a constant offset. Return the base and offset to the caller.
GetPointerBaseWithConstantOffset(Value * Ptr,int64_t & Offset,const DataLayout & DL)2559 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2560                                               const DataLayout &DL) {
2561   unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2562   APInt ByteOffset(BitWidth, 0);
2563   while (1) {
2564     if (Ptr->getType()->isVectorTy())
2565       break;
2566 
2567     if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2568       APInt GEPOffset(BitWidth, 0);
2569       if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2570         break;
2571 
2572       ByteOffset += GEPOffset;
2573 
2574       Ptr = GEP->getPointerOperand();
2575     } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2576                Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2577       Ptr = cast<Operator>(Ptr)->getOperand(0);
2578     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2579       if (GA->mayBeOverridden())
2580         break;
2581       Ptr = GA->getAliasee();
2582     } else {
2583       break;
2584     }
2585   }
2586   Offset = ByteOffset.getSExtValue();
2587   return Ptr;
2588 }
2589 
2590 
2591 /// This function computes the length of a null-terminated C string pointed to
2592 /// by V. If successful, it returns true and returns the string in Str.
2593 /// If unsuccessful, it returns false.
getConstantStringInfo(const Value * V,StringRef & Str,uint64_t Offset,bool TrimAtNul)2594 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2595                                  uint64_t Offset, bool TrimAtNul) {
2596   assert(V);
2597 
2598   // Look through bitcast instructions and geps.
2599   V = V->stripPointerCasts();
2600 
2601   // If the value is a GEP instruction or constant expression, treat it as an
2602   // offset.
2603   if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2604     // Make sure the GEP has exactly three arguments.
2605     if (GEP->getNumOperands() != 3)
2606       return false;
2607 
2608     // Make sure the index-ee is a pointer to array of i8.
2609     PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2610     ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2611     if (!AT || !AT->getElementType()->isIntegerTy(8))
2612       return false;
2613 
2614     // Check to make sure that the first operand of the GEP is an integer and
2615     // has value 0 so that we are sure we're indexing into the initializer.
2616     const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2617     if (!FirstIdx || !FirstIdx->isZero())
2618       return false;
2619 
2620     // If the second index isn't a ConstantInt, then this is a variable index
2621     // into the array.  If this occurs, we can't say anything meaningful about
2622     // the string.
2623     uint64_t StartIdx = 0;
2624     if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2625       StartIdx = CI->getZExtValue();
2626     else
2627       return false;
2628     return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2629                                  TrimAtNul);
2630   }
2631 
2632   // The GEP instruction, constant or instruction, must reference a global
2633   // variable that is a constant and is initialized. The referenced constant
2634   // initializer is the array that we'll use for optimization.
2635   const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2636   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2637     return false;
2638 
2639   // Handle the all-zeros case
2640   if (GV->getInitializer()->isNullValue()) {
2641     // This is a degenerate case. The initializer is constant zero so the
2642     // length of the string must be zero.
2643     Str = "";
2644     return true;
2645   }
2646 
2647   // Must be a Constant Array
2648   const ConstantDataArray *Array =
2649     dyn_cast<ConstantDataArray>(GV->getInitializer());
2650   if (!Array || !Array->isString())
2651     return false;
2652 
2653   // Get the number of elements in the array
2654   uint64_t NumElts = Array->getType()->getArrayNumElements();
2655 
2656   // Start out with the entire array in the StringRef.
2657   Str = Array->getAsString();
2658 
2659   if (Offset > NumElts)
2660     return false;
2661 
2662   // Skip over 'offset' bytes.
2663   Str = Str.substr(Offset);
2664 
2665   if (TrimAtNul) {
2666     // Trim off the \0 and anything after it.  If the array is not nul
2667     // terminated, we just return the whole end of string.  The client may know
2668     // some other way that the string is length-bound.
2669     Str = Str.substr(0, Str.find('\0'));
2670   }
2671   return true;
2672 }
2673 
2674 // These next two are very similar to the above, but also look through PHI
2675 // nodes.
2676 // TODO: See if we can integrate these two together.
2677 
2678 /// If we can compute the length of the string pointed to by
2679 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLengthH(Value * V,SmallPtrSetImpl<PHINode * > & PHIs)2680 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2681   // Look through noop bitcast instructions.
2682   V = V->stripPointerCasts();
2683 
2684   // If this is a PHI node, there are two cases: either we have already seen it
2685   // or we haven't.
2686   if (PHINode *PN = dyn_cast<PHINode>(V)) {
2687     if (!PHIs.insert(PN).second)
2688       return ~0ULL;  // already in the set.
2689 
2690     // If it was new, see if all the input strings are the same length.
2691     uint64_t LenSoFar = ~0ULL;
2692     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2693       uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2694       if (Len == 0) return 0; // Unknown length -> unknown.
2695 
2696       if (Len == ~0ULL) continue;
2697 
2698       if (Len != LenSoFar && LenSoFar != ~0ULL)
2699         return 0;    // Disagree -> unknown.
2700       LenSoFar = Len;
2701     }
2702 
2703     // Success, all agree.
2704     return LenSoFar;
2705   }
2706 
2707   // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2708   if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2709     uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2710     if (Len1 == 0) return 0;
2711     uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2712     if (Len2 == 0) return 0;
2713     if (Len1 == ~0ULL) return Len2;
2714     if (Len2 == ~0ULL) return Len1;
2715     if (Len1 != Len2) return 0;
2716     return Len1;
2717   }
2718 
2719   // Otherwise, see if we can read the string.
2720   StringRef StrData;
2721   if (!getConstantStringInfo(V, StrData))
2722     return 0;
2723 
2724   return StrData.size()+1;
2725 }
2726 
2727 /// If we can compute the length of the string pointed to by
2728 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLength(Value * V)2729 uint64_t llvm::GetStringLength(Value *V) {
2730   if (!V->getType()->isPointerTy()) return 0;
2731 
2732   SmallPtrSet<PHINode*, 32> PHIs;
2733   uint64_t Len = GetStringLengthH(V, PHIs);
2734   // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2735   // an empty string as a length.
2736   return Len == ~0ULL ? 1 : Len;
2737 }
2738 
GetUnderlyingObject(Value * V,const DataLayout & DL,unsigned MaxLookup)2739 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2740                                  unsigned MaxLookup) {
2741   if (!V->getType()->isPointerTy())
2742     return V;
2743   for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2744     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2745       V = GEP->getPointerOperand();
2746     } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2747                Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2748       V = cast<Operator>(V)->getOperand(0);
2749     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2750       if (GA->mayBeOverridden())
2751         return V;
2752       V = GA->getAliasee();
2753     } else {
2754       // See if InstructionSimplify knows any relevant tricks.
2755       if (Instruction *I = dyn_cast<Instruction>(V))
2756         // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2757         if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
2758           V = Simplified;
2759           continue;
2760         }
2761 
2762       return V;
2763     }
2764     assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2765   }
2766   return V;
2767 }
2768 
GetUnderlyingObjects(Value * V,SmallVectorImpl<Value * > & Objects,const DataLayout & DL,unsigned MaxLookup)2769 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
2770                                 const DataLayout &DL, unsigned MaxLookup) {
2771   SmallPtrSet<Value *, 4> Visited;
2772   SmallVector<Value *, 4> Worklist;
2773   Worklist.push_back(V);
2774   do {
2775     Value *P = Worklist.pop_back_val();
2776     P = GetUnderlyingObject(P, DL, MaxLookup);
2777 
2778     if (!Visited.insert(P).second)
2779       continue;
2780 
2781     if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2782       Worklist.push_back(SI->getTrueValue());
2783       Worklist.push_back(SI->getFalseValue());
2784       continue;
2785     }
2786 
2787     if (PHINode *PN = dyn_cast<PHINode>(P)) {
2788       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2789         Worklist.push_back(PN->getIncomingValue(i));
2790       continue;
2791     }
2792 
2793     Objects.push_back(P);
2794   } while (!Worklist.empty());
2795 }
2796 
2797 /// Return true if the only users of this pointer are lifetime markers.
onlyUsedByLifetimeMarkers(const Value * V)2798 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2799   for (const User *U : V->users()) {
2800     const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2801     if (!II) return false;
2802 
2803     if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2804         II->getIntrinsicID() != Intrinsic::lifetime_end)
2805       return false;
2806   }
2807   return true;
2808 }
2809 
isSafeToSpeculativelyExecute(const Value * V)2810 bool llvm::isSafeToSpeculativelyExecute(const Value *V) {
2811   const Operator *Inst = dyn_cast<Operator>(V);
2812   if (!Inst)
2813     return false;
2814 
2815   for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2816     if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2817       if (C->canTrap())
2818         return false;
2819 
2820   switch (Inst->getOpcode()) {
2821   default:
2822     return true;
2823   case Instruction::UDiv:
2824   case Instruction::URem: {
2825     // x / y is undefined if y == 0.
2826     const APInt *V;
2827     if (match(Inst->getOperand(1), m_APInt(V)))
2828       return *V != 0;
2829     return false;
2830   }
2831   case Instruction::SDiv:
2832   case Instruction::SRem: {
2833     // x / y is undefined if y == 0 or x == INT_MIN and y == -1
2834     const APInt *Numerator, *Denominator;
2835     if (!match(Inst->getOperand(1), m_APInt(Denominator)))
2836       return false;
2837     // We cannot hoist this division if the denominator is 0.
2838     if (*Denominator == 0)
2839       return false;
2840     // It's safe to hoist if the denominator is not 0 or -1.
2841     if (*Denominator != -1)
2842       return true;
2843     // At this point we know that the denominator is -1.  It is safe to hoist as
2844     // long we know that the numerator is not INT_MIN.
2845     if (match(Inst->getOperand(0), m_APInt(Numerator)))
2846       return !Numerator->isMinSignedValue();
2847     // The numerator *might* be MinSignedValue.
2848     return false;
2849   }
2850   case Instruction::Load: {
2851     const LoadInst *LI = cast<LoadInst>(Inst);
2852     if (!LI->isUnordered() ||
2853         // Speculative load may create a race that did not exist in the source.
2854         LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2855       return false;
2856     const DataLayout &DL = LI->getModule()->getDataLayout();
2857     return LI->getPointerOperand()->isDereferenceablePointer(DL);
2858   }
2859   case Instruction::Call: {
2860     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2861       switch (II->getIntrinsicID()) {
2862       // These synthetic intrinsics have no side-effects and just mark
2863       // information about their operands.
2864       // FIXME: There are other no-op synthetic instructions that potentially
2865       // should be considered at least *safe* to speculate...
2866       case Intrinsic::dbg_declare:
2867       case Intrinsic::dbg_value:
2868         return true;
2869 
2870       case Intrinsic::bswap:
2871       case Intrinsic::ctlz:
2872       case Intrinsic::ctpop:
2873       case Intrinsic::cttz:
2874       case Intrinsic::objectsize:
2875       case Intrinsic::sadd_with_overflow:
2876       case Intrinsic::smul_with_overflow:
2877       case Intrinsic::ssub_with_overflow:
2878       case Intrinsic::uadd_with_overflow:
2879       case Intrinsic::umul_with_overflow:
2880       case Intrinsic::usub_with_overflow:
2881         return true;
2882       // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2883       // errno like libm sqrt would.
2884       case Intrinsic::sqrt:
2885       case Intrinsic::fma:
2886       case Intrinsic::fmuladd:
2887       case Intrinsic::fabs:
2888       case Intrinsic::minnum:
2889       case Intrinsic::maxnum:
2890         return true;
2891       // TODO: some fp intrinsics are marked as having the same error handling
2892       // as libm. They're safe to speculate when they won't error.
2893       // TODO: are convert_{from,to}_fp16 safe?
2894       // TODO: can we list target-specific intrinsics here?
2895       default: break;
2896       }
2897     }
2898     return false; // The called function could have undefined behavior or
2899                   // side-effects, even if marked readnone nounwind.
2900   }
2901   case Instruction::VAArg:
2902   case Instruction::Alloca:
2903   case Instruction::Invoke:
2904   case Instruction::PHI:
2905   case Instruction::Store:
2906   case Instruction::Ret:
2907   case Instruction::Br:
2908   case Instruction::IndirectBr:
2909   case Instruction::Switch:
2910   case Instruction::Unreachable:
2911   case Instruction::Fence:
2912   case Instruction::LandingPad:
2913   case Instruction::AtomicRMW:
2914   case Instruction::AtomicCmpXchg:
2915   case Instruction::Resume:
2916     return false; // Misc instructions which have effects
2917   }
2918 }
2919 
2920 /// Return true if we know that the specified value is never null.
isKnownNonNull(const Value * V,const TargetLibraryInfo * TLI)2921 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2922   // Alloca never returns null, malloc might.
2923   if (isa<AllocaInst>(V)) return true;
2924 
2925   // A byval, inalloca, or nonnull argument is never null.
2926   if (const Argument *A = dyn_cast<Argument>(V))
2927     return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2928 
2929   // Global values are not null unless extern weak.
2930   if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2931     return !GV->hasExternalWeakLinkage();
2932 
2933   // A Load tagged w/nonnull metadata is never null.
2934   if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2935     return LI->getMetadata(LLVMContext::MD_nonnull);
2936 
2937   if (auto CS = ImmutableCallSite(V))
2938     if (CS.isReturnNonNull())
2939       return true;
2940 
2941   // operator new never returns null.
2942   if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
2943     return true;
2944 
2945   return false;
2946 }
2947 
computeOverflowForUnsignedMul(Value * LHS,Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)2948 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
2949                                                    const DataLayout &DL,
2950                                                    AssumptionCache *AC,
2951                                                    const Instruction *CxtI,
2952                                                    const DominatorTree *DT) {
2953   // Multiplying n * m significant bits yields a result of n + m significant
2954   // bits. If the total number of significant bits does not exceed the
2955   // result bit width (minus 1), there is no overflow.
2956   // This means if we have enough leading zero bits in the operands
2957   // we can guarantee that the result does not overflow.
2958   // Ref: "Hacker's Delight" by Henry Warren
2959   unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
2960   APInt LHSKnownZero(BitWidth, 0);
2961   APInt LHSKnownOne(BitWidth, 0);
2962   APInt RHSKnownZero(BitWidth, 0);
2963   APInt RHSKnownOne(BitWidth, 0);
2964   computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
2965                    DT);
2966   computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
2967                    DT);
2968   // Note that underestimating the number of zero bits gives a more
2969   // conservative answer.
2970   unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
2971                       RHSKnownZero.countLeadingOnes();
2972   // First handle the easy case: if we have enough zero bits there's
2973   // definitely no overflow.
2974   if (ZeroBits >= BitWidth)
2975     return OverflowResult::NeverOverflows;
2976 
2977   // Get the largest possible values for each operand.
2978   APInt LHSMax = ~LHSKnownZero;
2979   APInt RHSMax = ~RHSKnownZero;
2980 
2981   // We know the multiply operation doesn't overflow if the maximum values for
2982   // each operand will not overflow after we multiply them together.
2983   bool MaxOverflow;
2984   LHSMax.umul_ov(RHSMax, MaxOverflow);
2985   if (!MaxOverflow)
2986     return OverflowResult::NeverOverflows;
2987 
2988   // We know it always overflows if multiplying the smallest possible values for
2989   // the operands also results in overflow.
2990   bool MinOverflow;
2991   LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
2992   if (MinOverflow)
2993     return OverflowResult::AlwaysOverflows;
2994 
2995   return OverflowResult::MayOverflow;
2996 }
2997 
computeOverflowForUnsignedAdd(Value * LHS,Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)2998 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
2999                                                    const DataLayout &DL,
3000                                                    AssumptionCache *AC,
3001                                                    const Instruction *CxtI,
3002                                                    const DominatorTree *DT) {
3003   bool LHSKnownNonNegative, LHSKnownNegative;
3004   ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3005                  AC, CxtI, DT);
3006   if (LHSKnownNonNegative || LHSKnownNegative) {
3007     bool RHSKnownNonNegative, RHSKnownNegative;
3008     ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3009                    AC, CxtI, DT);
3010 
3011     if (LHSKnownNegative && RHSKnownNegative) {
3012       // The sign bit is set in both cases: this MUST overflow.
3013       // Create a simple add instruction, and insert it into the struct.
3014       return OverflowResult::AlwaysOverflows;
3015     }
3016 
3017     if (LHSKnownNonNegative && RHSKnownNonNegative) {
3018       // The sign bit is clear in both cases: this CANNOT overflow.
3019       // Create a simple add instruction, and insert it into the struct.
3020       return OverflowResult::NeverOverflows;
3021     }
3022   }
3023 
3024   return OverflowResult::MayOverflow;
3025 }
3026