1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
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/Analysis/LoopInfo.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/ConstantRange.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/GetElementPtrTypeIterator.h"
27 #include "llvm/IR/GlobalAlias.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/Metadata.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/Statepoint.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/MathExtras.h"
40 using namespace llvm::PatternMatch;
42 const unsigned MaxDepth = 6;
44 /// Enable an experimental feature to leverage information about dominating
45 /// conditions to compute known bits. The individual options below control how
46 /// hard we search. The defaults are chosen to be fairly aggressive. If you
47 /// run into compile time problems when testing, scale them back and report
49 static cl::opt<bool> EnableDomConditions("value-tracking-dom-conditions",
50 cl::Hidden, cl::init(false));
52 // This is expensive, so we only do it for the top level query value.
53 // (TODO: evaluate cost vs profit, consider higher thresholds)
54 static cl::opt<unsigned> DomConditionsMaxDepth("dom-conditions-max-depth",
55 cl::Hidden, cl::init(1));
57 /// How many dominating blocks should be scanned looking for dominating
59 static cl::opt<unsigned> DomConditionsMaxDomBlocks("dom-conditions-dom-blocks",
63 // Controls the number of uses of the value searched for possible
64 // dominating comparisons.
65 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
66 cl::Hidden, cl::init(20));
68 // If true, don't consider only compares whose only use is a branch.
69 static cl::opt<bool> DomConditionsSingleCmpUse("dom-conditions-single-cmp-use",
70 cl::Hidden, cl::init(false));
72 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
73 /// 0). For vector types, returns the element type's bitwidth.
74 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
75 if (unsigned BitWidth = Ty->getScalarSizeInBits())
78 return DL.getPointerTypeSizeInBits(Ty);
81 // Many of these functions have internal versions that take an assumption
82 // exclusion set. This is because of the potential for mutual recursion to
83 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
84 // classic case of this is assume(x = y), which will attempt to determine
85 // bits in x from bits in y, which will attempt to determine bits in y from
86 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
87 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
88 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
89 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
92 // Simplifying using an assume can only be done in a particular control-flow
93 // context (the context instruction provides that context). If an assume and
94 // the context instruction are not in the same block then the DT helps in
95 // figuring out if we can use it.
99 const Instruction *CxtI;
100 const DominatorTree *DT;
102 Query(AssumptionCache *AC = nullptr, const Instruction *CxtI = nullptr,
103 const DominatorTree *DT = nullptr)
104 : AC(AC), CxtI(CxtI), DT(DT) {}
106 Query(const Query &Q, const Value *NewExcl)
107 : ExclInvs(Q.ExclInvs), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT) {
108 ExclInvs.insert(NewExcl);
111 } // end anonymous namespace
113 // Given the provided Value and, potentially, a context instruction, return
114 // the preferred context instruction (if any).
115 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
116 // If we've been provided with a context instruction, then use that (provided
117 // it has been inserted).
118 if (CxtI && CxtI->getParent())
121 // If the value is really an already-inserted instruction, then use that.
122 CxtI = dyn_cast<Instruction>(V);
123 if (CxtI && CxtI->getParent())
129 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
130 const DataLayout &DL, unsigned Depth,
133 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
134 const DataLayout &DL, unsigned Depth,
135 AssumptionCache *AC, const Instruction *CxtI,
136 const DominatorTree *DT) {
137 ::computeKnownBits(V, KnownZero, KnownOne, DL, Depth,
138 Query(AC, safeCxtI(V, CxtI), DT));
141 bool llvm::haveNoCommonBitsSet(Value *LHS, Value *RHS, const DataLayout &DL,
142 AssumptionCache *AC, const Instruction *CxtI,
143 const DominatorTree *DT) {
144 assert(LHS->getType() == RHS->getType() &&
145 "LHS and RHS should have the same type");
146 assert(LHS->getType()->isIntOrIntVectorTy() &&
147 "LHS and RHS should be integers");
148 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
149 APInt LHSKnownZero(IT->getBitWidth(), 0), LHSKnownOne(IT->getBitWidth(), 0);
150 APInt RHSKnownZero(IT->getBitWidth(), 0), RHSKnownOne(IT->getBitWidth(), 0);
151 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, 0, AC, CxtI, DT);
152 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, 0, AC, CxtI, DT);
153 return (LHSKnownZero | RHSKnownZero).isAllOnesValue();
156 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
157 const DataLayout &DL, unsigned Depth,
160 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
161 const DataLayout &DL, unsigned Depth,
162 AssumptionCache *AC, const Instruction *CxtI,
163 const DominatorTree *DT) {
164 ::ComputeSignBit(V, KnownZero, KnownOne, DL, Depth,
165 Query(AC, safeCxtI(V, CxtI), DT));
168 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
169 const Query &Q, const DataLayout &DL);
171 bool llvm::isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, bool OrZero,
172 unsigned Depth, AssumptionCache *AC,
173 const Instruction *CxtI,
174 const DominatorTree *DT) {
175 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
176 Query(AC, safeCxtI(V, CxtI), DT), DL);
179 static bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
182 bool llvm::isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
183 AssumptionCache *AC, const Instruction *CxtI,
184 const DominatorTree *DT) {
185 return ::isKnownNonZero(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
188 bool llvm::isKnownNonNegative(Value *V, const DataLayout &DL, unsigned Depth,
189 AssumptionCache *AC, const Instruction *CxtI,
190 const DominatorTree *DT) {
191 bool NonNegative, Negative;
192 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
196 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
199 bool llvm::isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
200 AssumptionCache *AC, const Instruction *CxtI,
201 const DominatorTree *DT) {
202 return ::isKnownNonEqual(V1, V2, DL, Query(AC,
203 safeCxtI(V1, safeCxtI(V2, CxtI)),
207 static bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
208 unsigned Depth, const Query &Q);
210 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
211 unsigned Depth, AssumptionCache *AC,
212 const Instruction *CxtI, const DominatorTree *DT) {
213 return ::MaskedValueIsZero(V, Mask, DL, Depth,
214 Query(AC, safeCxtI(V, CxtI), DT));
217 static unsigned ComputeNumSignBits(Value *V, const DataLayout &DL,
218 unsigned Depth, const Query &Q);
220 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL,
221 unsigned Depth, AssumptionCache *AC,
222 const Instruction *CxtI,
223 const DominatorTree *DT) {
224 return ::ComputeNumSignBits(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
227 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
228 APInt &KnownZero, APInt &KnownOne,
229 APInt &KnownZero2, APInt &KnownOne2,
230 const DataLayout &DL, unsigned Depth,
233 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
234 // We know that the top bits of C-X are clear if X contains less bits
235 // than C (i.e. no wrap-around can happen). For example, 20-X is
236 // positive if we can prove that X is >= 0 and < 16.
237 if (!CLHS->getValue().isNegative()) {
238 unsigned BitWidth = KnownZero.getBitWidth();
239 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
240 // NLZ can't be BitWidth with no sign bit
241 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
242 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
244 // If all of the MaskV bits are known to be zero, then we know the
245 // output top bits are zero, because we now know that the output is
247 if ((KnownZero2 & MaskV) == MaskV) {
248 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
249 // Top bits known zero.
250 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
256 unsigned BitWidth = KnownZero.getBitWidth();
258 // If an initial sequence of bits in the result is not needed, the
259 // corresponding bits in the operands are not needed.
260 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
261 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, DL, Depth + 1, Q);
262 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
264 // Carry in a 1 for a subtract, rather than a 0.
265 APInt CarryIn(BitWidth, 0);
267 // Sum = LHS + ~RHS + 1
268 std::swap(KnownZero2, KnownOne2);
272 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
273 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
275 // Compute known bits of the carry.
276 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
277 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
279 // Compute set of known bits (where all three relevant bits are known).
280 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
281 APInt RHSKnown = KnownZero2 | KnownOne2;
282 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
283 APInt Known = LHSKnown & RHSKnown & CarryKnown;
285 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
286 "known bits of sum differ");
288 // Compute known bits of the result.
289 KnownZero = ~PossibleSumOne & Known;
290 KnownOne = PossibleSumOne & Known;
292 // Are we still trying to solve for the sign bit?
293 if (!Known.isNegative()) {
295 // Adding two non-negative numbers, or subtracting a negative number from
296 // a non-negative one, can't wrap into negative.
297 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
298 KnownZero |= APInt::getSignBit(BitWidth);
299 // Adding two negative numbers, or subtracting a non-negative number from
300 // a negative one, can't wrap into non-negative.
301 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
302 KnownOne |= APInt::getSignBit(BitWidth);
307 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
308 APInt &KnownZero, APInt &KnownOne,
309 APInt &KnownZero2, APInt &KnownOne2,
310 const DataLayout &DL, unsigned Depth,
312 unsigned BitWidth = KnownZero.getBitWidth();
313 computeKnownBits(Op1, KnownZero, KnownOne, DL, Depth + 1, Q);
314 computeKnownBits(Op0, KnownZero2, KnownOne2, DL, Depth + 1, Q);
316 bool isKnownNegative = false;
317 bool isKnownNonNegative = false;
318 // If the multiplication is known not to overflow, compute the sign bit.
321 // The product of a number with itself is non-negative.
322 isKnownNonNegative = true;
324 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
325 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
326 bool isKnownNegativeOp1 = KnownOne.isNegative();
327 bool isKnownNegativeOp0 = KnownOne2.isNegative();
328 // The product of two numbers with the same sign is non-negative.
329 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
330 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
331 // The product of a negative number and a non-negative number is either
333 if (!isKnownNonNegative)
334 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
335 isKnownNonZero(Op0, DL, Depth, Q)) ||
336 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
337 isKnownNonZero(Op1, DL, Depth, Q));
341 // If low bits are zero in either operand, output low known-0 bits.
342 // Also compute a conservative estimate for high known-0 bits.
343 // More trickiness is possible, but this is sufficient for the
344 // interesting case of alignment computation.
345 KnownOne.clearAllBits();
346 unsigned TrailZ = KnownZero.countTrailingOnes() +
347 KnownZero2.countTrailingOnes();
348 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
349 KnownZero2.countLeadingOnes(),
350 BitWidth) - BitWidth;
352 TrailZ = std::min(TrailZ, BitWidth);
353 LeadZ = std::min(LeadZ, BitWidth);
354 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
355 APInt::getHighBitsSet(BitWidth, LeadZ);
357 // Only make use of no-wrap flags if we failed to compute the sign bit
358 // directly. This matters if the multiplication always overflows, in
359 // which case we prefer to follow the result of the direct computation,
360 // though as the program is invoking undefined behaviour we can choose
361 // whatever we like here.
362 if (isKnownNonNegative && !KnownOne.isNegative())
363 KnownZero.setBit(BitWidth - 1);
364 else if (isKnownNegative && !KnownZero.isNegative())
365 KnownOne.setBit(BitWidth - 1);
368 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
370 unsigned BitWidth = KnownZero.getBitWidth();
371 unsigned NumRanges = Ranges.getNumOperands() / 2;
372 assert(NumRanges >= 1);
374 // Use the high end of the ranges to find leading zeros.
375 unsigned MinLeadingZeros = BitWidth;
376 for (unsigned i = 0; i < NumRanges; ++i) {
378 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
380 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
381 ConstantRange Range(Lower->getValue(), Upper->getValue());
382 if (Range.isWrappedSet())
383 MinLeadingZeros = 0; // -1 has no zeros
384 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
385 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
388 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
391 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
392 SmallVector<const Value *, 16> WorkSet(1, I);
393 SmallPtrSet<const Value *, 32> Visited;
394 SmallPtrSet<const Value *, 16> EphValues;
396 while (!WorkSet.empty()) {
397 const Value *V = WorkSet.pop_back_val();
398 if (!Visited.insert(V).second)
401 // If all uses of this value are ephemeral, then so is this value.
402 bool FoundNEUse = false;
403 for (const User *I : V->users())
404 if (!EphValues.count(I)) {
414 if (const User *U = dyn_cast<User>(V))
415 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
417 if (isSafeToSpeculativelyExecute(*J))
418 WorkSet.push_back(*J);
426 // Is this an intrinsic that cannot be speculated but also cannot trap?
427 static bool isAssumeLikeIntrinsic(const Instruction *I) {
428 if (const CallInst *CI = dyn_cast<CallInst>(I))
429 if (Function *F = CI->getCalledFunction())
430 switch (F->getIntrinsicID()) {
432 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
433 case Intrinsic::assume:
434 case Intrinsic::dbg_declare:
435 case Intrinsic::dbg_value:
436 case Intrinsic::invariant_start:
437 case Intrinsic::invariant_end:
438 case Intrinsic::lifetime_start:
439 case Intrinsic::lifetime_end:
440 case Intrinsic::objectsize:
441 case Intrinsic::ptr_annotation:
442 case Intrinsic::var_annotation:
449 static bool isValidAssumeForContext(Value *V, const Query &Q) {
450 Instruction *Inv = cast<Instruction>(V);
452 // There are two restrictions on the use of an assume:
453 // 1. The assume must dominate the context (or the control flow must
454 // reach the assume whenever it reaches the context).
455 // 2. The context must not be in the assume's set of ephemeral values
456 // (otherwise we will use the assume to prove that the condition
457 // feeding the assume is trivially true, thus causing the removal of
461 if (Q.DT->dominates(Inv, Q.CxtI)) {
463 } else if (Inv->getParent() == Q.CxtI->getParent()) {
464 // The context comes first, but they're both in the same block. Make sure
465 // there is nothing in between that might interrupt the control flow.
466 for (BasicBlock::const_iterator I =
467 std::next(BasicBlock::const_iterator(Q.CxtI)),
468 IE(Inv); I != IE; ++I)
469 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
472 return !isEphemeralValueOf(Inv, Q.CxtI);
478 // When we don't have a DT, we do a limited search...
479 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
481 } else if (Inv->getParent() == Q.CxtI->getParent()) {
482 // Search forward from the assume until we reach the context (or the end
483 // of the block); the common case is that the assume will come first.
484 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
485 IE = Inv->getParent()->end(); I != IE; ++I)
489 // The context must come first...
490 for (BasicBlock::const_iterator I =
491 std::next(BasicBlock::const_iterator(Q.CxtI)),
492 IE(Inv); I != IE; ++I)
493 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
496 return !isEphemeralValueOf(Inv, Q.CxtI);
502 bool llvm::isValidAssumeForContext(const Instruction *I,
503 const Instruction *CxtI,
504 const DominatorTree *DT) {
505 return ::isValidAssumeForContext(const_cast<Instruction *>(I),
506 Query(nullptr, CxtI, DT));
509 template<typename LHS, typename RHS>
510 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
511 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
512 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
513 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
516 template<typename LHS, typename RHS>
517 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
518 BinaryOp_match<RHS, LHS, Instruction::And>>
519 m_c_And(const LHS &L, const RHS &R) {
520 return m_CombineOr(m_And(L, R), m_And(R, L));
523 template<typename LHS, typename RHS>
524 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
525 BinaryOp_match<RHS, LHS, Instruction::Or>>
526 m_c_Or(const LHS &L, const RHS &R) {
527 return m_CombineOr(m_Or(L, R), m_Or(R, L));
530 template<typename LHS, typename RHS>
531 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
532 BinaryOp_match<RHS, LHS, Instruction::Xor>>
533 m_c_Xor(const LHS &L, const RHS &R) {
534 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
537 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
538 /// true (at the context instruction.) This is mostly a utility function for
539 /// the prototype dominating conditions reasoning below.
540 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
543 const DataLayout &DL,
544 unsigned Depth, const Query &Q) {
545 Value *LHS = Cmp->getOperand(0);
546 Value *RHS = Cmp->getOperand(1);
547 // TODO: We could potentially be more aggressive here. This would be worth
548 // evaluating. If we can, explore commoning this code with the assume
550 if (LHS != V && RHS != V)
553 const unsigned BitWidth = KnownZero.getBitWidth();
555 switch (Cmp->getPredicate()) {
557 // We know nothing from this condition
559 // TODO: implement unsigned bound from below (known one bits)
560 // TODO: common condition check implementations with assumes
561 // TODO: implement other patterns from assume (e.g. V & B == A)
562 case ICmpInst::ICMP_SGT:
564 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
565 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
566 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
567 // We know that the sign bit is zero.
568 KnownZero |= APInt::getSignBit(BitWidth);
572 case ICmpInst::ICMP_EQ:
574 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
576 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
578 computeKnownBits(LHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
580 llvm_unreachable("missing use?");
581 KnownZero |= KnownZeroTemp;
582 KnownOne |= KnownOneTemp;
585 case ICmpInst::ICMP_ULE:
587 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
588 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
589 // The known zero bits carry over
590 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
591 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
594 case ICmpInst::ICMP_ULT:
596 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
597 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
598 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
599 // power of 2, then one more).
600 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
601 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
603 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
609 /// Compute known bits in 'V' from conditions which are known to be true along
610 /// all paths leading to the context instruction. In particular, look for
611 /// cases where one branch of an interesting condition dominates the context
612 /// instruction. This does not do general dataflow.
613 /// NOTE: This code is EXPERIMENTAL and currently off by default.
614 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
616 const DataLayout &DL,
619 // Need both the dominator tree and the query location to do anything useful
620 if (!Q.DT || !Q.CxtI)
622 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
623 // The context instruction might be in a statically unreachable block. If
624 // so, asking dominator queries may yield suprising results. (e.g. the block
625 // may not have a dom tree node)
626 if (!Q.DT->isReachableFromEntry(Cxt->getParent()))
629 // Avoid useless work
630 if (auto VI = dyn_cast<Instruction>(V))
631 if (VI->getParent() == Cxt->getParent())
634 // Note: We currently implement two options. It's not clear which of these
635 // will survive long term, we need data for that.
636 // Option 1 - Try walking the dominator tree looking for conditions which
637 // might apply. This works well for local conditions (loop guards, etc..),
638 // but not as well for things far from the context instruction (presuming a
639 // low max blocks explored). If we can set an high enough limit, this would
641 // Option 2 - We restrict out search to those conditions which are uses of
642 // the value we're interested in. This is independent of dom structure,
643 // but is slightly less powerful without looking through lots of use chains.
644 // It does handle conditions far from the context instruction (e.g. early
645 // function exits on entry) really well though.
647 // Option 1 - Search the dom tree
648 unsigned NumBlocksExplored = 0;
649 BasicBlock *Current = Cxt->getParent();
651 // Stop searching if we've gone too far up the chain
652 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
656 if (!Q.DT->getNode(Current)->getIDom())
658 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
660 // found function entry
663 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
664 if (!BI || BI->isUnconditional())
666 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
670 // We're looking for conditions that are guaranteed to hold at the context
671 // instruction. Finding a condition where one path dominates the context
672 // isn't enough because both the true and false cases could merge before
673 // the context instruction we're actually interested in. Instead, we need
674 // to ensure that the taken *edge* dominates the context instruction. We
675 // know that the edge must be reachable since we started from a reachable
677 BasicBlock *BB0 = BI->getSuccessor(0);
678 BasicBlockEdge Edge(BI->getParent(), BB0);
679 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
682 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
686 // Option 2 - Search the other uses of V
687 unsigned NumUsesExplored = 0;
688 for (auto U : V->users()) {
689 // Avoid massive lists
690 if (NumUsesExplored >= DomConditionsMaxUses)
693 // Consider only compare instructions uniquely controlling a branch
694 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
698 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
701 for (auto *CmpU : Cmp->users()) {
702 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
703 if (!BI || BI->isUnconditional())
705 // We're looking for conditions that are guaranteed to hold at the
706 // context instruction. Finding a condition where one path dominates
707 // the context isn't enough because both the true and false cases could
708 // merge before the context instruction we're actually interested in.
709 // Instead, we need to ensure that the taken *edge* dominates the context
711 BasicBlock *BB0 = BI->getSuccessor(0);
712 BasicBlockEdge Edge(BI->getParent(), BB0);
713 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
716 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
722 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
723 APInt &KnownOne, const DataLayout &DL,
724 unsigned Depth, const Query &Q) {
725 // Use of assumptions is context-sensitive. If we don't have a context, we
727 if (!Q.AC || !Q.CxtI)
730 unsigned BitWidth = KnownZero.getBitWidth();
732 for (auto &AssumeVH : Q.AC->assumptions()) {
735 CallInst *I = cast<CallInst>(AssumeVH);
736 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
737 "Got assumption for the wrong function!");
738 if (Q.ExclInvs.count(I))
741 // Warning: This loop can end up being somewhat performance sensetive.
742 // We're running this loop for once for each value queried resulting in a
743 // runtime of ~O(#assumes * #values).
745 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
746 "must be an assume intrinsic");
748 Value *Arg = I->getArgOperand(0);
750 if (Arg == V && isValidAssumeForContext(I, Q)) {
751 assert(BitWidth == 1 && "assume operand is not i1?");
752 KnownZero.clearAllBits();
753 KnownOne.setAllBits();
757 // The remaining tests are all recursive, so bail out if we hit the limit.
758 if (Depth == MaxDepth)
762 auto m_V = m_CombineOr(m_Specific(V),
763 m_CombineOr(m_PtrToInt(m_Specific(V)),
764 m_BitCast(m_Specific(V))));
766 CmpInst::Predicate Pred;
769 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
770 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
771 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
772 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
773 KnownZero |= RHSKnownZero;
774 KnownOne |= RHSKnownOne;
776 } else if (match(Arg,
777 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
778 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
779 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
780 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
781 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
782 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
784 // For those bits in the mask that are known to be one, we can propagate
785 // known bits from the RHS to V.
786 KnownZero |= RHSKnownZero & MaskKnownOne;
787 KnownOne |= RHSKnownOne & MaskKnownOne;
788 // assume(~(v & b) = a)
789 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
791 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
792 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
793 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
794 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
795 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
797 // For those bits in the mask that are known to be one, we can propagate
798 // inverted known bits from the RHS to V.
799 KnownZero |= RHSKnownOne & MaskKnownOne;
800 KnownOne |= RHSKnownZero & MaskKnownOne;
802 } else if (match(Arg,
803 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
804 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
805 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
806 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
807 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
808 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
810 // For those bits in B that are known to be zero, we can propagate known
811 // bits from the RHS to V.
812 KnownZero |= RHSKnownZero & BKnownZero;
813 KnownOne |= RHSKnownOne & BKnownZero;
814 // assume(~(v | b) = a)
815 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
817 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
818 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
819 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
820 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
821 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
823 // For those bits in B that are known to be zero, we can propagate
824 // inverted known bits from the RHS to V.
825 KnownZero |= RHSKnownOne & BKnownZero;
826 KnownOne |= RHSKnownZero & BKnownZero;
828 } else if (match(Arg,
829 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
830 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
831 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
832 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
833 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
834 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
836 // For those bits in B that are known to be zero, we can propagate known
837 // bits from the RHS to V. For those bits in B that are known to be one,
838 // we can propagate inverted known bits from the RHS to V.
839 KnownZero |= RHSKnownZero & BKnownZero;
840 KnownOne |= RHSKnownOne & BKnownZero;
841 KnownZero |= RHSKnownOne & BKnownOne;
842 KnownOne |= RHSKnownZero & BKnownOne;
843 // assume(~(v ^ b) = a)
844 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
846 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
847 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
848 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
849 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
850 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
852 // For those bits in B that are known to be zero, we can propagate
853 // inverted known bits from the RHS to V. For those bits in B that are
854 // known to be one, we can propagate known bits from the RHS to V.
855 KnownZero |= RHSKnownOne & BKnownZero;
856 KnownOne |= RHSKnownZero & BKnownZero;
857 KnownZero |= RHSKnownZero & BKnownOne;
858 KnownOne |= RHSKnownOne & BKnownOne;
859 // assume(v << c = a)
860 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
862 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
863 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
864 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
865 // For those bits in RHS that are known, we can propagate them to known
866 // bits in V shifted to the right by C.
867 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
868 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
869 // assume(~(v << c) = a)
870 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
872 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
873 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
874 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
875 // For those bits in RHS that are known, we can propagate them inverted
876 // to known bits in V shifted to the right by C.
877 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
878 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
879 // assume(v >> c = a)
880 } else if (match(Arg,
881 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
882 m_AShr(m_V, m_ConstantInt(C))),
884 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
885 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
886 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
887 // For those bits in RHS that are known, we can propagate them to known
888 // bits in V shifted to the right by C.
889 KnownZero |= RHSKnownZero << C->getZExtValue();
890 KnownOne |= RHSKnownOne << C->getZExtValue();
891 // assume(~(v >> c) = a)
892 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
893 m_LShr(m_V, m_ConstantInt(C)),
894 m_AShr(m_V, m_ConstantInt(C)))),
896 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
897 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
898 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
899 // For those bits in RHS that are known, we can propagate them inverted
900 // to known bits in V shifted to the right by C.
901 KnownZero |= RHSKnownOne << C->getZExtValue();
902 KnownOne |= RHSKnownZero << C->getZExtValue();
903 // assume(v >=_s c) where c is non-negative
904 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
905 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
906 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
907 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
909 if (RHSKnownZero.isNegative()) {
910 // We know that the sign bit is zero.
911 KnownZero |= APInt::getSignBit(BitWidth);
913 // assume(v >_s c) where c is at least -1.
914 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
915 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
916 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
917 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
919 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
920 // We know that the sign bit is zero.
921 KnownZero |= APInt::getSignBit(BitWidth);
923 // assume(v <=_s c) where c is negative
924 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
925 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
926 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
927 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
929 if (RHSKnownOne.isNegative()) {
930 // We know that the sign bit is one.
931 KnownOne |= APInt::getSignBit(BitWidth);
933 // assume(v <_s c) where c is non-positive
934 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
935 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
936 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
937 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
939 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
940 // We know that the sign bit is one.
941 KnownOne |= APInt::getSignBit(BitWidth);
944 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
945 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
946 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
947 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
949 // Whatever high bits in c are zero are known to be zero.
951 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
953 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
954 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
955 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
956 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
958 // Whatever high bits in c are zero are known to be zero (if c is a power
959 // of 2, then one more).
960 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
962 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
965 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
970 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
971 APInt &KnownOne, const DataLayout &DL,
972 unsigned Depth, const Query &Q) {
973 unsigned BitWidth = KnownZero.getBitWidth();
975 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
976 switch (I->getOpcode()) {
978 case Instruction::Load:
979 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
980 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
982 case Instruction::And: {
983 // If either the LHS or the RHS are Zero, the result is zero.
984 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
985 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
987 // Output known-1 bits are only known if set in both the LHS & RHS.
988 KnownOne &= KnownOne2;
989 // Output known-0 are known to be clear if zero in either the LHS | RHS.
990 KnownZero |= KnownZero2;
993 case Instruction::Or: {
994 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
995 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
997 // Output known-0 bits are only known if clear in both the LHS & RHS.
998 KnownZero &= KnownZero2;
999 // Output known-1 are known to be set if set in either the LHS | RHS.
1000 KnownOne |= KnownOne2;
1003 case Instruction::Xor: {
1004 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1005 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1007 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1008 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1009 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1010 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1011 KnownZero = KnownZeroOut;
1014 case Instruction::Mul: {
1015 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1016 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1017 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1020 case Instruction::UDiv: {
1021 // For the purposes of computing leading zeros we can conservatively
1022 // treat a udiv as a logical right shift by the power of 2 known to
1023 // be less than the denominator.
1024 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1025 unsigned LeadZ = KnownZero2.countLeadingOnes();
1027 KnownOne2.clearAllBits();
1028 KnownZero2.clearAllBits();
1029 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1030 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1031 if (RHSUnknownLeadingOnes != BitWidth)
1032 LeadZ = std::min(BitWidth,
1033 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1035 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1038 case Instruction::Select:
1039 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1040 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1042 // Only known if known in both the LHS and RHS.
1043 KnownOne &= KnownOne2;
1044 KnownZero &= KnownZero2;
1046 case Instruction::FPTrunc:
1047 case Instruction::FPExt:
1048 case Instruction::FPToUI:
1049 case Instruction::FPToSI:
1050 case Instruction::SIToFP:
1051 case Instruction::UIToFP:
1052 break; // Can't work with floating point.
1053 case Instruction::PtrToInt:
1054 case Instruction::IntToPtr:
1055 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1056 // FALL THROUGH and handle them the same as zext/trunc.
1057 case Instruction::ZExt:
1058 case Instruction::Trunc: {
1059 Type *SrcTy = I->getOperand(0)->getType();
1061 unsigned SrcBitWidth;
1062 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1063 // which fall through here.
1064 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1066 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1067 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1068 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1069 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1070 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1071 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1072 // Any top bits are known to be zero.
1073 if (BitWidth > SrcBitWidth)
1074 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1077 case Instruction::BitCast: {
1078 Type *SrcTy = I->getOperand(0)->getType();
1079 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy() ||
1080 SrcTy->isFloatingPointTy()) &&
1081 // TODO: For now, not handling conversions like:
1082 // (bitcast i64 %x to <2 x i32>)
1083 !I->getType()->isVectorTy()) {
1084 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1089 case Instruction::SExt: {
1090 // Compute the bits in the result that are not present in the input.
1091 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1093 KnownZero = KnownZero.trunc(SrcBitWidth);
1094 KnownOne = KnownOne.trunc(SrcBitWidth);
1095 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1096 KnownZero = KnownZero.zext(BitWidth);
1097 KnownOne = KnownOne.zext(BitWidth);
1099 // If the sign bit of the input is known set or clear, then we know the
1100 // top bits of the result.
1101 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1102 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1103 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1104 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1107 case Instruction::Shl:
1108 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1109 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1110 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1111 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1112 KnownZero <<= ShiftAmt;
1113 KnownOne <<= ShiftAmt;
1114 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1117 case Instruction::LShr:
1118 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1119 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1120 // Compute the new bits that are at the top now.
1121 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1123 // Unsigned shift right.
1124 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1125 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1126 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1127 // high bits known zero.
1128 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1131 case Instruction::AShr:
1132 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1133 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1134 // Compute the new bits that are at the top now.
1135 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1137 // Signed shift right.
1138 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1139 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1140 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1142 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1143 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1144 KnownZero |= HighBits;
1145 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1146 KnownOne |= HighBits;
1149 case Instruction::Sub: {
1150 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1151 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1152 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1156 case Instruction::Add: {
1157 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1158 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1159 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1163 case Instruction::SRem:
1164 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1165 APInt RA = Rem->getValue().abs();
1166 if (RA.isPowerOf2()) {
1167 APInt LowBits = RA - 1;
1168 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1171 // The low bits of the first operand are unchanged by the srem.
1172 KnownZero = KnownZero2 & LowBits;
1173 KnownOne = KnownOne2 & LowBits;
1175 // If the first operand is non-negative or has all low bits zero, then
1176 // the upper bits are all zero.
1177 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1178 KnownZero |= ~LowBits;
1180 // If the first operand is negative and not all low bits are zero, then
1181 // the upper bits are all one.
1182 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1183 KnownOne |= ~LowBits;
1185 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1189 // The sign bit is the LHS's sign bit, except when the result of the
1190 // remainder is zero.
1191 if (KnownZero.isNonNegative()) {
1192 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1193 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1195 // If it's known zero, our sign bit is also zero.
1196 if (LHSKnownZero.isNegative())
1197 KnownZero.setBit(BitWidth - 1);
1201 case Instruction::URem: {
1202 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1203 APInt RA = Rem->getValue();
1204 if (RA.isPowerOf2()) {
1205 APInt LowBits = (RA - 1);
1206 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1208 KnownZero |= ~LowBits;
1209 KnownOne &= LowBits;
1214 // Since the result is less than or equal to either operand, any leading
1215 // zero bits in either operand must also exist in the result.
1216 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1217 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1219 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1220 KnownZero2.countLeadingOnes());
1221 KnownOne.clearAllBits();
1222 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1226 case Instruction::Alloca: {
1227 AllocaInst *AI = cast<AllocaInst>(I);
1228 unsigned Align = AI->getAlignment();
1230 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1233 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1236 case Instruction::GetElementPtr: {
1237 // Analyze all of the subscripts of this getelementptr instruction
1238 // to determine if we can prove known low zero bits.
1239 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1240 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1242 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1244 gep_type_iterator GTI = gep_type_begin(I);
1245 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1246 Value *Index = I->getOperand(i);
1247 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1248 // Handle struct member offset arithmetic.
1250 // Handle case when index is vector zeroinitializer
1251 Constant *CIndex = cast<Constant>(Index);
1252 if (CIndex->isZeroValue())
1255 if (CIndex->getType()->isVectorTy())
1256 Index = CIndex->getSplatValue();
1258 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1259 const StructLayout *SL = DL.getStructLayout(STy);
1260 uint64_t Offset = SL->getElementOffset(Idx);
1261 TrailZ = std::min<unsigned>(TrailZ,
1262 countTrailingZeros(Offset));
1264 // Handle array index arithmetic.
1265 Type *IndexedTy = GTI.getIndexedType();
1266 if (!IndexedTy->isSized()) {
1270 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1271 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1272 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1273 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1275 TrailZ = std::min(TrailZ,
1276 unsigned(countTrailingZeros(TypeSize) +
1277 LocalKnownZero.countTrailingOnes()));
1281 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1284 case Instruction::PHI: {
1285 PHINode *P = cast<PHINode>(I);
1286 // Handle the case of a simple two-predecessor recurrence PHI.
1287 // There's a lot more that could theoretically be done here, but
1288 // this is sufficient to catch some interesting cases.
1289 if (P->getNumIncomingValues() == 2) {
1290 for (unsigned i = 0; i != 2; ++i) {
1291 Value *L = P->getIncomingValue(i);
1292 Value *R = P->getIncomingValue(!i);
1293 Operator *LU = dyn_cast<Operator>(L);
1296 unsigned Opcode = LU->getOpcode();
1297 // Check for operations that have the property that if
1298 // both their operands have low zero bits, the result
1299 // will have low zero bits.
1300 if (Opcode == Instruction::Add ||
1301 Opcode == Instruction::Sub ||
1302 Opcode == Instruction::And ||
1303 Opcode == Instruction::Or ||
1304 Opcode == Instruction::Mul) {
1305 Value *LL = LU->getOperand(0);
1306 Value *LR = LU->getOperand(1);
1307 // Find a recurrence.
1314 // Ok, we have a PHI of the form L op= R. Check for low
1316 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1318 // We need to take the minimum number of known bits
1319 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1320 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1322 KnownZero = APInt::getLowBitsSet(BitWidth,
1323 std::min(KnownZero2.countTrailingOnes(),
1324 KnownZero3.countTrailingOnes()));
1330 // Unreachable blocks may have zero-operand PHI nodes.
1331 if (P->getNumIncomingValues() == 0)
1334 // Otherwise take the unions of the known bit sets of the operands,
1335 // taking conservative care to avoid excessive recursion.
1336 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1337 // Skip if every incoming value references to ourself.
1338 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1341 KnownZero = APInt::getAllOnesValue(BitWidth);
1342 KnownOne = APInt::getAllOnesValue(BitWidth);
1343 for (Value *IncValue : P->incoming_values()) {
1344 // Skip direct self references.
1345 if (IncValue == P) continue;
1347 KnownZero2 = APInt(BitWidth, 0);
1348 KnownOne2 = APInt(BitWidth, 0);
1349 // Recurse, but cap the recursion to one level, because we don't
1350 // want to waste time spinning around in loops.
1351 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1353 KnownZero &= KnownZero2;
1354 KnownOne &= KnownOne2;
1355 // If all bits have been ruled out, there's no need to check
1357 if (!KnownZero && !KnownOne)
1363 case Instruction::Call:
1364 case Instruction::Invoke:
1365 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1366 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1367 // If a range metadata is attached to this IntrinsicInst, intersect the
1368 // explicit range specified by the metadata and the implicit range of
1370 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1371 switch (II->getIntrinsicID()) {
1373 case Intrinsic::bswap:
1374 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1376 KnownZero |= KnownZero2.byteSwap();
1377 KnownOne |= KnownOne2.byteSwap();
1379 case Intrinsic::ctlz:
1380 case Intrinsic::cttz: {
1381 unsigned LowBits = Log2_32(BitWidth)+1;
1382 // If this call is undefined for 0, the result will be less than 2^n.
1383 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1385 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1388 case Intrinsic::ctpop: {
1389 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1391 // We can bound the space the count needs. Also, bits known to be zero
1392 // can't contribute to the population.
1393 unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation();
1394 unsigned LeadingZeros =
1395 APInt(BitWidth, BitsPossiblySet).countLeadingZeros();
1396 assert(LeadingZeros <= BitWidth);
1397 KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros);
1398 KnownOne &= ~KnownZero;
1399 // TODO: we could bound KnownOne using the lower bound on the number
1400 // of bits which might be set provided by popcnt KnownOne2.
1403 case Intrinsic::fabs: {
1404 Type *Ty = II->getType();
1405 APInt SignBit = APInt::getSignBit(Ty->getScalarSizeInBits());
1406 KnownZero |= APInt::getSplat(Ty->getPrimitiveSizeInBits(), SignBit);
1409 case Intrinsic::x86_sse42_crc32_64_64:
1410 KnownZero |= APInt::getHighBitsSet(64, 32);
1415 case Instruction::ExtractValue:
1416 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1417 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1418 if (EVI->getNumIndices() != 1) break;
1419 if (EVI->getIndices()[0] == 0) {
1420 switch (II->getIntrinsicID()) {
1422 case Intrinsic::uadd_with_overflow:
1423 case Intrinsic::sadd_with_overflow:
1424 computeKnownBitsAddSub(true, II->getArgOperand(0),
1425 II->getArgOperand(1), false, KnownZero,
1426 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1428 case Intrinsic::usub_with_overflow:
1429 case Intrinsic::ssub_with_overflow:
1430 computeKnownBitsAddSub(false, II->getArgOperand(0),
1431 II->getArgOperand(1), false, KnownZero,
1432 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1434 case Intrinsic::umul_with_overflow:
1435 case Intrinsic::smul_with_overflow:
1436 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1437 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1446 static unsigned getAlignment(const Value *V, const DataLayout &DL) {
1448 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1449 Align = GO->getAlignment();
1451 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1452 Type *ObjectType = GVar->getType()->getElementType();
1453 if (ObjectType->isSized()) {
1454 // If the object is defined in the current Module, we'll be giving
1455 // it the preferred alignment. Otherwise, we have to assume that it
1456 // may only have the minimum ABI alignment.
1457 if (GVar->isStrongDefinitionForLinker())
1458 Align = DL.getPreferredAlignment(GVar);
1460 Align = DL.getABITypeAlignment(ObjectType);
1464 } else if (const Argument *A = dyn_cast<Argument>(V)) {
1465 Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1467 if (!Align && A->hasStructRetAttr()) {
1468 // An sret parameter has at least the ABI alignment of the return type.
1469 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1470 if (EltTy->isSized())
1471 Align = DL.getABITypeAlignment(EltTy);
1473 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
1474 Align = AI->getAlignment();
1475 else if (auto CS = ImmutableCallSite(V))
1476 Align = CS.getAttributes().getParamAlignment(AttributeSet::ReturnIndex);
1477 else if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1478 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_align)) {
1479 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
1480 Align = CI->getLimitedValue();
1486 /// Determine which bits of V are known to be either zero or one and return
1487 /// them in the KnownZero/KnownOne bit sets.
1489 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1490 /// we cannot optimize based on the assumption that it is zero without changing
1491 /// it to be an explicit zero. If we don't change it to zero, other code could
1492 /// optimized based on the contradictory assumption that it is non-zero.
1493 /// Because instcombine aggressively folds operations with undef args anyway,
1494 /// this won't lose us code quality.
1496 /// This function is defined on values with integer type, values with pointer
1497 /// type, and vectors of integers. In the case
1498 /// where V is a vector, known zero, and known one values are the
1499 /// same width as the vector element, and the bit is set only if it is true
1500 /// for all of the elements in the vector.
1501 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1502 const DataLayout &DL, unsigned Depth, const Query &Q) {
1503 assert(V && "No Value?");
1504 assert(Depth <= MaxDepth && "Limit Search Depth");
1505 unsigned BitWidth = KnownZero.getBitWidth();
1507 assert((V->getType()->isIntOrIntVectorTy() ||
1508 V->getType()->isFPOrFPVectorTy() ||
1509 V->getType()->getScalarType()->isPointerTy()) &&
1510 "Not integer, floating point, or pointer type!");
1511 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1512 (!V->getType()->isIntOrIntVectorTy() ||
1513 V->getType()->getScalarSizeInBits() == BitWidth) &&
1514 KnownZero.getBitWidth() == BitWidth &&
1515 KnownOne.getBitWidth() == BitWidth &&
1516 "V, KnownOne and KnownZero should have same BitWidth");
1518 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1519 // We know all of the bits for a constant!
1520 KnownOne = CI->getValue();
1521 KnownZero = ~KnownOne;
1524 // Null and aggregate-zero are all-zeros.
1525 if (isa<ConstantPointerNull>(V) ||
1526 isa<ConstantAggregateZero>(V)) {
1527 KnownOne.clearAllBits();
1528 KnownZero = APInt::getAllOnesValue(BitWidth);
1531 // Handle a constant vector by taking the intersection of the known bits of
1532 // each element. There is no real need to handle ConstantVector here, because
1533 // we don't handle undef in any particularly useful way.
1534 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1535 // We know that CDS must be a vector of integers. Take the intersection of
1537 KnownZero.setAllBits(); KnownOne.setAllBits();
1538 APInt Elt(KnownZero.getBitWidth(), 0);
1539 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1540 Elt = CDS->getElementAsInteger(i);
1547 // Start out not knowing anything.
1548 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1550 // Limit search depth.
1551 // All recursive calls that increase depth must come after this.
1552 if (Depth == MaxDepth)
1555 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1556 // the bits of its aliasee.
1557 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1558 if (!GA->mayBeOverridden())
1559 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1563 if (Operator *I = dyn_cast<Operator>(V))
1564 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1566 // Aligned pointers have trailing zeros - refine KnownZero set
1567 if (V->getType()->isPointerTy()) {
1568 unsigned Align = getAlignment(V, DL);
1570 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1573 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1574 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1575 // computeKnownBitsFromOperator.
1577 // Check whether a nearby assume intrinsic can determine some known bits.
1578 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1580 // Check whether there's a dominating condition which implies something about
1581 // this value at the given context.
1582 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1583 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1586 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1589 /// Determine whether the sign bit is known to be zero or one.
1590 /// Convenience wrapper around computeKnownBits.
1591 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1592 const DataLayout &DL, unsigned Depth, const Query &Q) {
1593 unsigned BitWidth = getBitWidth(V->getType(), DL);
1599 APInt ZeroBits(BitWidth, 0);
1600 APInt OneBits(BitWidth, 0);
1601 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1602 KnownOne = OneBits[BitWidth - 1];
1603 KnownZero = ZeroBits[BitWidth - 1];
1606 /// Return true if the given value is known to have exactly one
1607 /// bit set when defined. For vectors return true if every element is known to
1608 /// be a power of two when defined. Supports values with integer or pointer
1609 /// types and vectors of integers.
1610 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1611 const Query &Q, const DataLayout &DL) {
1612 if (Constant *C = dyn_cast<Constant>(V)) {
1613 if (C->isNullValue())
1615 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1616 return CI->getValue().isPowerOf2();
1617 // TODO: Handle vector constants.
1620 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1621 // it is shifted off the end then the result is undefined.
1622 if (match(V, m_Shl(m_One(), m_Value())))
1625 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1626 // bottom. If it is shifted off the bottom then the result is undefined.
1627 if (match(V, m_LShr(m_SignBit(), m_Value())))
1630 // The remaining tests are all recursive, so bail out if we hit the limit.
1631 if (Depth++ == MaxDepth)
1634 Value *X = nullptr, *Y = nullptr;
1635 // A shift of a power of two is a power of two or zero.
1636 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1637 match(V, m_Shr(m_Value(X), m_Value()))))
1638 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1640 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1641 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1643 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1644 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1645 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1647 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1648 // A power of two and'd with anything is a power of two or zero.
1649 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1650 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1652 // X & (-X) is always a power of two or zero.
1653 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1658 // Adding a power-of-two or zero to the same power-of-two or zero yields
1659 // either the original power-of-two, a larger power-of-two or zero.
1660 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1661 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1662 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1663 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1664 match(X, m_And(m_Value(), m_Specific(Y))))
1665 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1667 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1668 match(Y, m_And(m_Value(), m_Specific(X))))
1669 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1672 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1673 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1674 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1676 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1677 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1678 // If i8 V is a power of two or zero:
1679 // ZeroBits: 1 1 1 0 1 1 1 1
1680 // ~ZeroBits: 0 0 0 1 0 0 0 0
1681 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1682 // If OrZero isn't set, we cannot give back a zero result.
1683 // Make sure either the LHS or RHS has a bit set.
1684 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1689 // An exact divide or right shift can only shift off zero bits, so the result
1690 // is a power of two only if the first operand is a power of two and not
1691 // copying a sign bit (sdiv int_min, 2).
1692 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1693 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1694 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1701 /// \brief Test whether a GEP's result is known to be non-null.
1703 /// Uses properties inherent in a GEP to try to determine whether it is known
1706 /// Currently this routine does not support vector GEPs.
1707 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1708 unsigned Depth, const Query &Q) {
1709 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1712 // FIXME: Support vector-GEPs.
1713 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1715 // If the base pointer is non-null, we cannot walk to a null address with an
1716 // inbounds GEP in address space zero.
1717 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1720 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1721 // If so, then the GEP cannot produce a null pointer, as doing so would
1722 // inherently violate the inbounds contract within address space zero.
1723 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1724 GTI != GTE; ++GTI) {
1725 // Struct types are easy -- they must always be indexed by a constant.
1726 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1727 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1728 unsigned ElementIdx = OpC->getZExtValue();
1729 const StructLayout *SL = DL.getStructLayout(STy);
1730 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1731 if (ElementOffset > 0)
1736 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1737 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1740 // Fast path the constant operand case both for efficiency and so we don't
1741 // increment Depth when just zipping down an all-constant GEP.
1742 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1748 // We post-increment Depth here because while isKnownNonZero increments it
1749 // as well, when we pop back up that increment won't persist. We don't want
1750 // to recurse 10k times just because we have 10k GEP operands. We don't
1751 // bail completely out because we want to handle constant GEPs regardless
1753 if (Depth++ >= MaxDepth)
1756 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1763 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1764 /// ensure that the value it's attached to is never Value? 'RangeType' is
1765 /// is the type of the value described by the range.
1766 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1767 const APInt& Value) {
1768 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1769 assert(NumRanges >= 1);
1770 for (unsigned i = 0; i < NumRanges; ++i) {
1771 ConstantInt *Lower =
1772 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1773 ConstantInt *Upper =
1774 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1775 ConstantRange Range(Lower->getValue(), Upper->getValue());
1776 if (Range.contains(Value))
1782 /// Return true if the given value is known to be non-zero when defined.
1783 /// For vectors return true if every element is known to be non-zero when
1784 /// defined. Supports values with integer or pointer type and vectors of
1786 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1788 if (Constant *C = dyn_cast<Constant>(V)) {
1789 if (C->isNullValue())
1791 if (isa<ConstantInt>(C))
1792 // Must be non-zero due to null test above.
1794 // TODO: Handle vectors
1798 if (Instruction* I = dyn_cast<Instruction>(V)) {
1799 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1800 // If the possible ranges don't contain zero, then the value is
1801 // definitely non-zero.
1802 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1803 const APInt ZeroValue(Ty->getBitWidth(), 0);
1804 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1810 // The remaining tests are all recursive, so bail out if we hit the limit.
1811 if (Depth++ >= MaxDepth)
1814 // Check for pointer simplifications.
1815 if (V->getType()->isPointerTy()) {
1816 if (isKnownNonNull(V))
1818 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1819 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1823 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1825 // X | Y != 0 if X != 0 or Y != 0.
1826 Value *X = nullptr, *Y = nullptr;
1827 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1828 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1830 // ext X != 0 if X != 0.
1831 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1832 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1834 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1835 // if the lowest bit is shifted off the end.
1836 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1837 // shl nuw can't remove any non-zero bits.
1838 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1839 if (BO->hasNoUnsignedWrap())
1840 return isKnownNonZero(X, DL, Depth, Q);
1842 APInt KnownZero(BitWidth, 0);
1843 APInt KnownOne(BitWidth, 0);
1844 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1848 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1849 // defined if the sign bit is shifted off the end.
1850 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1851 // shr exact can only shift out zero bits.
1852 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1854 return isKnownNonZero(X, DL, Depth, Q);
1856 bool XKnownNonNegative, XKnownNegative;
1857 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1861 // If the shifter operand is a constant, and all of the bits shifted
1862 // out are known to be zero, and X is known non-zero then at least one
1863 // non-zero bit must remain.
1864 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1865 APInt KnownZero(BitWidth, 0);
1866 APInt KnownOne(BitWidth, 0);
1867 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1869 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1870 // Is there a known one in the portion not shifted out?
1871 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1873 // Are all the bits to be shifted out known zero?
1874 if (KnownZero.countTrailingOnes() >= ShiftVal)
1875 return isKnownNonZero(X, DL, Depth, Q);
1878 // div exact can only produce a zero if the dividend is zero.
1879 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1880 return isKnownNonZero(X, DL, Depth, Q);
1883 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1884 bool XKnownNonNegative, XKnownNegative;
1885 bool YKnownNonNegative, YKnownNegative;
1886 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1887 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1889 // If X and Y are both non-negative (as signed values) then their sum is not
1890 // zero unless both X and Y are zero.
1891 if (XKnownNonNegative && YKnownNonNegative)
1892 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1895 // If X and Y are both negative (as signed values) then their sum is not
1896 // zero unless both X and Y equal INT_MIN.
1897 if (BitWidth && XKnownNegative && YKnownNegative) {
1898 APInt KnownZero(BitWidth, 0);
1899 APInt KnownOne(BitWidth, 0);
1900 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1901 // The sign bit of X is set. If some other bit is set then X is not equal
1903 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1904 if ((KnownOne & Mask) != 0)
1906 // The sign bit of Y is set. If some other bit is set then Y is not equal
1908 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1909 if ((KnownOne & Mask) != 0)
1913 // The sum of a non-negative number and a power of two is not zero.
1914 if (XKnownNonNegative &&
1915 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1917 if (YKnownNonNegative &&
1918 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1922 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1923 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1924 // If X and Y are non-zero then so is X * Y as long as the multiplication
1925 // does not overflow.
1926 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1927 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
1930 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1931 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1932 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
1933 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
1937 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1938 // Try and detect a recurrence that monotonically increases from a
1939 // starting value, as these are common as induction variables.
1940 if (PN->getNumIncomingValues() == 2) {
1941 Value *Start = PN->getIncomingValue(0);
1942 Value *Induction = PN->getIncomingValue(1);
1943 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
1944 std::swap(Start, Induction);
1945 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
1946 if (!C->isZero() && !C->isNegative()) {
1948 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
1949 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
1957 if (!BitWidth) return false;
1958 APInt KnownZero(BitWidth, 0);
1959 APInt KnownOne(BitWidth, 0);
1960 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1961 return KnownOne != 0;
1964 /// Return true if V2 == V1 + X, where X is known non-zero.
1965 static bool isAddOfNonZero(Value *V1, Value *V2, const DataLayout &DL,
1967 BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
1968 if (!BO || BO->getOpcode() != Instruction::Add)
1970 Value *Op = nullptr;
1971 if (V2 == BO->getOperand(0))
1972 Op = BO->getOperand(1);
1973 else if (V2 == BO->getOperand(1))
1974 Op = BO->getOperand(0);
1977 return isKnownNonZero(Op, DL, 0, Q);
1980 /// Return true if it is known that V1 != V2.
1981 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
1983 if (V1->getType()->isVectorTy() || V1 == V2)
1985 if (V1->getType() != V2->getType())
1986 // We can't look through casts yet.
1988 if (isAddOfNonZero(V1, V2, DL, Q) || isAddOfNonZero(V2, V1, DL, Q))
1991 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
1992 // Are any known bits in V1 contradictory to known bits in V2? If V1
1993 // has a known zero where V2 has a known one, they must not be equal.
1994 auto BitWidth = Ty->getBitWidth();
1995 APInt KnownZero1(BitWidth, 0);
1996 APInt KnownOne1(BitWidth, 0);
1997 computeKnownBits(V1, KnownZero1, KnownOne1, DL, 0, Q);
1998 APInt KnownZero2(BitWidth, 0);
1999 APInt KnownOne2(BitWidth, 0);
2000 computeKnownBits(V2, KnownZero2, KnownOne2, DL, 0, Q);
2002 auto OppositeBits = (KnownZero1 & KnownOne2) | (KnownZero2 & KnownOne1);
2003 if (OppositeBits.getBoolValue())
2009 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2010 /// simplify operations downstream. Mask is known to be zero for bits that V
2013 /// This function is defined on values with integer type, values with pointer
2014 /// type, and vectors of integers. In the case
2015 /// where V is a vector, the mask, known zero, and known one values are the
2016 /// same width as the vector element, and the bit is set only if it is true
2017 /// for all of the elements in the vector.
2018 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
2019 unsigned Depth, const Query &Q) {
2020 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
2021 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2022 return (KnownZero & Mask) == Mask;
2027 /// Return the number of times the sign bit of the register is replicated into
2028 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2029 /// (itself), but other cases can give us information. For example, immediately
2030 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2031 /// other, so we return 3.
2033 /// 'Op' must have a scalar integer type.
2035 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
2037 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
2039 unsigned FirstAnswer = 1;
2041 // Note that ConstantInt is handled by the general computeKnownBits case
2045 return 1; // Limit search depth.
2047 Operator *U = dyn_cast<Operator>(V);
2048 switch (Operator::getOpcode(V)) {
2050 case Instruction::SExt:
2051 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2052 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
2054 case Instruction::SDiv: {
2055 const APInt *Denominator;
2056 // sdiv X, C -> adds log(C) sign bits.
2057 if (match(U->getOperand(1), m_APInt(Denominator))) {
2059 // Ignore non-positive denominator.
2060 if (!Denominator->isStrictlyPositive())
2063 // Calculate the incoming numerator bits.
2064 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2066 // Add floor(log(C)) bits to the numerator bits.
2067 return std::min(TyBits, NumBits + Denominator->logBase2());
2072 case Instruction::SRem: {
2073 const APInt *Denominator;
2074 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2075 // positive constant. This let us put a lower bound on the number of sign
2077 if (match(U->getOperand(1), m_APInt(Denominator))) {
2079 // Ignore non-positive denominator.
2080 if (!Denominator->isStrictlyPositive())
2083 // Calculate the incoming numerator bits. SRem by a positive constant
2084 // can't lower the number of sign bits.
2086 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2088 // Calculate the leading sign bit constraints by examining the
2089 // denominator. Given that the denominator is positive, there are two
2092 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2093 // (1 << ceilLogBase2(C)).
2095 // 2. the numerator is negative. Then the result range is (-C,0] and
2096 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2098 // Thus a lower bound on the number of sign bits is `TyBits -
2099 // ceilLogBase2(C)`.
2101 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2102 return std::max(NumrBits, ResBits);
2107 case Instruction::AShr: {
2108 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2109 // ashr X, C -> adds C sign bits. Vectors too.
2111 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2112 Tmp += ShAmt->getZExtValue();
2113 if (Tmp > TyBits) Tmp = TyBits;
2117 case Instruction::Shl: {
2119 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2120 // shl destroys sign bits.
2121 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2122 Tmp2 = ShAmt->getZExtValue();
2123 if (Tmp2 >= TyBits || // Bad shift.
2124 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2129 case Instruction::And:
2130 case Instruction::Or:
2131 case Instruction::Xor: // NOT is handled here.
2132 // Logical binary ops preserve the number of sign bits at the worst.
2133 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2135 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2136 FirstAnswer = std::min(Tmp, Tmp2);
2137 // We computed what we know about the sign bits as our first
2138 // answer. Now proceed to the generic code that uses
2139 // computeKnownBits, and pick whichever answer is better.
2143 case Instruction::Select:
2144 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2145 if (Tmp == 1) return 1; // Early out.
2146 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2147 return std::min(Tmp, Tmp2);
2149 case Instruction::Add:
2150 // Add can have at most one carry bit. Thus we know that the output
2151 // is, at worst, one more bit than the inputs.
2152 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2153 if (Tmp == 1) return 1; // Early out.
2155 // Special case decrementing a value (ADD X, -1):
2156 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2157 if (CRHS->isAllOnesValue()) {
2158 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2159 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2162 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2164 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2167 // If we are subtracting one from a positive number, there is no carry
2168 // out of the result.
2169 if (KnownZero.isNegative())
2173 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2174 if (Tmp2 == 1) return 1;
2175 return std::min(Tmp, Tmp2)-1;
2177 case Instruction::Sub:
2178 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2179 if (Tmp2 == 1) return 1;
2182 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2183 if (CLHS->isNullValue()) {
2184 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2185 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2187 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2189 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2192 // If the input is known to be positive (the sign bit is known clear),
2193 // the output of the NEG has the same number of sign bits as the input.
2194 if (KnownZero.isNegative())
2197 // Otherwise, we treat this like a SUB.
2200 // Sub can have at most one carry bit. Thus we know that the output
2201 // is, at worst, one more bit than the inputs.
2202 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2203 if (Tmp == 1) return 1; // Early out.
2204 return std::min(Tmp, Tmp2)-1;
2206 case Instruction::PHI: {
2207 PHINode *PN = cast<PHINode>(U);
2208 unsigned NumIncomingValues = PN->getNumIncomingValues();
2209 // Don't analyze large in-degree PHIs.
2210 if (NumIncomingValues > 4) break;
2211 // Unreachable blocks may have zero-operand PHI nodes.
2212 if (NumIncomingValues == 0) break;
2214 // Take the minimum of all incoming values. This can't infinitely loop
2215 // because of our depth threshold.
2216 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2217 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2218 if (Tmp == 1) return Tmp;
2220 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2225 case Instruction::Trunc:
2226 // FIXME: it's tricky to do anything useful for this, but it is an important
2227 // case for targets like X86.
2231 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2232 // use this information.
2233 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2235 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2237 if (KnownZero.isNegative()) { // sign bit is 0
2239 } else if (KnownOne.isNegative()) { // sign bit is 1;
2246 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2247 // the number of identical bits in the top of the input value.
2249 Mask <<= Mask.getBitWidth()-TyBits;
2250 // Return # leading zeros. We use 'min' here in case Val was zero before
2251 // shifting. We don't want to return '64' as for an i32 "0".
2252 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2255 /// This function computes the integer multiple of Base that equals V.
2256 /// If successful, it returns true and returns the multiple in
2257 /// Multiple. If unsuccessful, it returns false. It looks
2258 /// through SExt instructions only if LookThroughSExt is true.
2259 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2260 bool LookThroughSExt, unsigned Depth) {
2261 const unsigned MaxDepth = 6;
2263 assert(V && "No Value?");
2264 assert(Depth <= MaxDepth && "Limit Search Depth");
2265 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2267 Type *T = V->getType();
2269 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2279 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2280 Constant *BaseVal = ConstantInt::get(T, Base);
2281 if (CO && CO == BaseVal) {
2283 Multiple = ConstantInt::get(T, 1);
2287 if (CI && CI->getZExtValue() % Base == 0) {
2288 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2292 if (Depth == MaxDepth) return false; // Limit search depth.
2294 Operator *I = dyn_cast<Operator>(V);
2295 if (!I) return false;
2297 switch (I->getOpcode()) {
2299 case Instruction::SExt:
2300 if (!LookThroughSExt) return false;
2301 // otherwise fall through to ZExt
2302 case Instruction::ZExt:
2303 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2304 LookThroughSExt, Depth+1);
2305 case Instruction::Shl:
2306 case Instruction::Mul: {
2307 Value *Op0 = I->getOperand(0);
2308 Value *Op1 = I->getOperand(1);
2310 if (I->getOpcode() == Instruction::Shl) {
2311 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2312 if (!Op1CI) return false;
2313 // Turn Op0 << Op1 into Op0 * 2^Op1
2314 APInt Op1Int = Op1CI->getValue();
2315 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2316 APInt API(Op1Int.getBitWidth(), 0);
2317 API.setBit(BitToSet);
2318 Op1 = ConstantInt::get(V->getContext(), API);
2321 Value *Mul0 = nullptr;
2322 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2323 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2324 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2325 if (Op1C->getType()->getPrimitiveSizeInBits() <
2326 MulC->getType()->getPrimitiveSizeInBits())
2327 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2328 if (Op1C->getType()->getPrimitiveSizeInBits() >
2329 MulC->getType()->getPrimitiveSizeInBits())
2330 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2332 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2333 Multiple = ConstantExpr::getMul(MulC, Op1C);
2337 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2338 if (Mul0CI->getValue() == 1) {
2339 // V == Base * Op1, so return Op1
2345 Value *Mul1 = nullptr;
2346 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2347 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2348 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2349 if (Op0C->getType()->getPrimitiveSizeInBits() <
2350 MulC->getType()->getPrimitiveSizeInBits())
2351 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2352 if (Op0C->getType()->getPrimitiveSizeInBits() >
2353 MulC->getType()->getPrimitiveSizeInBits())
2354 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2356 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2357 Multiple = ConstantExpr::getMul(MulC, Op0C);
2361 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2362 if (Mul1CI->getValue() == 1) {
2363 // V == Base * Op0, so return Op0
2371 // We could not determine if V is a multiple of Base.
2375 /// Return true if we can prove that the specified FP value is never equal to
2378 /// NOTE: this function will need to be revisited when we support non-default
2381 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2382 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2383 return !CFP->getValueAPF().isNegZero();
2385 // FIXME: Magic number! At the least, this should be given a name because it's
2386 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2387 // expose it as a parameter, so it can be used for testing / experimenting.
2389 return false; // Limit search depth.
2391 const Operator *I = dyn_cast<Operator>(V);
2392 if (!I) return false;
2394 // Check if the nsz fast-math flag is set
2395 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2396 if (FPO->hasNoSignedZeros())
2399 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2400 if (I->getOpcode() == Instruction::FAdd)
2401 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2402 if (CFP->isNullValue())
2405 // sitofp and uitofp turn into +0.0 for zero.
2406 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2409 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2410 // sqrt(-0.0) = -0.0, no other negative results are possible.
2411 if (II->getIntrinsicID() == Intrinsic::sqrt)
2412 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2414 if (const CallInst *CI = dyn_cast<CallInst>(I))
2415 if (const Function *F = CI->getCalledFunction()) {
2416 if (F->isDeclaration()) {
2418 if (F->getName() == "abs") return true;
2419 // fabs[lf](x) != -0.0
2420 if (F->getName() == "fabs") return true;
2421 if (F->getName() == "fabsf") return true;
2422 if (F->getName() == "fabsl") return true;
2423 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2424 F->getName() == "sqrtl")
2425 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2432 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2433 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2434 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2436 // FIXME: Magic number! At the least, this should be given a name because it's
2437 // used similarly in CannotBeNegativeZero(). A better fix may be to
2438 // expose it as a parameter, so it can be used for testing / experimenting.
2440 return false; // Limit search depth.
2442 const Operator *I = dyn_cast<Operator>(V);
2443 if (!I) return false;
2445 switch (I->getOpcode()) {
2447 case Instruction::FMul:
2448 // x*x is always non-negative or a NaN.
2449 if (I->getOperand(0) == I->getOperand(1))
2452 case Instruction::FAdd:
2453 case Instruction::FDiv:
2454 case Instruction::FRem:
2455 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2456 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2457 case Instruction::FPExt:
2458 case Instruction::FPTrunc:
2459 // Widening/narrowing never change sign.
2460 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2461 case Instruction::Call:
2462 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2463 switch (II->getIntrinsicID()) {
2465 case Intrinsic::exp:
2466 case Intrinsic::exp2:
2467 case Intrinsic::fabs:
2468 case Intrinsic::sqrt:
2470 case Intrinsic::powi:
2471 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2472 // powi(x,n) is non-negative if n is even.
2473 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2476 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2477 case Intrinsic::fma:
2478 case Intrinsic::fmuladd:
2479 // x*x+y is non-negative if y is non-negative.
2480 return I->getOperand(0) == I->getOperand(1) &&
2481 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2488 /// If the specified value can be set by repeating the same byte in memory,
2489 /// return the i8 value that it is represented with. This is
2490 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2491 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2492 /// byte store (e.g. i16 0x1234), return null.
2493 Value *llvm::isBytewiseValue(Value *V) {
2494 // All byte-wide stores are splatable, even of arbitrary variables.
2495 if (V->getType()->isIntegerTy(8)) return V;
2497 // Handle 'null' ConstantArrayZero etc.
2498 if (Constant *C = dyn_cast<Constant>(V))
2499 if (C->isNullValue())
2500 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2502 // Constant float and double values can be handled as integer values if the
2503 // corresponding integer value is "byteable". An important case is 0.0.
2504 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2505 if (CFP->getType()->isFloatTy())
2506 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2507 if (CFP->getType()->isDoubleTy())
2508 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2509 // Don't handle long double formats, which have strange constraints.
2512 // We can handle constant integers that are multiple of 8 bits.
2513 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2514 if (CI->getBitWidth() % 8 == 0) {
2515 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2517 if (!CI->getValue().isSplat(8))
2519 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2523 // A ConstantDataArray/Vector is splatable if all its members are equal and
2525 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2526 Value *Elt = CA->getElementAsConstant(0);
2527 Value *Val = isBytewiseValue(Elt);
2531 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2532 if (CA->getElementAsConstant(I) != Elt)
2538 // Conceptually, we could handle things like:
2539 // %a = zext i8 %X to i16
2540 // %b = shl i16 %a, 8
2541 // %c = or i16 %a, %b
2542 // but until there is an example that actually needs this, it doesn't seem
2543 // worth worrying about.
2548 // This is the recursive version of BuildSubAggregate. It takes a few different
2549 // arguments. Idxs is the index within the nested struct From that we are
2550 // looking at now (which is of type IndexedType). IdxSkip is the number of
2551 // indices from Idxs that should be left out when inserting into the resulting
2552 // struct. To is the result struct built so far, new insertvalue instructions
2554 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2555 SmallVectorImpl<unsigned> &Idxs,
2557 Instruction *InsertBefore) {
2558 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2560 // Save the original To argument so we can modify it
2562 // General case, the type indexed by Idxs is a struct
2563 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2564 // Process each struct element recursively
2567 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2571 // Couldn't find any inserted value for this index? Cleanup
2572 while (PrevTo != OrigTo) {
2573 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2574 PrevTo = Del->getAggregateOperand();
2575 Del->eraseFromParent();
2577 // Stop processing elements
2581 // If we successfully found a value for each of our subaggregates
2585 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2586 // the struct's elements had a value that was inserted directly. In the latter
2587 // case, perhaps we can't determine each of the subelements individually, but
2588 // we might be able to find the complete struct somewhere.
2590 // Find the value that is at that particular spot
2591 Value *V = FindInsertedValue(From, Idxs);
2596 // Insert the value in the new (sub) aggregrate
2597 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2598 "tmp", InsertBefore);
2601 // This helper takes a nested struct and extracts a part of it (which is again a
2602 // struct) into a new value. For example, given the struct:
2603 // { a, { b, { c, d }, e } }
2604 // and the indices "1, 1" this returns
2607 // It does this by inserting an insertvalue for each element in the resulting
2608 // struct, as opposed to just inserting a single struct. This will only work if
2609 // each of the elements of the substruct are known (ie, inserted into From by an
2610 // insertvalue instruction somewhere).
2612 // All inserted insertvalue instructions are inserted before InsertBefore
2613 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2614 Instruction *InsertBefore) {
2615 assert(InsertBefore && "Must have someplace to insert!");
2616 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2618 Value *To = UndefValue::get(IndexedType);
2619 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2620 unsigned IdxSkip = Idxs.size();
2622 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2625 /// Given an aggregrate and an sequence of indices, see if
2626 /// the scalar value indexed is already around as a register, for example if it
2627 /// were inserted directly into the aggregrate.
2629 /// If InsertBefore is not null, this function will duplicate (modified)
2630 /// insertvalues when a part of a nested struct is extracted.
2631 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2632 Instruction *InsertBefore) {
2633 // Nothing to index? Just return V then (this is useful at the end of our
2635 if (idx_range.empty())
2637 // We have indices, so V should have an indexable type.
2638 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2639 "Not looking at a struct or array?");
2640 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2641 "Invalid indices for type?");
2643 if (Constant *C = dyn_cast<Constant>(V)) {
2644 C = C->getAggregateElement(idx_range[0]);
2645 if (!C) return nullptr;
2646 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2649 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2650 // Loop the indices for the insertvalue instruction in parallel with the
2651 // requested indices
2652 const unsigned *req_idx = idx_range.begin();
2653 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2654 i != e; ++i, ++req_idx) {
2655 if (req_idx == idx_range.end()) {
2656 // We can't handle this without inserting insertvalues
2660 // The requested index identifies a part of a nested aggregate. Handle
2661 // this specially. For example,
2662 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2663 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2664 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2665 // This can be changed into
2666 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2667 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2668 // which allows the unused 0,0 element from the nested struct to be
2670 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2674 // This insert value inserts something else than what we are looking for.
2675 // See if the (aggregate) value inserted into has the value we are
2676 // looking for, then.
2678 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2681 // If we end up here, the indices of the insertvalue match with those
2682 // requested (though possibly only partially). Now we recursively look at
2683 // the inserted value, passing any remaining indices.
2684 return FindInsertedValue(I->getInsertedValueOperand(),
2685 makeArrayRef(req_idx, idx_range.end()),
2689 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2690 // If we're extracting a value from an aggregate that was extracted from
2691 // something else, we can extract from that something else directly instead.
2692 // However, we will need to chain I's indices with the requested indices.
2694 // Calculate the number of indices required
2695 unsigned size = I->getNumIndices() + idx_range.size();
2696 // Allocate some space to put the new indices in
2697 SmallVector<unsigned, 5> Idxs;
2699 // Add indices from the extract value instruction
2700 Idxs.append(I->idx_begin(), I->idx_end());
2702 // Add requested indices
2703 Idxs.append(idx_range.begin(), idx_range.end());
2705 assert(Idxs.size() == size
2706 && "Number of indices added not correct?");
2708 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2710 // Otherwise, we don't know (such as, extracting from a function return value
2711 // or load instruction)
2715 /// Analyze the specified pointer to see if it can be expressed as a base
2716 /// pointer plus a constant offset. Return the base and offset to the caller.
2717 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2718 const DataLayout &DL) {
2719 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2720 APInt ByteOffset(BitWidth, 0);
2722 if (Ptr->getType()->isVectorTy())
2725 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2726 APInt GEPOffset(BitWidth, 0);
2727 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2730 ByteOffset += GEPOffset;
2732 Ptr = GEP->getPointerOperand();
2733 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2734 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2735 Ptr = cast<Operator>(Ptr)->getOperand(0);
2736 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2737 if (GA->mayBeOverridden())
2739 Ptr = GA->getAliasee();
2744 Offset = ByteOffset.getSExtValue();
2749 /// This function computes the length of a null-terminated C string pointed to
2750 /// by V. If successful, it returns true and returns the string in Str.
2751 /// If unsuccessful, it returns false.
2752 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2753 uint64_t Offset, bool TrimAtNul) {
2756 // Look through bitcast instructions and geps.
2757 V = V->stripPointerCasts();
2759 // If the value is a GEP instruction or constant expression, treat it as an
2761 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2762 // Make sure the GEP has exactly three arguments.
2763 if (GEP->getNumOperands() != 3)
2766 // Make sure the index-ee is a pointer to array of i8.
2767 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2768 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2769 if (!AT || !AT->getElementType()->isIntegerTy(8))
2772 // Check to make sure that the first operand of the GEP is an integer and
2773 // has value 0 so that we are sure we're indexing into the initializer.
2774 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2775 if (!FirstIdx || !FirstIdx->isZero())
2778 // If the second index isn't a ConstantInt, then this is a variable index
2779 // into the array. If this occurs, we can't say anything meaningful about
2781 uint64_t StartIdx = 0;
2782 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2783 StartIdx = CI->getZExtValue();
2786 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2790 // The GEP instruction, constant or instruction, must reference a global
2791 // variable that is a constant and is initialized. The referenced constant
2792 // initializer is the array that we'll use for optimization.
2793 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2794 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2797 // Handle the all-zeros case
2798 if (GV->getInitializer()->isNullValue()) {
2799 // This is a degenerate case. The initializer is constant zero so the
2800 // length of the string must be zero.
2805 // Must be a Constant Array
2806 const ConstantDataArray *Array =
2807 dyn_cast<ConstantDataArray>(GV->getInitializer());
2808 if (!Array || !Array->isString())
2811 // Get the number of elements in the array
2812 uint64_t NumElts = Array->getType()->getArrayNumElements();
2814 // Start out with the entire array in the StringRef.
2815 Str = Array->getAsString();
2817 if (Offset > NumElts)
2820 // Skip over 'offset' bytes.
2821 Str = Str.substr(Offset);
2824 // Trim off the \0 and anything after it. If the array is not nul
2825 // terminated, we just return the whole end of string. The client may know
2826 // some other way that the string is length-bound.
2827 Str = Str.substr(0, Str.find('\0'));
2832 // These next two are very similar to the above, but also look through PHI
2834 // TODO: See if we can integrate these two together.
2836 /// If we can compute the length of the string pointed to by
2837 /// the specified pointer, return 'len+1'. If we can't, return 0.
2838 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2839 // Look through noop bitcast instructions.
2840 V = V->stripPointerCasts();
2842 // If this is a PHI node, there are two cases: either we have already seen it
2844 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2845 if (!PHIs.insert(PN).second)
2846 return ~0ULL; // already in the set.
2848 // If it was new, see if all the input strings are the same length.
2849 uint64_t LenSoFar = ~0ULL;
2850 for (Value *IncValue : PN->incoming_values()) {
2851 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2852 if (Len == 0) return 0; // Unknown length -> unknown.
2854 if (Len == ~0ULL) continue;
2856 if (Len != LenSoFar && LenSoFar != ~0ULL)
2857 return 0; // Disagree -> unknown.
2861 // Success, all agree.
2865 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2866 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2867 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2868 if (Len1 == 0) return 0;
2869 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2870 if (Len2 == 0) return 0;
2871 if (Len1 == ~0ULL) return Len2;
2872 if (Len2 == ~0ULL) return Len1;
2873 if (Len1 != Len2) return 0;
2877 // Otherwise, see if we can read the string.
2879 if (!getConstantStringInfo(V, StrData))
2882 return StrData.size()+1;
2885 /// If we can compute the length of the string pointed to by
2886 /// the specified pointer, return 'len+1'. If we can't, return 0.
2887 uint64_t llvm::GetStringLength(Value *V) {
2888 if (!V->getType()->isPointerTy()) return 0;
2890 SmallPtrSet<PHINode*, 32> PHIs;
2891 uint64_t Len = GetStringLengthH(V, PHIs);
2892 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2893 // an empty string as a length.
2894 return Len == ~0ULL ? 1 : Len;
2897 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2898 /// previous iteration of the loop was referring to the same object as \p PN.
2899 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2900 // Find the loop-defined value.
2901 Loop *L = LI->getLoopFor(PN->getParent());
2902 if (PN->getNumIncomingValues() != 2)
2905 // Find the value from previous iteration.
2906 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2907 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2908 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2909 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2912 // If a new pointer is loaded in the loop, the pointer references a different
2913 // object in every iteration. E.g.:
2917 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2918 if (!L->isLoopInvariant(Load->getPointerOperand()))
2923 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2924 unsigned MaxLookup) {
2925 if (!V->getType()->isPointerTy())
2927 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2928 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2929 V = GEP->getPointerOperand();
2930 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2931 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2932 V = cast<Operator>(V)->getOperand(0);
2933 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2934 if (GA->mayBeOverridden())
2936 V = GA->getAliasee();
2938 // See if InstructionSimplify knows any relevant tricks.
2939 if (Instruction *I = dyn_cast<Instruction>(V))
2940 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2941 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
2948 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2953 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
2954 const DataLayout &DL, LoopInfo *LI,
2955 unsigned MaxLookup) {
2956 SmallPtrSet<Value *, 4> Visited;
2957 SmallVector<Value *, 4> Worklist;
2958 Worklist.push_back(V);
2960 Value *P = Worklist.pop_back_val();
2961 P = GetUnderlyingObject(P, DL, MaxLookup);
2963 if (!Visited.insert(P).second)
2966 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2967 Worklist.push_back(SI->getTrueValue());
2968 Worklist.push_back(SI->getFalseValue());
2972 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2973 // If this PHI changes the underlying object in every iteration of the
2974 // loop, don't look through it. Consider:
2977 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
2981 // Prev is tracking Curr one iteration behind so they refer to different
2982 // underlying objects.
2983 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
2984 isSameUnderlyingObjectInLoop(PN, LI))
2985 for (Value *IncValue : PN->incoming_values())
2986 Worklist.push_back(IncValue);
2990 Objects.push_back(P);
2991 } while (!Worklist.empty());
2994 /// Return true if the only users of this pointer are lifetime markers.
2995 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2996 for (const User *U : V->users()) {
2997 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2998 if (!II) return false;
3000 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3001 II->getIntrinsicID() != Intrinsic::lifetime_end)
3007 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
3008 Type *Ty, const DataLayout &DL,
3009 const Instruction *CtxI,
3010 const DominatorTree *DT,
3011 const TargetLibraryInfo *TLI) {
3012 assert(Offset.isNonNegative() && "offset can't be negative");
3013 assert(Ty->isSized() && "must be sized");
3015 APInt DerefBytes(Offset.getBitWidth(), 0);
3016 bool CheckForNonNull = false;
3017 if (const Argument *A = dyn_cast<Argument>(BV)) {
3018 DerefBytes = A->getDereferenceableBytes();
3019 if (!DerefBytes.getBoolValue()) {
3020 DerefBytes = A->getDereferenceableOrNullBytes();
3021 CheckForNonNull = true;
3023 } else if (auto CS = ImmutableCallSite(BV)) {
3024 DerefBytes = CS.getDereferenceableBytes(0);
3025 if (!DerefBytes.getBoolValue()) {
3026 DerefBytes = CS.getDereferenceableOrNullBytes(0);
3027 CheckForNonNull = true;
3029 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
3030 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
3031 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3032 DerefBytes = CI->getLimitedValue();
3034 if (!DerefBytes.getBoolValue()) {
3036 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
3037 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3038 DerefBytes = CI->getLimitedValue();
3040 CheckForNonNull = true;
3044 if (DerefBytes.getBoolValue())
3045 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
3046 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
3052 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
3053 const Instruction *CtxI,
3054 const DominatorTree *DT,
3055 const TargetLibraryInfo *TLI) {
3056 Type *VTy = V->getType();
3057 Type *Ty = VTy->getPointerElementType();
3061 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3062 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
3065 static bool isAligned(const Value *Base, APInt Offset, unsigned Align,
3066 const DataLayout &DL) {
3067 APInt BaseAlign(Offset.getBitWidth(), getAlignment(Base, DL));
3070 Type *Ty = Base->getType()->getPointerElementType();
3071 BaseAlign = DL.getABITypeAlignment(Ty);
3074 APInt Alignment(Offset.getBitWidth(), Align);
3076 assert(Alignment.isPowerOf2() && "must be a power of 2!");
3077 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1));
3080 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) {
3081 APInt Offset(DL.getTypeStoreSizeInBits(Base->getType()), 0);
3082 return isAligned(Base, Offset, Align, DL);
3085 /// Test if V is always a pointer to allocated and suitably aligned memory for
3086 /// a simple load or store.
3087 static bool isDereferenceableAndAlignedPointer(
3088 const Value *V, unsigned Align, const DataLayout &DL,
3089 const Instruction *CtxI, const DominatorTree *DT,
3090 const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited) {
3091 // Note that it is not safe to speculate into a malloc'd region because
3092 // malloc may return null.
3094 // These are obviously ok if aligned.
3095 if (isa<AllocaInst>(V))
3096 return isAligned(V, Align, DL);
3098 // It's not always safe to follow a bitcast, for example:
3099 // bitcast i8* (alloca i8) to i32*
3100 // would result in a 4-byte load from a 1-byte alloca. However,
3101 // if we're casting from a pointer from a type of larger size
3102 // to a type of smaller size (or the same size), and the alignment
3103 // is at least as large as for the resulting pointer type, then
3104 // we can look through the bitcast.
3105 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
3106 Type *STy = BC->getSrcTy()->getPointerElementType(),
3107 *DTy = BC->getDestTy()->getPointerElementType();
3108 if (STy->isSized() && DTy->isSized() &&
3109 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
3110 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
3111 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, DL,
3112 CtxI, DT, TLI, Visited);
3115 // Global variables which can't collapse to null are ok.
3116 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
3117 if (!GV->hasExternalWeakLinkage())
3118 return isAligned(V, Align, DL);
3120 // byval arguments are okay.
3121 if (const Argument *A = dyn_cast<Argument>(V))
3122 if (A->hasByValAttr())
3123 return isAligned(V, Align, DL);
3125 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
3126 return isAligned(V, Align, DL);
3128 // For GEPs, determine if the indexing lands within the allocated object.
3129 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3130 Type *VTy = GEP->getType();
3131 Type *Ty = VTy->getPointerElementType();
3132 const Value *Base = GEP->getPointerOperand();
3134 // Conservatively require that the base pointer be fully dereferenceable
3136 if (!Visited.insert(Base).second)
3138 if (!isDereferenceableAndAlignedPointer(Base, Align, DL, CtxI, DT, TLI,
3142 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
3143 if (!GEP->accumulateConstantOffset(DL, Offset))
3146 // Check if the load is within the bounds of the underlying object
3147 // and offset is aligned.
3148 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3149 Type *BaseType = Base->getType()->getPointerElementType();
3150 assert(isPowerOf2_32(Align) && "must be a power of 2!");
3151 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType)) &&
3152 !(Offset & APInt(Offset.getBitWidth(), Align-1));
3155 // For gc.relocate, look through relocations
3156 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3157 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3158 GCRelocateOperands RelocateInst(I);
3159 return isDereferenceableAndAlignedPointer(
3160 RelocateInst.getDerivedPtr(), Align, DL, CtxI, DT, TLI, Visited);
3163 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3164 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, DL,
3165 CtxI, DT, TLI, Visited);
3167 // If we don't know, assume the worst.
3171 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align,
3172 const DataLayout &DL,
3173 const Instruction *CtxI,
3174 const DominatorTree *DT,
3175 const TargetLibraryInfo *TLI) {
3176 // When dereferenceability information is provided by a dereferenceable
3177 // attribute, we know exactly how many bytes are dereferenceable. If we can
3178 // determine the exact offset to the attributed variable, we can use that
3179 // information here.
3180 Type *VTy = V->getType();
3181 Type *Ty = VTy->getPointerElementType();
3183 // Require ABI alignment for loads without alignment specification
3185 Align = DL.getABITypeAlignment(Ty);
3187 if (Ty->isSized()) {
3188 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3189 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3191 if (Offset.isNonNegative())
3192 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL, CtxI, DT, TLI) &&
3193 isAligned(BV, Offset, Align, DL))
3197 SmallPtrSet<const Value *, 32> Visited;
3198 return ::isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT, TLI,
3202 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3203 const Instruction *CtxI,
3204 const DominatorTree *DT,
3205 const TargetLibraryInfo *TLI) {
3206 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT, TLI);
3209 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3210 const Instruction *CtxI,
3211 const DominatorTree *DT,
3212 const TargetLibraryInfo *TLI) {
3213 const Operator *Inst = dyn_cast<Operator>(V);
3217 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3218 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3222 switch (Inst->getOpcode()) {
3225 case Instruction::UDiv:
3226 case Instruction::URem: {
3227 // x / y is undefined if y == 0.
3229 if (match(Inst->getOperand(1), m_APInt(V)))
3233 case Instruction::SDiv:
3234 case Instruction::SRem: {
3235 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3236 const APInt *Numerator, *Denominator;
3237 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3239 // We cannot hoist this division if the denominator is 0.
3240 if (*Denominator == 0)
3242 // It's safe to hoist if the denominator is not 0 or -1.
3243 if (*Denominator != -1)
3245 // At this point we know that the denominator is -1. It is safe to hoist as
3246 // long we know that the numerator is not INT_MIN.
3247 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3248 return !Numerator->isMinSignedValue();
3249 // The numerator *might* be MinSignedValue.
3252 case Instruction::Load: {
3253 const LoadInst *LI = cast<LoadInst>(Inst);
3254 if (!LI->isUnordered() ||
3255 // Speculative load may create a race that did not exist in the source.
3256 LI->getParent()->getParent()->hasFnAttribute(
3257 Attribute::SanitizeThread) ||
3258 // Speculative load may load data from dirty regions.
3259 LI->getParent()->getParent()->hasFnAttribute(
3260 Attribute::SanitizeAddress))
3262 const DataLayout &DL = LI->getModule()->getDataLayout();
3263 return isDereferenceableAndAlignedPointer(
3264 LI->getPointerOperand(), LI->getAlignment(), DL, CtxI, DT, TLI);
3266 case Instruction::Call: {
3267 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3268 switch (II->getIntrinsicID()) {
3269 // These synthetic intrinsics have no side-effects and just mark
3270 // information about their operands.
3271 // FIXME: There are other no-op synthetic instructions that potentially
3272 // should be considered at least *safe* to speculate...
3273 case Intrinsic::dbg_declare:
3274 case Intrinsic::dbg_value:
3277 case Intrinsic::bswap:
3278 case Intrinsic::ctlz:
3279 case Intrinsic::ctpop:
3280 case Intrinsic::cttz:
3281 case Intrinsic::objectsize:
3282 case Intrinsic::sadd_with_overflow:
3283 case Intrinsic::smul_with_overflow:
3284 case Intrinsic::ssub_with_overflow:
3285 case Intrinsic::uadd_with_overflow:
3286 case Intrinsic::umul_with_overflow:
3287 case Intrinsic::usub_with_overflow:
3289 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3290 // errno like libm sqrt would.
3291 case Intrinsic::sqrt:
3292 case Intrinsic::fma:
3293 case Intrinsic::fmuladd:
3294 case Intrinsic::fabs:
3295 case Intrinsic::minnum:
3296 case Intrinsic::maxnum:
3298 // TODO: some fp intrinsics are marked as having the same error handling
3299 // as libm. They're safe to speculate when they won't error.
3300 // TODO: are convert_{from,to}_fp16 safe?
3301 // TODO: can we list target-specific intrinsics here?
3305 return false; // The called function could have undefined behavior or
3306 // side-effects, even if marked readnone nounwind.
3308 case Instruction::VAArg:
3309 case Instruction::Alloca:
3310 case Instruction::Invoke:
3311 case Instruction::PHI:
3312 case Instruction::Store:
3313 case Instruction::Ret:
3314 case Instruction::Br:
3315 case Instruction::IndirectBr:
3316 case Instruction::Switch:
3317 case Instruction::Unreachable:
3318 case Instruction::Fence:
3319 case Instruction::AtomicRMW:
3320 case Instruction::AtomicCmpXchg:
3321 case Instruction::LandingPad:
3322 case Instruction::Resume:
3323 case Instruction::CatchPad:
3324 case Instruction::CatchEndPad:
3325 case Instruction::CatchRet:
3326 case Instruction::CleanupPad:
3327 case Instruction::CleanupEndPad:
3328 case Instruction::CleanupRet:
3329 case Instruction::TerminatePad:
3330 return false; // Misc instructions which have effects
3334 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3335 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3338 /// Return true if we know that the specified value is never null.
3339 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3340 assert(V->getType()->isPointerTy() && "V must be pointer type");
3342 // Alloca never returns null, malloc might.
3343 if (isa<AllocaInst>(V)) return true;
3345 // A byval, inalloca, or nonnull argument is never null.
3346 if (const Argument *A = dyn_cast<Argument>(V))
3347 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3349 // A global variable in address space 0 is non null unless extern weak.
3350 // Other address spaces may have null as a valid address for a global,
3351 // so we can't assume anything.
3352 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3353 return !GV->hasExternalWeakLinkage() &&
3354 GV->getType()->getAddressSpace() == 0;
3356 // A Load tagged w/nonnull metadata is never null.
3357 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3358 return LI->getMetadata(LLVMContext::MD_nonnull);
3360 if (auto CS = ImmutableCallSite(V))
3361 if (CS.isReturnNonNull())
3364 // operator new never returns null.
3365 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3371 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3372 const Instruction *CtxI,
3373 const DominatorTree *DT) {
3374 assert(V->getType()->isPointerTy() && "V must be pointer type");
3376 unsigned NumUsesExplored = 0;
3377 for (auto U : V->users()) {
3378 // Avoid massive lists
3379 if (NumUsesExplored >= DomConditionsMaxUses)
3382 // Consider only compare instructions uniquely controlling a branch
3383 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3387 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3390 for (auto *CmpU : Cmp->users()) {
3391 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3395 assert(BI->isConditional() && "uses a comparison!");
3397 BasicBlock *NonNullSuccessor = nullptr;
3398 CmpInst::Predicate Pred;
3400 if (match(const_cast<ICmpInst*>(Cmp),
3401 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3402 if (Pred == ICmpInst::ICMP_EQ)
3403 NonNullSuccessor = BI->getSuccessor(1);
3404 else if (Pred == ICmpInst::ICMP_NE)
3405 NonNullSuccessor = BI->getSuccessor(0);
3408 if (NonNullSuccessor) {
3409 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3410 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3419 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3420 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3421 if (isKnownNonNull(V, TLI))
3424 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3427 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3428 const DataLayout &DL,
3429 AssumptionCache *AC,
3430 const Instruction *CxtI,
3431 const DominatorTree *DT) {
3432 // Multiplying n * m significant bits yields a result of n + m significant
3433 // bits. If the total number of significant bits does not exceed the
3434 // result bit width (minus 1), there is no overflow.
3435 // This means if we have enough leading zero bits in the operands
3436 // we can guarantee that the result does not overflow.
3437 // Ref: "Hacker's Delight" by Henry Warren
3438 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3439 APInt LHSKnownZero(BitWidth, 0);
3440 APInt LHSKnownOne(BitWidth, 0);
3441 APInt RHSKnownZero(BitWidth, 0);
3442 APInt RHSKnownOne(BitWidth, 0);
3443 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3445 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3447 // Note that underestimating the number of zero bits gives a more
3448 // conservative answer.
3449 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3450 RHSKnownZero.countLeadingOnes();
3451 // First handle the easy case: if we have enough zero bits there's
3452 // definitely no overflow.
3453 if (ZeroBits >= BitWidth)
3454 return OverflowResult::NeverOverflows;
3456 // Get the largest possible values for each operand.
3457 APInt LHSMax = ~LHSKnownZero;
3458 APInt RHSMax = ~RHSKnownZero;
3460 // We know the multiply operation doesn't overflow if the maximum values for
3461 // each operand will not overflow after we multiply them together.
3463 LHSMax.umul_ov(RHSMax, MaxOverflow);
3465 return OverflowResult::NeverOverflows;
3467 // We know it always overflows if multiplying the smallest possible values for
3468 // the operands also results in overflow.
3470 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3472 return OverflowResult::AlwaysOverflows;
3474 return OverflowResult::MayOverflow;
3477 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3478 const DataLayout &DL,
3479 AssumptionCache *AC,
3480 const Instruction *CxtI,
3481 const DominatorTree *DT) {
3482 bool LHSKnownNonNegative, LHSKnownNegative;
3483 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3485 if (LHSKnownNonNegative || LHSKnownNegative) {
3486 bool RHSKnownNonNegative, RHSKnownNegative;
3487 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3490 if (LHSKnownNegative && RHSKnownNegative) {
3491 // The sign bit is set in both cases: this MUST overflow.
3492 // Create a simple add instruction, and insert it into the struct.
3493 return OverflowResult::AlwaysOverflows;
3496 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3497 // The sign bit is clear in both cases: this CANNOT overflow.
3498 // Create a simple add instruction, and insert it into the struct.
3499 return OverflowResult::NeverOverflows;
3503 return OverflowResult::MayOverflow;
3506 static OverflowResult computeOverflowForSignedAdd(
3507 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3508 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3509 if (Add && Add->hasNoSignedWrap()) {
3510 return OverflowResult::NeverOverflows;
3513 bool LHSKnownNonNegative, LHSKnownNegative;
3514 bool RHSKnownNonNegative, RHSKnownNegative;
3515 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3517 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3520 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3521 (LHSKnownNegative && RHSKnownNonNegative)) {
3522 // The sign bits are opposite: this CANNOT overflow.
3523 return OverflowResult::NeverOverflows;
3526 // The remaining code needs Add to be available. Early returns if not so.
3528 return OverflowResult::MayOverflow;
3530 // If the sign of Add is the same as at least one of the operands, this add
3531 // CANNOT overflow. This is particularly useful when the sum is
3532 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3534 bool LHSOrRHSKnownNonNegative =
3535 (LHSKnownNonNegative || RHSKnownNonNegative);
3536 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3537 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3538 bool AddKnownNonNegative, AddKnownNegative;
3539 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3540 /*Depth=*/0, AC, CxtI, DT);
3541 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3542 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3543 return OverflowResult::NeverOverflows;
3547 return OverflowResult::MayOverflow;
3550 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3551 const DataLayout &DL,
3552 AssumptionCache *AC,
3553 const Instruction *CxtI,
3554 const DominatorTree *DT) {
3555 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3556 Add, DL, AC, CxtI, DT);
3559 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3560 const DataLayout &DL,
3561 AssumptionCache *AC,
3562 const Instruction *CxtI,
3563 const DominatorTree *DT) {
3564 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3567 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3568 // FIXME: This conservative implementation can be relaxed. E.g. most
3569 // atomic operations are guaranteed to terminate on most platforms
3570 // and most functions terminate.
3572 return !I->isAtomic() && // atomics may never succeed on some platforms
3573 !isa<CallInst>(I) && // could throw and might not terminate
3574 !isa<InvokeInst>(I) && // might not terminate and could throw to
3575 // non-successor (see bug 24185 for details).
3576 !isa<ResumeInst>(I) && // has no successors
3577 !isa<ReturnInst>(I); // has no successors
3580 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3582 // The loop header is guaranteed to be executed for every iteration.
3584 // FIXME: Relax this constraint to cover all basic blocks that are
3585 // guaranteed to be executed at every iteration.
3586 if (I->getParent() != L->getHeader()) return false;
3588 for (const Instruction &LI : *L->getHeader()) {
3589 if (&LI == I) return true;
3590 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3592 llvm_unreachable("Instruction not contained in its own parent basic block.");
3595 bool llvm::propagatesFullPoison(const Instruction *I) {
3596 switch (I->getOpcode()) {
3597 case Instruction::Add:
3598 case Instruction::Sub:
3599 case Instruction::Xor:
3600 case Instruction::Trunc:
3601 case Instruction::BitCast:
3602 case Instruction::AddrSpaceCast:
3603 // These operations all propagate poison unconditionally. Note that poison
3604 // is not any particular value, so xor or subtraction of poison with
3605 // itself still yields poison, not zero.
3608 case Instruction::AShr:
3609 case Instruction::SExt:
3610 // For these operations, one bit of the input is replicated across
3611 // multiple output bits. A replicated poison bit is still poison.
3614 case Instruction::Shl: {
3615 // Left shift *by* a poison value is poison. The number of
3616 // positions to shift is unsigned, so no negative values are
3617 // possible there. Left shift by zero places preserves poison. So
3618 // it only remains to consider left shift of poison by a positive
3619 // number of places.
3621 // A left shift by a positive number of places leaves the lowest order bit
3622 // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3623 // make the poison operand violate that flag, yielding a fresh full-poison
3625 auto *OBO = cast<OverflowingBinaryOperator>(I);
3626 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3629 case Instruction::Mul: {
3630 // A multiplication by zero yields a non-poison zero result, so we need to
3631 // rule out zero as an operand. Conservatively, multiplication by a
3632 // non-zero constant is not multiplication by zero.
3634 // Multiplication by a non-zero constant can leave some bits
3635 // non-poisoned. For example, a multiplication by 2 leaves the lowest
3636 // order bit unpoisoned. So we need to consider that.
3638 // Multiplication by 1 preserves poison. If the multiplication has a
3639 // no-wrap flag, then we can make the poison operand violate that flag
3640 // when multiplied by any integer other than 0 and 1.
3641 auto *OBO = cast<OverflowingBinaryOperator>(I);
3642 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3643 for (Value *V : OBO->operands()) {
3644 if (auto *CI = dyn_cast<ConstantInt>(V)) {
3645 // A ConstantInt cannot yield poison, so we can assume that it is
3646 // the other operand that is poison.
3647 return !CI->isZero();
3654 case Instruction::GetElementPtr:
3655 // A GEP implicitly represents a sequence of additions, subtractions,
3656 // truncations, sign extensions and multiplications. The multiplications
3657 // are by the non-zero sizes of some set of types, so we do not have to be
3658 // concerned with multiplication by zero. If the GEP is in-bounds, then
3659 // these operations are implicitly no-signed-wrap so poison is propagated
3660 // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3661 return cast<GEPOperator>(I)->isInBounds();
3668 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3669 switch (I->getOpcode()) {
3670 case Instruction::Store:
3671 return cast<StoreInst>(I)->getPointerOperand();
3673 case Instruction::Load:
3674 return cast<LoadInst>(I)->getPointerOperand();
3676 case Instruction::AtomicCmpXchg:
3677 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3679 case Instruction::AtomicRMW:
3680 return cast<AtomicRMWInst>(I)->getPointerOperand();
3682 case Instruction::UDiv:
3683 case Instruction::SDiv:
3684 case Instruction::URem:
3685 case Instruction::SRem:
3686 return I->getOperand(1);
3693 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3694 // We currently only look for uses of poison values within the same basic
3695 // block, as that makes it easier to guarantee that the uses will be
3696 // executed given that PoisonI is executed.
3698 // FIXME: Expand this to consider uses beyond the same basic block. To do
3699 // this, look out for the distinction between post-dominance and strong
3701 const BasicBlock *BB = PoisonI->getParent();
3703 // Set of instructions that we have proved will yield poison if PoisonI
3705 SmallSet<const Value *, 16> YieldsPoison;
3706 YieldsPoison.insert(PoisonI);
3708 for (BasicBlock::const_iterator I = PoisonI->getIterator(), E = BB->end();
3710 if (&*I != PoisonI) {
3711 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&*I);
3712 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) return true;
3713 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
3717 // Mark poison that propagates from I through uses of I.
3718 if (YieldsPoison.count(&*I)) {
3719 for (const User *User : I->users()) {
3720 const Instruction *UserI = cast<Instruction>(User);
3721 if (UserI->getParent() == BB && propagatesFullPoison(UserI))
3722 YieldsPoison.insert(User);
3729 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3733 if (auto *C = dyn_cast<ConstantFP>(V))
3738 static bool isKnownNonZero(Value *V) {
3739 if (auto *C = dyn_cast<ConstantFP>(V))
3740 return !C->isZero();
3744 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3746 Value *CmpLHS, Value *CmpRHS,
3747 Value *TrueVal, Value *FalseVal,
3748 Value *&LHS, Value *&RHS) {
3752 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3753 // return inconsistent results between implementations.
3754 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3755 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3756 // Therefore we behave conservatively and only proceed if at least one of the
3757 // operands is known to not be zero, or if we don't care about signed zeroes.
3760 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3761 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3762 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3763 !isKnownNonZero(CmpRHS))
3764 return {SPF_UNKNOWN, SPNB_NA, false};
3767 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3768 bool Ordered = false;
3770 // When given one NaN and one non-NaN input:
3771 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3772 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3773 // ordered comparison fails), which could be NaN or non-NaN.
3774 // so here we discover exactly what NaN behavior is required/accepted.
3775 if (CmpInst::isFPPredicate(Pred)) {
3776 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3777 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3779 if (LHSSafe && RHSSafe) {
3780 // Both operands are known non-NaN.
3781 NaNBehavior = SPNB_RETURNS_ANY;
3782 } else if (CmpInst::isOrdered(Pred)) {
3783 // An ordered comparison will return false when given a NaN, so it
3787 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3788 NaNBehavior = SPNB_RETURNS_NAN;
3790 NaNBehavior = SPNB_RETURNS_OTHER;
3792 // Completely unsafe.
3793 return {SPF_UNKNOWN, SPNB_NA, false};
3796 // An unordered comparison will return true when given a NaN, so it
3799 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3800 NaNBehavior = SPNB_RETURNS_OTHER;
3802 NaNBehavior = SPNB_RETURNS_NAN;
3804 // Completely unsafe.
3805 return {SPF_UNKNOWN, SPNB_NA, false};
3809 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3810 std::swap(CmpLHS, CmpRHS);
3811 Pred = CmpInst::getSwappedPredicate(Pred);
3812 if (NaNBehavior == SPNB_RETURNS_NAN)
3813 NaNBehavior = SPNB_RETURNS_OTHER;
3814 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3815 NaNBehavior = SPNB_RETURNS_NAN;
3819 // ([if]cmp X, Y) ? X : Y
3820 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3822 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3823 case ICmpInst::ICMP_UGT:
3824 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3825 case ICmpInst::ICMP_SGT:
3826 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3827 case ICmpInst::ICMP_ULT:
3828 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3829 case ICmpInst::ICMP_SLT:
3830 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3831 case FCmpInst::FCMP_UGT:
3832 case FCmpInst::FCMP_UGE:
3833 case FCmpInst::FCMP_OGT:
3834 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3835 case FCmpInst::FCMP_ULT:
3836 case FCmpInst::FCMP_ULE:
3837 case FCmpInst::FCMP_OLT:
3838 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3842 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3843 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3844 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3846 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3847 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3848 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3849 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3852 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3853 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3854 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3855 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3859 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3860 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3861 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3862 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3863 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3866 return {SPF_SMIN, SPNB_NA, false};
3871 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3873 return {SPF_UNKNOWN, SPNB_NA, false};
3876 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
3877 Instruction::CastOps *CastOp) {
3878 CastInst *CI = dyn_cast<CastInst>(V1);
3879 Constant *C = dyn_cast<Constant>(V2);
3880 CastInst *CI2 = dyn_cast<CastInst>(V2);
3883 *CastOp = CI->getOpcode();
3886 // If V1 and V2 are both the same cast from the same type, we can look
3888 if (CI2->getOpcode() == CI->getOpcode() &&
3889 CI2->getSrcTy() == CI->getSrcTy())
3890 return CI2->getOperand(0);
3896 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3897 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3898 // This is only valid if the truncated value can be sign-extended
3899 // back to the original value.
3900 if (ConstantExpr::getSExt(T, C->getType()) == C)
3904 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3905 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3907 if (isa<TruncInst>(CI))
3908 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3910 if (isa<FPToUIInst>(CI))
3911 return ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
3913 if (isa<FPToSIInst>(CI))
3914 return ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
3916 if (isa<UIToFPInst>(CI))
3917 return ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
3919 if (isa<SIToFPInst>(CI))
3920 return ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
3922 if (isa<FPTruncInst>(CI))
3923 return ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
3925 if (isa<FPExtInst>(CI))
3926 return ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
3931 SelectPatternResult llvm::matchSelectPattern(Value *V,
3932 Value *&LHS, Value *&RHS,
3933 Instruction::CastOps *CastOp) {
3934 SelectInst *SI = dyn_cast<SelectInst>(V);
3935 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
3937 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
3938 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
3940 CmpInst::Predicate Pred = CmpI->getPredicate();
3941 Value *CmpLHS = CmpI->getOperand(0);
3942 Value *CmpRHS = CmpI->getOperand(1);
3943 Value *TrueVal = SI->getTrueValue();
3944 Value *FalseVal = SI->getFalseValue();
3946 if (isa<FPMathOperator>(CmpI))
3947 FMF = CmpI->getFastMathFlags();
3950 if (CmpI->isEquality())
3951 return {SPF_UNKNOWN, SPNB_NA, false};
3953 // Deal with type mismatches.
3954 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
3955 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
3956 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3957 cast<CastInst>(TrueVal)->getOperand(0), C,
3959 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
3960 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3961 C, cast<CastInst>(FalseVal)->getOperand(0),
3964 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,