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/Analysis/AssumptionTracker.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/MemoryBuiltins.h"
20 #include "llvm/IR/CallSite.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/MathExtras.h"
38 using namespace llvm::PatternMatch;
40 const unsigned MaxDepth = 6;
42 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
43 /// unknown returns 0). For vector types, returns the element type's bitwidth.
44 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
45 if (unsigned BitWidth = Ty->getScalarSizeInBits())
48 return TD ? TD->getPointerTypeSizeInBits(Ty) : 0;
51 // Many of these functions have internal versions that take an assumption
52 // exclusion set. This is because of the potential for mutual recursion to
53 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
54 // classic case of this is assume(x = y), which will attempt to determine
55 // bits in x from bits in y, which will attempt to determine bits in y from
56 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
57 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
58 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
59 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
61 // Simplifying using an assume can only be done in a particular control-flow
62 // context (the context instruction provides that context). If an assume and
63 // the context instruction are not in the same block then the DT helps in
64 // figuring out if we can use it.
67 AssumptionTracker *AT;
68 const Instruction *CxtI;
69 const DominatorTree *DT;
71 Query(AssumptionTracker *AT = nullptr, const Instruction *CxtI = nullptr,
72 const DominatorTree *DT = nullptr)
73 : AT(AT), CxtI(CxtI), DT(DT) {}
75 Query(const Query &Q, const Value *NewExcl)
76 : ExclInvs(Q.ExclInvs), AT(Q.AT), CxtI(Q.CxtI), DT(Q.DT) {
77 ExclInvs.insert(NewExcl);
81 // Given the provided Value and, potentially, a context instruction, returned
82 // the preferred context instruction (if any).
83 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
84 // If we've been provided with a context instruction, then use that (provided
85 // it has been inserted).
86 if (CxtI && CxtI->getParent())
89 // If the value is really an already-inserted instruction, then use that.
90 CxtI = dyn_cast<Instruction>(V);
91 if (CxtI && CxtI->getParent())
97 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
98 const DataLayout *TD, unsigned Depth,
101 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
102 const DataLayout *TD, unsigned Depth,
103 AssumptionTracker *AT, const Instruction *CxtI,
104 const DominatorTree *DT) {
105 ::computeKnownBits(V, KnownZero, KnownOne, TD, Depth,
106 Query(AT, safeCxtI(V, CxtI), DT));
109 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
110 const DataLayout *TD, unsigned Depth,
113 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
114 const DataLayout *TD, unsigned Depth,
115 AssumptionTracker *AT, const Instruction *CxtI,
116 const DominatorTree *DT) {
117 ::ComputeSignBit(V, KnownZero, KnownOne, TD, Depth,
118 Query(AT, safeCxtI(V, CxtI), DT));
121 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
124 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
125 AssumptionTracker *AT,
126 const Instruction *CxtI,
127 const DominatorTree *DT) {
128 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
129 Query(AT, safeCxtI(V, CxtI), DT));
132 static bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
135 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
136 AssumptionTracker *AT, const Instruction *CxtI,
137 const DominatorTree *DT) {
138 return ::isKnownNonZero(V, TD, Depth, Query(AT, safeCxtI(V, CxtI), DT));
141 static bool MaskedValueIsZero(Value *V, const APInt &Mask,
142 const DataLayout *TD, unsigned Depth,
145 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
146 const DataLayout *TD, unsigned Depth,
147 AssumptionTracker *AT, const Instruction *CxtI,
148 const DominatorTree *DT) {
149 return ::MaskedValueIsZero(V, Mask, TD, Depth,
150 Query(AT, safeCxtI(V, CxtI), DT));
153 static unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
154 unsigned Depth, const Query &Q);
156 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
157 unsigned Depth, AssumptionTracker *AT,
158 const Instruction *CxtI,
159 const DominatorTree *DT) {
160 return ::ComputeNumSignBits(V, TD, Depth, Query(AT, safeCxtI(V, CxtI), DT));
163 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
164 APInt &KnownZero, APInt &KnownOne,
165 APInt &KnownZero2, APInt &KnownOne2,
166 const DataLayout *TD, unsigned Depth,
169 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
170 // We know that the top bits of C-X are clear if X contains less bits
171 // than C (i.e. no wrap-around can happen). For example, 20-X is
172 // positive if we can prove that X is >= 0 and < 16.
173 if (!CLHS->getValue().isNegative()) {
174 unsigned BitWidth = KnownZero.getBitWidth();
175 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
176 // NLZ can't be BitWidth with no sign bit
177 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
178 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
180 // If all of the MaskV bits are known to be zero, then we know the
181 // output top bits are zero, because we now know that the output is
183 if ((KnownZero2 & MaskV) == MaskV) {
184 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
185 // Top bits known zero.
186 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
192 unsigned BitWidth = KnownZero.getBitWidth();
194 // If an initial sequence of bits in the result is not needed, the
195 // corresponding bits in the operands are not needed.
196 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
197 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1, Q);
198 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
200 // Carry in a 1 for a subtract, rather than a 0.
201 APInt CarryIn(BitWidth, 0);
203 // Sum = LHS + ~RHS + 1
204 std::swap(KnownZero2, KnownOne2);
208 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
209 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
211 // Compute known bits of the carry.
212 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
213 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
215 // Compute set of known bits (where all three relevant bits are known).
216 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
217 APInt RHSKnown = KnownZero2 | KnownOne2;
218 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
219 APInt Known = LHSKnown & RHSKnown & CarryKnown;
221 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
222 "known bits of sum differ");
224 // Compute known bits of the result.
225 KnownZero = ~PossibleSumOne & Known;
226 KnownOne = PossibleSumOne & Known;
228 // Are we still trying to solve for the sign bit?
229 if (!Known.isNegative()) {
231 // Adding two non-negative numbers, or subtracting a negative number from
232 // a non-negative one, can't wrap into negative.
233 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
234 KnownZero |= APInt::getSignBit(BitWidth);
235 // Adding two negative numbers, or subtracting a non-negative number from
236 // a negative one, can't wrap into non-negative.
237 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
238 KnownOne |= APInt::getSignBit(BitWidth);
243 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
244 APInt &KnownZero, APInt &KnownOne,
245 APInt &KnownZero2, APInt &KnownOne2,
246 const DataLayout *TD, unsigned Depth,
248 unsigned BitWidth = KnownZero.getBitWidth();
249 computeKnownBits(Op1, KnownZero, KnownOne, TD, Depth+1, Q);
250 computeKnownBits(Op0, KnownZero2, KnownOne2, TD, Depth+1, Q);
252 bool isKnownNegative = false;
253 bool isKnownNonNegative = false;
254 // If the multiplication is known not to overflow, compute the sign bit.
257 // The product of a number with itself is non-negative.
258 isKnownNonNegative = true;
260 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
261 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
262 bool isKnownNegativeOp1 = KnownOne.isNegative();
263 bool isKnownNegativeOp0 = KnownOne2.isNegative();
264 // The product of two numbers with the same sign is non-negative.
265 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
266 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
267 // The product of a negative number and a non-negative number is either
269 if (!isKnownNonNegative)
270 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
271 isKnownNonZero(Op0, TD, Depth, Q)) ||
272 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
273 isKnownNonZero(Op1, TD, Depth, Q));
277 // If low bits are zero in either operand, output low known-0 bits.
278 // Also compute a conserative estimate for high known-0 bits.
279 // More trickiness is possible, but this is sufficient for the
280 // interesting case of alignment computation.
281 KnownOne.clearAllBits();
282 unsigned TrailZ = KnownZero.countTrailingOnes() +
283 KnownZero2.countTrailingOnes();
284 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
285 KnownZero2.countLeadingOnes(),
286 BitWidth) - BitWidth;
288 TrailZ = std::min(TrailZ, BitWidth);
289 LeadZ = std::min(LeadZ, BitWidth);
290 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
291 APInt::getHighBitsSet(BitWidth, LeadZ);
293 // Only make use of no-wrap flags if we failed to compute the sign bit
294 // directly. This matters if the multiplication always overflows, in
295 // which case we prefer to follow the result of the direct computation,
296 // though as the program is invoking undefined behaviour we can choose
297 // whatever we like here.
298 if (isKnownNonNegative && !KnownOne.isNegative())
299 KnownZero.setBit(BitWidth - 1);
300 else if (isKnownNegative && !KnownZero.isNegative())
301 KnownOne.setBit(BitWidth - 1);
304 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
306 unsigned BitWidth = KnownZero.getBitWidth();
307 unsigned NumRanges = Ranges.getNumOperands() / 2;
308 assert(NumRanges >= 1);
310 // Use the high end of the ranges to find leading zeros.
311 unsigned MinLeadingZeros = BitWidth;
312 for (unsigned i = 0; i < NumRanges; ++i) {
313 ConstantInt *Lower = cast<ConstantInt>(Ranges.getOperand(2*i + 0));
314 ConstantInt *Upper = cast<ConstantInt>(Ranges.getOperand(2*i + 1));
315 ConstantRange Range(Lower->getValue(), Upper->getValue());
316 if (Range.isWrappedSet())
317 MinLeadingZeros = 0; // -1 has no zeros
318 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
319 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
322 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
325 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
326 SmallVector<const Value *, 16> WorkSet(1, I);
327 SmallPtrSet<const Value *, 32> Visited;
328 SmallPtrSet<const Value *, 16> EphValues;
330 while (!WorkSet.empty()) {
331 const Value *V = WorkSet.pop_back_val();
332 if (!Visited.insert(V))
335 // If all uses of this value are ephemeral, then so is this value.
336 bool FoundNEUse = false;
337 for (const User *I : V->users())
338 if (!EphValues.count(I)) {
348 if (const User *U = dyn_cast<User>(V))
349 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
351 if (isSafeToSpeculativelyExecute(*J))
352 WorkSet.push_back(*J);
360 // Is this an intrinsic that cannot be speculated but also cannot trap?
361 static bool isAssumeLikeIntrinsic(const Instruction *I) {
362 if (const CallInst *CI = dyn_cast<CallInst>(I))
363 if (Function *F = CI->getCalledFunction())
364 switch (F->getIntrinsicID()) {
366 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
367 case Intrinsic::assume:
368 case Intrinsic::dbg_declare:
369 case Intrinsic::dbg_value:
370 case Intrinsic::invariant_start:
371 case Intrinsic::invariant_end:
372 case Intrinsic::lifetime_start:
373 case Intrinsic::lifetime_end:
374 case Intrinsic::objectsize:
375 case Intrinsic::ptr_annotation:
376 case Intrinsic::var_annotation:
383 static bool isValidAssumeForContext(Value *V, const Query &Q,
384 const DataLayout *DL) {
385 Instruction *Inv = cast<Instruction>(V);
387 // There are two restrictions on the use of an assume:
388 // 1. The assume must dominate the context (or the control flow must
389 // reach the assume whenever it reaches the context).
390 // 2. The context must not be in the assume's set of ephemeral values
391 // (otherwise we will use the assume to prove that the condition
392 // feeding the assume is trivially true, thus causing the removal of
396 if (Q.DT->dominates(Inv, Q.CxtI)) {
398 } else if (Inv->getParent() == Q.CxtI->getParent()) {
399 // The context comes first, but they're both in the same block. Make sure
400 // there is nothing in between that might interrupt the control flow.
401 for (BasicBlock::const_iterator I =
402 std::next(BasicBlock::const_iterator(Q.CxtI)),
403 IE(Inv); I != IE; ++I)
404 if (!isSafeToSpeculativelyExecute(I, DL) &&
405 !isAssumeLikeIntrinsic(I))
408 return !isEphemeralValueOf(Inv, Q.CxtI);
414 // When we don't have a DT, we do a limited search...
415 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
417 } else if (Inv->getParent() == Q.CxtI->getParent()) {
418 // Search forward from the assume until we reach the context (or the end
419 // of the block); the common case is that the assume will come first.
420 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
421 IE = Inv->getParent()->end(); I != IE; ++I)
425 // The context must come first...
426 for (BasicBlock::const_iterator I =
427 std::next(BasicBlock::const_iterator(Q.CxtI)),
428 IE(Inv); I != IE; ++I)
429 if (!isSafeToSpeculativelyExecute(I, DL) &&
430 !isAssumeLikeIntrinsic(I))
433 return !isEphemeralValueOf(Inv, Q.CxtI);
439 bool llvm::isValidAssumeForContext(const Instruction *I,
440 const Instruction *CxtI,
441 const DataLayout *DL,
442 const DominatorTree *DT) {
443 return ::isValidAssumeForContext(const_cast<Instruction*>(I),
444 Query(nullptr, CxtI, DT), DL);
447 template<typename LHS, typename RHS>
448 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
449 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
450 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
451 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
454 template<typename LHS, typename RHS>
455 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
456 BinaryOp_match<RHS, LHS, Instruction::And>>
457 m_c_And(const LHS &L, const RHS &R) {
458 return m_CombineOr(m_And(L, R), m_And(R, L));
461 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
463 const DataLayout *DL,
464 unsigned Depth, const Query &Q) {
465 // Use of assumptions is context-sensitive. If we don't have a context, we
467 if (!Q.AT || !Q.CxtI)
470 unsigned BitWidth = KnownZero.getBitWidth();
472 Function *F = const_cast<Function*>(Q.CxtI->getParent()->getParent());
473 for (auto &CI : Q.AT->assumptions(F)) {
475 if (Q.ExclInvs.count(I))
478 if (match(I, m_Intrinsic<Intrinsic::assume>(m_Specific(V))) &&
479 isValidAssumeForContext(I, Q, DL)) {
480 assert(BitWidth == 1 && "assume operand is not i1?");
481 KnownZero.clearAllBits();
482 KnownOne.setAllBits();
487 auto m_V = m_CombineOr(m_Specific(V),
488 m_CombineOr(m_PtrToInt(m_Specific(V)),
489 m_BitCast(m_Specific(V))));
491 CmpInst::Predicate Pred;
493 if (match(I, m_Intrinsic<Intrinsic::assume>(
494 m_c_ICmp(Pred, m_V, m_Value(A)))) &&
495 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
496 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
497 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
498 KnownZero |= RHSKnownZero;
499 KnownOne |= RHSKnownOne;
501 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
502 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A)))) &&
503 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
504 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
505 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
506 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
507 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
509 // For those bits in the mask that are known to be one, we can propagate
510 // known bits from the RHS to V.
511 KnownZero |= RHSKnownZero & MaskKnownOne;
512 KnownOne |= RHSKnownOne & MaskKnownOne;
517 /// Determine which bits of V are known to be either zero or one and return
518 /// them in the KnownZero/KnownOne bit sets.
520 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
521 /// we cannot optimize based on the assumption that it is zero without changing
522 /// it to be an explicit zero. If we don't change it to zero, other code could
523 /// optimized based on the contradictory assumption that it is non-zero.
524 /// Because instcombine aggressively folds operations with undef args anyway,
525 /// this won't lose us code quality.
527 /// This function is defined on values with integer type, values with pointer
528 /// type (but only if TD is non-null), and vectors of integers. In the case
529 /// where V is a vector, known zero, and known one values are the
530 /// same width as the vector element, and the bit is set only if it is true
531 /// for all of the elements in the vector.
532 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
533 const DataLayout *TD, unsigned Depth,
535 assert(V && "No Value?");
536 assert(Depth <= MaxDepth && "Limit Search Depth");
537 unsigned BitWidth = KnownZero.getBitWidth();
539 assert((V->getType()->isIntOrIntVectorTy() ||
540 V->getType()->getScalarType()->isPointerTy()) &&
541 "Not integer or pointer type!");
543 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
544 (!V->getType()->isIntOrIntVectorTy() ||
545 V->getType()->getScalarSizeInBits() == BitWidth) &&
546 KnownZero.getBitWidth() == BitWidth &&
547 KnownOne.getBitWidth() == BitWidth &&
548 "V, KnownOne and KnownZero should have same BitWidth");
550 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
551 // We know all of the bits for a constant!
552 KnownOne = CI->getValue();
553 KnownZero = ~KnownOne;
556 // Null and aggregate-zero are all-zeros.
557 if (isa<ConstantPointerNull>(V) ||
558 isa<ConstantAggregateZero>(V)) {
559 KnownOne.clearAllBits();
560 KnownZero = APInt::getAllOnesValue(BitWidth);
563 // Handle a constant vector by taking the intersection of the known bits of
564 // each element. There is no real need to handle ConstantVector here, because
565 // we don't handle undef in any particularly useful way.
566 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
567 // We know that CDS must be a vector of integers. Take the intersection of
569 KnownZero.setAllBits(); KnownOne.setAllBits();
570 APInt Elt(KnownZero.getBitWidth(), 0);
571 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
572 Elt = CDS->getElementAsInteger(i);
579 // The address of an aligned GlobalValue has trailing zeros.
580 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
581 unsigned Align = GV->getAlignment();
582 if (Align == 0 && TD) {
583 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
584 Type *ObjectType = GVar->getType()->getElementType();
585 if (ObjectType->isSized()) {
586 // If the object is defined in the current Module, we'll be giving
587 // it the preferred alignment. Otherwise, we have to assume that it
588 // may only have the minimum ABI alignment.
589 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
590 Align = TD->getPreferredAlignment(GVar);
592 Align = TD->getABITypeAlignment(ObjectType);
597 KnownZero = APInt::getLowBitsSet(BitWidth,
598 countTrailingZeros(Align));
600 KnownZero.clearAllBits();
601 KnownOne.clearAllBits();
604 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
605 // the bits of its aliasee.
606 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
607 if (GA->mayBeOverridden()) {
608 KnownZero.clearAllBits(); KnownOne.clearAllBits();
610 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1, Q);
615 if (Argument *A = dyn_cast<Argument>(V)) {
616 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
618 if (!Align && TD && A->hasStructRetAttr()) {
619 // An sret parameter has at least the ABI alignment of the return type.
620 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
621 if (EltTy->isSized())
622 Align = TD->getABITypeAlignment(EltTy);
626 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
628 // Don't give up yet... there might be an assumption that provides more
630 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
634 // Start out not knowing anything.
635 KnownZero.clearAllBits(); KnownOne.clearAllBits();
637 if (Depth == MaxDepth)
638 return; // Limit search depth.
640 // Check whether a nearby assume intrinsic can determine some known bits.
641 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
643 Operator *I = dyn_cast<Operator>(V);
646 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
647 switch (I->getOpcode()) {
649 case Instruction::Load:
650 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
651 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
653 case Instruction::And: {
654 // If either the LHS or the RHS are Zero, the result is zero.
655 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
656 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
658 // Output known-1 bits are only known if set in both the LHS & RHS.
659 KnownOne &= KnownOne2;
660 // Output known-0 are known to be clear if zero in either the LHS | RHS.
661 KnownZero |= KnownZero2;
664 case Instruction::Or: {
665 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
666 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
668 // Output known-0 bits are only known if clear in both the LHS & RHS.
669 KnownZero &= KnownZero2;
670 // Output known-1 are known to be set if set in either the LHS | RHS.
671 KnownOne |= KnownOne2;
674 case Instruction::Xor: {
675 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
676 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
678 // Output known-0 bits are known if clear or set in both the LHS & RHS.
679 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
680 // Output known-1 are known to be set if set in only one of the LHS, RHS.
681 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
682 KnownZero = KnownZeroOut;
685 case Instruction::Mul: {
686 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
687 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
688 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
692 case Instruction::UDiv: {
693 // For the purposes of computing leading zeros we can conservatively
694 // treat a udiv as a logical right shift by the power of 2 known to
695 // be less than the denominator.
696 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
697 unsigned LeadZ = KnownZero2.countLeadingOnes();
699 KnownOne2.clearAllBits();
700 KnownZero2.clearAllBits();
701 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
702 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
703 if (RHSUnknownLeadingOnes != BitWidth)
704 LeadZ = std::min(BitWidth,
705 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
707 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
710 case Instruction::Select:
711 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1, Q);
712 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
714 // Only known if known in both the LHS and RHS.
715 KnownOne &= KnownOne2;
716 KnownZero &= KnownZero2;
718 case Instruction::FPTrunc:
719 case Instruction::FPExt:
720 case Instruction::FPToUI:
721 case Instruction::FPToSI:
722 case Instruction::SIToFP:
723 case Instruction::UIToFP:
724 break; // Can't work with floating point.
725 case Instruction::PtrToInt:
726 case Instruction::IntToPtr:
727 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
728 // We can't handle these if we don't know the pointer size.
730 // FALL THROUGH and handle them the same as zext/trunc.
731 case Instruction::ZExt:
732 case Instruction::Trunc: {
733 Type *SrcTy = I->getOperand(0)->getType();
735 unsigned SrcBitWidth;
736 // Note that we handle pointer operands here because of inttoptr/ptrtoint
737 // which fall through here.
739 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
741 SrcBitWidth = SrcTy->getScalarSizeInBits();
742 if (!SrcBitWidth) break;
745 assert(SrcBitWidth && "SrcBitWidth can't be zero");
746 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
747 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
748 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
749 KnownZero = KnownZero.zextOrTrunc(BitWidth);
750 KnownOne = KnownOne.zextOrTrunc(BitWidth);
751 // Any top bits are known to be zero.
752 if (BitWidth > SrcBitWidth)
753 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
756 case Instruction::BitCast: {
757 Type *SrcTy = I->getOperand(0)->getType();
758 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
759 // TODO: For now, not handling conversions like:
760 // (bitcast i64 %x to <2 x i32>)
761 !I->getType()->isVectorTy()) {
762 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
767 case Instruction::SExt: {
768 // Compute the bits in the result that are not present in the input.
769 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
771 KnownZero = KnownZero.trunc(SrcBitWidth);
772 KnownOne = KnownOne.trunc(SrcBitWidth);
773 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
774 KnownZero = KnownZero.zext(BitWidth);
775 KnownOne = KnownOne.zext(BitWidth);
777 // If the sign bit of the input is known set or clear, then we know the
778 // top bits of the result.
779 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
780 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
781 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
782 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
785 case Instruction::Shl:
786 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
787 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
788 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
789 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
790 KnownZero <<= ShiftAmt;
791 KnownOne <<= ShiftAmt;
792 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
796 case Instruction::LShr:
797 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
798 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
799 // Compute the new bits that are at the top now.
800 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
802 // Unsigned shift right.
803 computeKnownBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1, Q);
804 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
805 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
806 // high bits known zero.
807 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
811 case Instruction::AShr:
812 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
813 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
814 // Compute the new bits that are at the top now.
815 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
817 // Signed shift right.
818 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
819 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
820 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
822 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
823 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
824 KnownZero |= HighBits;
825 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
826 KnownOne |= HighBits;
830 case Instruction::Sub: {
831 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
832 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
833 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
837 case Instruction::Add: {
838 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
839 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
840 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
844 case Instruction::SRem:
845 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
846 APInt RA = Rem->getValue().abs();
847 if (RA.isPowerOf2()) {
848 APInt LowBits = RA - 1;
849 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD,
852 // The low bits of the first operand are unchanged by the srem.
853 KnownZero = KnownZero2 & LowBits;
854 KnownOne = KnownOne2 & LowBits;
856 // If the first operand is non-negative or has all low bits zero, then
857 // the upper bits are all zero.
858 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
859 KnownZero |= ~LowBits;
861 // If the first operand is negative and not all low bits are zero, then
862 // the upper bits are all one.
863 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
864 KnownOne |= ~LowBits;
866 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
870 // The sign bit is the LHS's sign bit, except when the result of the
871 // remainder is zero.
872 if (KnownZero.isNonNegative()) {
873 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
874 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
876 // If it's known zero, our sign bit is also zero.
877 if (LHSKnownZero.isNegative())
878 KnownZero.setBit(BitWidth - 1);
882 case Instruction::URem: {
883 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
884 APInt RA = Rem->getValue();
885 if (RA.isPowerOf2()) {
886 APInt LowBits = (RA - 1);
887 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
889 KnownZero |= ~LowBits;
895 // Since the result is less than or equal to either operand, any leading
896 // zero bits in either operand must also exist in the result.
897 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
898 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
900 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
901 KnownZero2.countLeadingOnes());
902 KnownOne.clearAllBits();
903 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
907 case Instruction::Alloca: {
908 AllocaInst *AI = cast<AllocaInst>(V);
909 unsigned Align = AI->getAlignment();
910 if (Align == 0 && TD)
911 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
914 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
917 case Instruction::GetElementPtr: {
918 // Analyze all of the subscripts of this getelementptr instruction
919 // to determine if we can prove known low zero bits.
920 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
921 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
923 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
925 gep_type_iterator GTI = gep_type_begin(I);
926 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
927 Value *Index = I->getOperand(i);
928 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
929 // Handle struct member offset arithmetic.
935 // Handle case when index is vector zeroinitializer
936 Constant *CIndex = cast<Constant>(Index);
937 if (CIndex->isZeroValue())
940 if (CIndex->getType()->isVectorTy())
941 Index = CIndex->getSplatValue();
943 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
944 const StructLayout *SL = TD->getStructLayout(STy);
945 uint64_t Offset = SL->getElementOffset(Idx);
946 TrailZ = std::min<unsigned>(TrailZ,
947 countTrailingZeros(Offset));
949 // Handle array index arithmetic.
950 Type *IndexedTy = GTI.getIndexedType();
951 if (!IndexedTy->isSized()) {
955 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
956 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
957 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
958 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1, Q);
959 TrailZ = std::min(TrailZ,
960 unsigned(countTrailingZeros(TypeSize) +
961 LocalKnownZero.countTrailingOnes()));
965 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
968 case Instruction::PHI: {
969 PHINode *P = cast<PHINode>(I);
970 // Handle the case of a simple two-predecessor recurrence PHI.
971 // There's a lot more that could theoretically be done here, but
972 // this is sufficient to catch some interesting cases.
973 if (P->getNumIncomingValues() == 2) {
974 for (unsigned i = 0; i != 2; ++i) {
975 Value *L = P->getIncomingValue(i);
976 Value *R = P->getIncomingValue(!i);
977 Operator *LU = dyn_cast<Operator>(L);
980 unsigned Opcode = LU->getOpcode();
981 // Check for operations that have the property that if
982 // both their operands have low zero bits, the result
983 // will have low zero bits.
984 if (Opcode == Instruction::Add ||
985 Opcode == Instruction::Sub ||
986 Opcode == Instruction::And ||
987 Opcode == Instruction::Or ||
988 Opcode == Instruction::Mul) {
989 Value *LL = LU->getOperand(0);
990 Value *LR = LU->getOperand(1);
991 // Find a recurrence.
998 // Ok, we have a PHI of the form L op= R. Check for low
1000 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1, Q);
1002 // We need to take the minimum number of known bits
1003 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1004 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1, Q);
1006 KnownZero = APInt::getLowBitsSet(BitWidth,
1007 std::min(KnownZero2.countTrailingOnes(),
1008 KnownZero3.countTrailingOnes()));
1014 // Unreachable blocks may have zero-operand PHI nodes.
1015 if (P->getNumIncomingValues() == 0)
1018 // Otherwise take the unions of the known bit sets of the operands,
1019 // taking conservative care to avoid excessive recursion.
1020 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1021 // Skip if every incoming value references to ourself.
1022 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1025 KnownZero = APInt::getAllOnesValue(BitWidth);
1026 KnownOne = APInt::getAllOnesValue(BitWidth);
1027 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1028 // Skip direct self references.
1029 if (P->getIncomingValue(i) == P) continue;
1031 KnownZero2 = APInt(BitWidth, 0);
1032 KnownOne2 = APInt(BitWidth, 0);
1033 // Recurse, but cap the recursion to one level, because we don't
1034 // want to waste time spinning around in loops.
1035 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
1037 KnownZero &= KnownZero2;
1038 KnownOne &= KnownOne2;
1039 // If all bits have been ruled out, there's no need to check
1041 if (!KnownZero && !KnownOne)
1047 case Instruction::Call:
1048 case Instruction::Invoke:
1049 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1050 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1051 // If a range metadata is attached to this IntrinsicInst, intersect the
1052 // explicit range specified by the metadata and the implicit range of
1054 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1055 switch (II->getIntrinsicID()) {
1057 case Intrinsic::ctlz:
1058 case Intrinsic::cttz: {
1059 unsigned LowBits = Log2_32(BitWidth)+1;
1060 // If this call is undefined for 0, the result will be less than 2^n.
1061 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1063 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1066 case Intrinsic::ctpop: {
1067 unsigned LowBits = Log2_32(BitWidth)+1;
1068 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1071 case Intrinsic::x86_sse42_crc32_64_64:
1072 KnownZero |= APInt::getHighBitsSet(64, 32);
1077 case Instruction::ExtractValue:
1078 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1079 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1080 if (EVI->getNumIndices() != 1) break;
1081 if (EVI->getIndices()[0] == 0) {
1082 switch (II->getIntrinsicID()) {
1084 case Intrinsic::uadd_with_overflow:
1085 case Intrinsic::sadd_with_overflow:
1086 computeKnownBitsAddSub(true, II->getArgOperand(0),
1087 II->getArgOperand(1), false, KnownZero,
1088 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1090 case Intrinsic::usub_with_overflow:
1091 case Intrinsic::ssub_with_overflow:
1092 computeKnownBitsAddSub(false, II->getArgOperand(0),
1093 II->getArgOperand(1), false, KnownZero,
1094 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1096 case Intrinsic::umul_with_overflow:
1097 case Intrinsic::smul_with_overflow:
1098 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
1099 false, KnownZero, KnownOne,
1100 KnownZero2, KnownOne2, TD, Depth, Q);
1107 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1110 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
1111 /// one. Convenience wrapper around computeKnownBits.
1112 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1113 const DataLayout *TD, unsigned Depth,
1115 unsigned BitWidth = getBitWidth(V->getType(), TD);
1121 APInt ZeroBits(BitWidth, 0);
1122 APInt OneBits(BitWidth, 0);
1123 computeKnownBits(V, ZeroBits, OneBits, TD, Depth, Q);
1124 KnownOne = OneBits[BitWidth - 1];
1125 KnownZero = ZeroBits[BitWidth - 1];
1128 /// isKnownToBeAPowerOfTwo - Return true if the given value is known to have exactly one
1129 /// bit set when defined. For vectors return true if every element is known to
1130 /// be a power of two when defined. Supports values with integer or pointer
1131 /// types and vectors of integers.
1132 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1134 if (Constant *C = dyn_cast<Constant>(V)) {
1135 if (C->isNullValue())
1137 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1138 return CI->getValue().isPowerOf2();
1139 // TODO: Handle vector constants.
1142 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1143 // it is shifted off the end then the result is undefined.
1144 if (match(V, m_Shl(m_One(), m_Value())))
1147 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1148 // bottom. If it is shifted off the bottom then the result is undefined.
1149 if (match(V, m_LShr(m_SignBit(), m_Value())))
1152 // The remaining tests are all recursive, so bail out if we hit the limit.
1153 if (Depth++ == MaxDepth)
1156 Value *X = nullptr, *Y = nullptr;
1157 // A shift of a power of two is a power of two or zero.
1158 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1159 match(V, m_Shr(m_Value(X), m_Value()))))
1160 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q);
1162 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1163 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1165 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1167 isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1168 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1170 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1171 // A power of two and'd with anything is a power of two or zero.
1172 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q) ||
1173 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth, Q))
1175 // X & (-X) is always a power of two or zero.
1176 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1181 // Adding a power-of-two or zero to the same power-of-two or zero yields
1182 // either the original power-of-two, a larger power-of-two or zero.
1183 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1184 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1185 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1186 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1187 match(X, m_And(m_Value(), m_Specific(Y))))
1188 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1190 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1191 match(Y, m_And(m_Value(), m_Specific(X))))
1192 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1195 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1196 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1197 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth, Q);
1199 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1200 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth, Q);
1201 // If i8 V is a power of two or zero:
1202 // ZeroBits: 1 1 1 0 1 1 1 1
1203 // ~ZeroBits: 0 0 0 1 0 0 0 0
1204 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1205 // If OrZero isn't set, we cannot give back a zero result.
1206 // Make sure either the LHS or RHS has a bit set.
1207 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1212 // An exact divide or right shift can only shift off zero bits, so the result
1213 // is a power of two only if the first operand is a power of two and not
1214 // copying a sign bit (sdiv int_min, 2).
1215 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1216 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1217 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1224 /// \brief Test whether a GEP's result is known to be non-null.
1226 /// Uses properties inherent in a GEP to try to determine whether it is known
1229 /// Currently this routine does not support vector GEPs.
1230 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
1231 unsigned Depth, const Query &Q) {
1232 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1235 // FIXME: Support vector-GEPs.
1236 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1238 // If the base pointer is non-null, we cannot walk to a null address with an
1239 // inbounds GEP in address space zero.
1240 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1243 // Past this, if we don't have DataLayout, we can't do much.
1247 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1248 // If so, then the GEP cannot produce a null pointer, as doing so would
1249 // inherently violate the inbounds contract within address space zero.
1250 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1251 GTI != GTE; ++GTI) {
1252 // Struct types are easy -- they must always be indexed by a constant.
1253 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1254 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1255 unsigned ElementIdx = OpC->getZExtValue();
1256 const StructLayout *SL = DL->getStructLayout(STy);
1257 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1258 if (ElementOffset > 0)
1263 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1264 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
1267 // Fast path the constant operand case both for efficiency and so we don't
1268 // increment Depth when just zipping down an all-constant GEP.
1269 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1275 // We post-increment Depth here because while isKnownNonZero increments it
1276 // as well, when we pop back up that increment won't persist. We don't want
1277 // to recurse 10k times just because we have 10k GEP operands. We don't
1278 // bail completely out because we want to handle constant GEPs regardless
1280 if (Depth++ >= MaxDepth)
1283 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1290 /// isKnownNonZero - Return true if the given value is known to be non-zero
1291 /// when defined. For vectors return true if every element is known to be
1292 /// non-zero when defined. Supports values with integer or pointer type and
1293 /// vectors of integers.
1294 bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
1296 if (Constant *C = dyn_cast<Constant>(V)) {
1297 if (C->isNullValue())
1299 if (isa<ConstantInt>(C))
1300 // Must be non-zero due to null test above.
1302 // TODO: Handle vectors
1306 // The remaining tests are all recursive, so bail out if we hit the limit.
1307 if (Depth++ >= MaxDepth)
1310 // Check for pointer simplifications.
1311 if (V->getType()->isPointerTy()) {
1312 if (isKnownNonNull(V))
1314 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1315 if (isGEPKnownNonNull(GEP, TD, Depth, Q))
1319 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1321 // X | Y != 0 if X != 0 or Y != 0.
1322 Value *X = nullptr, *Y = nullptr;
1323 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1324 return isKnownNonZero(X, TD, Depth, Q) ||
1325 isKnownNonZero(Y, TD, Depth, Q);
1327 // ext X != 0 if X != 0.
1328 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1329 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth, Q);
1331 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1332 // if the lowest bit is shifted off the end.
1333 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1334 // shl nuw can't remove any non-zero bits.
1335 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1336 if (BO->hasNoUnsignedWrap())
1337 return isKnownNonZero(X, TD, Depth, Q);
1339 APInt KnownZero(BitWidth, 0);
1340 APInt KnownOne(BitWidth, 0);
1341 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1345 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1346 // defined if the sign bit is shifted off the end.
1347 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1348 // shr exact can only shift out zero bits.
1349 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1351 return isKnownNonZero(X, TD, Depth, Q);
1353 bool XKnownNonNegative, XKnownNegative;
1354 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1358 // div exact can only produce a zero if the dividend is zero.
1359 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1360 return isKnownNonZero(X, TD, Depth, Q);
1363 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1364 bool XKnownNonNegative, XKnownNegative;
1365 bool YKnownNonNegative, YKnownNegative;
1366 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1367 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth, Q);
1369 // If X and Y are both non-negative (as signed values) then their sum is not
1370 // zero unless both X and Y are zero.
1371 if (XKnownNonNegative && YKnownNonNegative)
1372 if (isKnownNonZero(X, TD, Depth, Q) ||
1373 isKnownNonZero(Y, TD, Depth, Q))
1376 // If X and Y are both negative (as signed values) then their sum is not
1377 // zero unless both X and Y equal INT_MIN.
1378 if (BitWidth && XKnownNegative && YKnownNegative) {
1379 APInt KnownZero(BitWidth, 0);
1380 APInt KnownOne(BitWidth, 0);
1381 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1382 // The sign bit of X is set. If some other bit is set then X is not equal
1384 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1385 if ((KnownOne & Mask) != 0)
1387 // The sign bit of Y is set. If some other bit is set then Y is not equal
1389 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth, Q);
1390 if ((KnownOne & Mask) != 0)
1394 // The sum of a non-negative number and a power of two is not zero.
1395 if (XKnownNonNegative &&
1396 isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth, Q))
1398 if (YKnownNonNegative &&
1399 isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth, Q))
1403 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1404 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1405 // If X and Y are non-zero then so is X * Y as long as the multiplication
1406 // does not overflow.
1407 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1408 isKnownNonZero(X, TD, Depth, Q) &&
1409 isKnownNonZero(Y, TD, Depth, Q))
1412 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1413 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1414 if (isKnownNonZero(SI->getTrueValue(), TD, Depth, Q) &&
1415 isKnownNonZero(SI->getFalseValue(), TD, Depth, Q))
1419 if (!BitWidth) return false;
1420 APInt KnownZero(BitWidth, 0);
1421 APInt KnownOne(BitWidth, 0);
1422 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1423 return KnownOne != 0;
1426 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1427 /// this predicate to simplify operations downstream. Mask is known to be zero
1428 /// for bits that V cannot have.
1430 /// This function is defined on values with integer type, values with pointer
1431 /// type (but only if TD is non-null), and vectors of integers. In the case
1432 /// where V is a vector, the mask, known zero, and known one values are the
1433 /// same width as the vector element, and the bit is set only if it is true
1434 /// for all of the elements in the vector.
1435 bool MaskedValueIsZero(Value *V, const APInt &Mask,
1436 const DataLayout *TD, unsigned Depth,
1438 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1439 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1440 return (KnownZero & Mask) == Mask;
1445 /// ComputeNumSignBits - Return the number of times the sign bit of the
1446 /// register is replicated into the other bits. We know that at least 1 bit
1447 /// is always equal to the sign bit (itself), but other cases can give us
1448 /// information. For example, immediately after an "ashr X, 2", we know that
1449 /// the top 3 bits are all equal to each other, so we return 3.
1451 /// 'Op' must have a scalar integer type.
1453 unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
1454 unsigned Depth, const Query &Q) {
1455 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1456 "ComputeNumSignBits requires a DataLayout object to operate "
1457 "on non-integer values!");
1458 Type *Ty = V->getType();
1459 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1460 Ty->getScalarSizeInBits();
1462 unsigned FirstAnswer = 1;
1464 // Note that ConstantInt is handled by the general computeKnownBits case
1468 return 1; // Limit search depth.
1470 Operator *U = dyn_cast<Operator>(V);
1471 switch (Operator::getOpcode(V)) {
1473 case Instruction::SExt:
1474 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1475 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q) + Tmp;
1477 case Instruction::AShr: {
1478 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1479 // ashr X, C -> adds C sign bits. Vectors too.
1481 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1482 Tmp += ShAmt->getZExtValue();
1483 if (Tmp > TyBits) Tmp = TyBits;
1487 case Instruction::Shl: {
1489 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1490 // shl destroys sign bits.
1491 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1492 Tmp2 = ShAmt->getZExtValue();
1493 if (Tmp2 >= TyBits || // Bad shift.
1494 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1499 case Instruction::And:
1500 case Instruction::Or:
1501 case Instruction::Xor: // NOT is handled here.
1502 // Logical binary ops preserve the number of sign bits at the worst.
1503 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1505 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1506 FirstAnswer = std::min(Tmp, Tmp2);
1507 // We computed what we know about the sign bits as our first
1508 // answer. Now proceed to the generic code that uses
1509 // computeKnownBits, and pick whichever answer is better.
1513 case Instruction::Select:
1514 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1515 if (Tmp == 1) return 1; // Early out.
1516 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1, Q);
1517 return std::min(Tmp, Tmp2);
1519 case Instruction::Add:
1520 // Add can have at most one carry bit. Thus we know that the output
1521 // is, at worst, one more bit than the inputs.
1522 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1523 if (Tmp == 1) return 1; // Early out.
1525 // Special case decrementing a value (ADD X, -1):
1526 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1527 if (CRHS->isAllOnesValue()) {
1528 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1529 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1531 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1533 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1536 // If we are subtracting one from a positive number, there is no carry
1537 // out of the result.
1538 if (KnownZero.isNegative())
1542 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1543 if (Tmp2 == 1) return 1;
1544 return std::min(Tmp, Tmp2)-1;
1546 case Instruction::Sub:
1547 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1548 if (Tmp2 == 1) return 1;
1551 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1552 if (CLHS->isNullValue()) {
1553 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1554 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
1555 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1557 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1560 // If the input is known to be positive (the sign bit is known clear),
1561 // the output of the NEG has the same number of sign bits as the input.
1562 if (KnownZero.isNegative())
1565 // Otherwise, we treat this like a SUB.
1568 // Sub can have at most one carry bit. Thus we know that the output
1569 // is, at worst, one more bit than the inputs.
1570 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1571 if (Tmp == 1) return 1; // Early out.
1572 return std::min(Tmp, Tmp2)-1;
1574 case Instruction::PHI: {
1575 PHINode *PN = cast<PHINode>(U);
1576 // Don't analyze large in-degree PHIs.
1577 if (PN->getNumIncomingValues() > 4) break;
1579 // Take the minimum of all incoming values. This can't infinitely loop
1580 // because of our depth threshold.
1581 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1, Q);
1582 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1583 if (Tmp == 1) return Tmp;
1585 ComputeNumSignBits(PN->getIncomingValue(i), TD,
1591 case Instruction::Trunc:
1592 // FIXME: it's tricky to do anything useful for this, but it is an important
1593 // case for targets like X86.
1597 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1598 // use this information.
1599 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1601 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1603 if (KnownZero.isNegative()) { // sign bit is 0
1605 } else if (KnownOne.isNegative()) { // sign bit is 1;
1612 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1613 // the number of identical bits in the top of the input value.
1615 Mask <<= Mask.getBitWidth()-TyBits;
1616 // Return # leading zeros. We use 'min' here in case Val was zero before
1617 // shifting. We don't want to return '64' as for an i32 "0".
1618 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1621 /// ComputeMultiple - This function computes the integer multiple of Base that
1622 /// equals V. If successful, it returns true and returns the multiple in
1623 /// Multiple. If unsuccessful, it returns false. It looks
1624 /// through SExt instructions only if LookThroughSExt is true.
1625 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1626 bool LookThroughSExt, unsigned Depth) {
1627 const unsigned MaxDepth = 6;
1629 assert(V && "No Value?");
1630 assert(Depth <= MaxDepth && "Limit Search Depth");
1631 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1633 Type *T = V->getType();
1635 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1645 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1646 Constant *BaseVal = ConstantInt::get(T, Base);
1647 if (CO && CO == BaseVal) {
1649 Multiple = ConstantInt::get(T, 1);
1653 if (CI && CI->getZExtValue() % Base == 0) {
1654 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1658 if (Depth == MaxDepth) return false; // Limit search depth.
1660 Operator *I = dyn_cast<Operator>(V);
1661 if (!I) return false;
1663 switch (I->getOpcode()) {
1665 case Instruction::SExt:
1666 if (!LookThroughSExt) return false;
1667 // otherwise fall through to ZExt
1668 case Instruction::ZExt:
1669 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1670 LookThroughSExt, Depth+1);
1671 case Instruction::Shl:
1672 case Instruction::Mul: {
1673 Value *Op0 = I->getOperand(0);
1674 Value *Op1 = I->getOperand(1);
1676 if (I->getOpcode() == Instruction::Shl) {
1677 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1678 if (!Op1CI) return false;
1679 // Turn Op0 << Op1 into Op0 * 2^Op1
1680 APInt Op1Int = Op1CI->getValue();
1681 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1682 APInt API(Op1Int.getBitWidth(), 0);
1683 API.setBit(BitToSet);
1684 Op1 = ConstantInt::get(V->getContext(), API);
1687 Value *Mul0 = nullptr;
1688 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1689 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1690 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1691 if (Op1C->getType()->getPrimitiveSizeInBits() <
1692 MulC->getType()->getPrimitiveSizeInBits())
1693 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1694 if (Op1C->getType()->getPrimitiveSizeInBits() >
1695 MulC->getType()->getPrimitiveSizeInBits())
1696 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1698 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1699 Multiple = ConstantExpr::getMul(MulC, Op1C);
1703 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1704 if (Mul0CI->getValue() == 1) {
1705 // V == Base * Op1, so return Op1
1711 Value *Mul1 = nullptr;
1712 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1713 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1714 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1715 if (Op0C->getType()->getPrimitiveSizeInBits() <
1716 MulC->getType()->getPrimitiveSizeInBits())
1717 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1718 if (Op0C->getType()->getPrimitiveSizeInBits() >
1719 MulC->getType()->getPrimitiveSizeInBits())
1720 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1722 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1723 Multiple = ConstantExpr::getMul(MulC, Op0C);
1727 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1728 if (Mul1CI->getValue() == 1) {
1729 // V == Base * Op0, so return Op0
1737 // We could not determine if V is a multiple of Base.
1741 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1742 /// value is never equal to -0.0.
1744 /// NOTE: this function will need to be revisited when we support non-default
1747 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1748 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1749 return !CFP->getValueAPF().isNegZero();
1752 return 1; // Limit search depth.
1754 const Operator *I = dyn_cast<Operator>(V);
1755 if (!I) return false;
1757 // Check if the nsz fast-math flag is set
1758 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1759 if (FPO->hasNoSignedZeros())
1762 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1763 if (I->getOpcode() == Instruction::FAdd)
1764 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
1765 if (CFP->isNullValue())
1768 // sitofp and uitofp turn into +0.0 for zero.
1769 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1772 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1773 // sqrt(-0.0) = -0.0, no other negative results are possible.
1774 if (II->getIntrinsicID() == Intrinsic::sqrt)
1775 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1777 if (const CallInst *CI = dyn_cast<CallInst>(I))
1778 if (const Function *F = CI->getCalledFunction()) {
1779 if (F->isDeclaration()) {
1781 if (F->getName() == "abs") return true;
1782 // fabs[lf](x) != -0.0
1783 if (F->getName() == "fabs") return true;
1784 if (F->getName() == "fabsf") return true;
1785 if (F->getName() == "fabsl") return true;
1786 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
1787 F->getName() == "sqrtl")
1788 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
1795 /// isBytewiseValue - If the specified value can be set by repeating the same
1796 /// byte in memory, return the i8 value that it is represented with. This is
1797 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
1798 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
1799 /// byte store (e.g. i16 0x1234), return null.
1800 Value *llvm::isBytewiseValue(Value *V) {
1801 // All byte-wide stores are splatable, even of arbitrary variables.
1802 if (V->getType()->isIntegerTy(8)) return V;
1804 // Handle 'null' ConstantArrayZero etc.
1805 if (Constant *C = dyn_cast<Constant>(V))
1806 if (C->isNullValue())
1807 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
1809 // Constant float and double values can be handled as integer values if the
1810 // corresponding integer value is "byteable". An important case is 0.0.
1811 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1812 if (CFP->getType()->isFloatTy())
1813 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
1814 if (CFP->getType()->isDoubleTy())
1815 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
1816 // Don't handle long double formats, which have strange constraints.
1819 // We can handle constant integers that are power of two in size and a
1820 // multiple of 8 bits.
1821 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1822 unsigned Width = CI->getBitWidth();
1823 if (isPowerOf2_32(Width) && Width > 8) {
1824 // We can handle this value if the recursive binary decomposition is the
1825 // same at all levels.
1826 APInt Val = CI->getValue();
1828 while (Val.getBitWidth() != 8) {
1829 unsigned NextWidth = Val.getBitWidth()/2;
1830 Val2 = Val.lshr(NextWidth);
1831 Val2 = Val2.trunc(Val.getBitWidth()/2);
1832 Val = Val.trunc(Val.getBitWidth()/2);
1834 // If the top/bottom halves aren't the same, reject it.
1838 return ConstantInt::get(V->getContext(), Val);
1842 // A ConstantDataArray/Vector is splatable if all its members are equal and
1844 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
1845 Value *Elt = CA->getElementAsConstant(0);
1846 Value *Val = isBytewiseValue(Elt);
1850 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
1851 if (CA->getElementAsConstant(I) != Elt)
1857 // Conceptually, we could handle things like:
1858 // %a = zext i8 %X to i16
1859 // %b = shl i16 %a, 8
1860 // %c = or i16 %a, %b
1861 // but until there is an example that actually needs this, it doesn't seem
1862 // worth worrying about.
1867 // This is the recursive version of BuildSubAggregate. It takes a few different
1868 // arguments. Idxs is the index within the nested struct From that we are
1869 // looking at now (which is of type IndexedType). IdxSkip is the number of
1870 // indices from Idxs that should be left out when inserting into the resulting
1871 // struct. To is the result struct built so far, new insertvalue instructions
1873 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
1874 SmallVectorImpl<unsigned> &Idxs,
1876 Instruction *InsertBefore) {
1877 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
1879 // Save the original To argument so we can modify it
1881 // General case, the type indexed by Idxs is a struct
1882 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1883 // Process each struct element recursively
1886 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
1890 // Couldn't find any inserted value for this index? Cleanup
1891 while (PrevTo != OrigTo) {
1892 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
1893 PrevTo = Del->getAggregateOperand();
1894 Del->eraseFromParent();
1896 // Stop processing elements
1900 // If we successfully found a value for each of our subaggregates
1904 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
1905 // the struct's elements had a value that was inserted directly. In the latter
1906 // case, perhaps we can't determine each of the subelements individually, but
1907 // we might be able to find the complete struct somewhere.
1909 // Find the value that is at that particular spot
1910 Value *V = FindInsertedValue(From, Idxs);
1915 // Insert the value in the new (sub) aggregrate
1916 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
1917 "tmp", InsertBefore);
1920 // This helper takes a nested struct and extracts a part of it (which is again a
1921 // struct) into a new value. For example, given the struct:
1922 // { a, { b, { c, d }, e } }
1923 // and the indices "1, 1" this returns
1926 // It does this by inserting an insertvalue for each element in the resulting
1927 // struct, as opposed to just inserting a single struct. This will only work if
1928 // each of the elements of the substruct are known (ie, inserted into From by an
1929 // insertvalue instruction somewhere).
1931 // All inserted insertvalue instructions are inserted before InsertBefore
1932 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
1933 Instruction *InsertBefore) {
1934 assert(InsertBefore && "Must have someplace to insert!");
1935 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
1937 Value *To = UndefValue::get(IndexedType);
1938 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
1939 unsigned IdxSkip = Idxs.size();
1941 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
1944 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
1945 /// the scalar value indexed is already around as a register, for example if it
1946 /// were inserted directly into the aggregrate.
1948 /// If InsertBefore is not null, this function will duplicate (modified)
1949 /// insertvalues when a part of a nested struct is extracted.
1950 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
1951 Instruction *InsertBefore) {
1952 // Nothing to index? Just return V then (this is useful at the end of our
1954 if (idx_range.empty())
1956 // We have indices, so V should have an indexable type.
1957 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
1958 "Not looking at a struct or array?");
1959 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
1960 "Invalid indices for type?");
1962 if (Constant *C = dyn_cast<Constant>(V)) {
1963 C = C->getAggregateElement(idx_range[0]);
1964 if (!C) return nullptr;
1965 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
1968 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
1969 // Loop the indices for the insertvalue instruction in parallel with the
1970 // requested indices
1971 const unsigned *req_idx = idx_range.begin();
1972 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
1973 i != e; ++i, ++req_idx) {
1974 if (req_idx == idx_range.end()) {
1975 // We can't handle this without inserting insertvalues
1979 // The requested index identifies a part of a nested aggregate. Handle
1980 // this specially. For example,
1981 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
1982 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
1983 // %C = extractvalue {i32, { i32, i32 } } %B, 1
1984 // This can be changed into
1985 // %A = insertvalue {i32, i32 } undef, i32 10, 0
1986 // %C = insertvalue {i32, i32 } %A, i32 11, 1
1987 // which allows the unused 0,0 element from the nested struct to be
1989 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
1993 // This insert value inserts something else than what we are looking for.
1994 // See if the (aggregrate) value inserted into has the value we are
1995 // looking for, then.
1997 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2000 // If we end up here, the indices of the insertvalue match with those
2001 // requested (though possibly only partially). Now we recursively look at
2002 // the inserted value, passing any remaining indices.
2003 return FindInsertedValue(I->getInsertedValueOperand(),
2004 makeArrayRef(req_idx, idx_range.end()),
2008 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2009 // If we're extracting a value from an aggregrate that was extracted from
2010 // something else, we can extract from that something else directly instead.
2011 // However, we will need to chain I's indices with the requested indices.
2013 // Calculate the number of indices required
2014 unsigned size = I->getNumIndices() + idx_range.size();
2015 // Allocate some space to put the new indices in
2016 SmallVector<unsigned, 5> Idxs;
2018 // Add indices from the extract value instruction
2019 Idxs.append(I->idx_begin(), I->idx_end());
2021 // Add requested indices
2022 Idxs.append(idx_range.begin(), idx_range.end());
2024 assert(Idxs.size() == size
2025 && "Number of indices added not correct?");
2027 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2029 // Otherwise, we don't know (such as, extracting from a function return value
2030 // or load instruction)
2034 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
2035 /// it can be expressed as a base pointer plus a constant offset. Return the
2036 /// base and offset to the caller.
2037 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2038 const DataLayout *DL) {
2039 // Without DataLayout, conservatively assume 64-bit offsets, which is
2040 // the widest we support.
2041 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
2042 APInt ByteOffset(BitWidth, 0);
2044 if (Ptr->getType()->isVectorTy())
2047 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2049 APInt GEPOffset(BitWidth, 0);
2050 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
2053 ByteOffset += GEPOffset;
2056 Ptr = GEP->getPointerOperand();
2057 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2058 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2059 Ptr = cast<Operator>(Ptr)->getOperand(0);
2060 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2061 if (GA->mayBeOverridden())
2063 Ptr = GA->getAliasee();
2068 Offset = ByteOffset.getSExtValue();
2073 /// getConstantStringInfo - This function computes the length of a
2074 /// null-terminated C string pointed to by V. If successful, it returns true
2075 /// and returns the string in Str. If unsuccessful, it returns false.
2076 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2077 uint64_t Offset, bool TrimAtNul) {
2080 // Look through bitcast instructions and geps.
2081 V = V->stripPointerCasts();
2083 // If the value is a GEP instructionor constant expression, treat it as an
2085 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2086 // Make sure the GEP has exactly three arguments.
2087 if (GEP->getNumOperands() != 3)
2090 // Make sure the index-ee is a pointer to array of i8.
2091 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2092 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2093 if (!AT || !AT->getElementType()->isIntegerTy(8))
2096 // Check to make sure that the first operand of the GEP is an integer and
2097 // has value 0 so that we are sure we're indexing into the initializer.
2098 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2099 if (!FirstIdx || !FirstIdx->isZero())
2102 // If the second index isn't a ConstantInt, then this is a variable index
2103 // into the array. If this occurs, we can't say anything meaningful about
2105 uint64_t StartIdx = 0;
2106 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2107 StartIdx = CI->getZExtValue();
2110 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
2113 // The GEP instruction, constant or instruction, must reference a global
2114 // variable that is a constant and is initialized. The referenced constant
2115 // initializer is the array that we'll use for optimization.
2116 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2117 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2120 // Handle the all-zeros case
2121 if (GV->getInitializer()->isNullValue()) {
2122 // This is a degenerate case. The initializer is constant zero so the
2123 // length of the string must be zero.
2128 // Must be a Constant Array
2129 const ConstantDataArray *Array =
2130 dyn_cast<ConstantDataArray>(GV->getInitializer());
2131 if (!Array || !Array->isString())
2134 // Get the number of elements in the array
2135 uint64_t NumElts = Array->getType()->getArrayNumElements();
2137 // Start out with the entire array in the StringRef.
2138 Str = Array->getAsString();
2140 if (Offset > NumElts)
2143 // Skip over 'offset' bytes.
2144 Str = Str.substr(Offset);
2147 // Trim off the \0 and anything after it. If the array is not nul
2148 // terminated, we just return the whole end of string. The client may know
2149 // some other way that the string is length-bound.
2150 Str = Str.substr(0, Str.find('\0'));
2155 // These next two are very similar to the above, but also look through PHI
2157 // TODO: See if we can integrate these two together.
2159 /// GetStringLengthH - If we can compute the length of the string pointed to by
2160 /// the specified pointer, return 'len+1'. If we can't, return 0.
2161 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2162 // Look through noop bitcast instructions.
2163 V = V->stripPointerCasts();
2165 // If this is a PHI node, there are two cases: either we have already seen it
2167 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2168 if (!PHIs.insert(PN))
2169 return ~0ULL; // already in the set.
2171 // If it was new, see if all the input strings are the same length.
2172 uint64_t LenSoFar = ~0ULL;
2173 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2174 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2175 if (Len == 0) return 0; // Unknown length -> unknown.
2177 if (Len == ~0ULL) continue;
2179 if (Len != LenSoFar && LenSoFar != ~0ULL)
2180 return 0; // Disagree -> unknown.
2184 // Success, all agree.
2188 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2189 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2190 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2191 if (Len1 == 0) return 0;
2192 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2193 if (Len2 == 0) return 0;
2194 if (Len1 == ~0ULL) return Len2;
2195 if (Len2 == ~0ULL) return Len1;
2196 if (Len1 != Len2) return 0;
2200 // Otherwise, see if we can read the string.
2202 if (!getConstantStringInfo(V, StrData))
2205 return StrData.size()+1;
2208 /// GetStringLength - If we can compute the length of the string pointed to by
2209 /// the specified pointer, return 'len+1'. If we can't, return 0.
2210 uint64_t llvm::GetStringLength(Value *V) {
2211 if (!V->getType()->isPointerTy()) return 0;
2213 SmallPtrSet<PHINode*, 32> PHIs;
2214 uint64_t Len = GetStringLengthH(V, PHIs);
2215 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2216 // an empty string as a length.
2217 return Len == ~0ULL ? 1 : Len;
2221 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
2222 if (!V->getType()->isPointerTy())
2224 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2225 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2226 V = GEP->getPointerOperand();
2227 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2228 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2229 V = cast<Operator>(V)->getOperand(0);
2230 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2231 if (GA->mayBeOverridden())
2233 V = GA->getAliasee();
2235 // See if InstructionSimplify knows any relevant tricks.
2236 if (Instruction *I = dyn_cast<Instruction>(V))
2237 // TODO: Acquire a DominatorTree and AssumptionTracker and use them.
2238 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
2245 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2251 llvm::GetUnderlyingObjects(Value *V,
2252 SmallVectorImpl<Value *> &Objects,
2253 const DataLayout *TD,
2254 unsigned MaxLookup) {
2255 SmallPtrSet<Value *, 4> Visited;
2256 SmallVector<Value *, 4> Worklist;
2257 Worklist.push_back(V);
2259 Value *P = Worklist.pop_back_val();
2260 P = GetUnderlyingObject(P, TD, MaxLookup);
2262 if (!Visited.insert(P))
2265 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2266 Worklist.push_back(SI->getTrueValue());
2267 Worklist.push_back(SI->getFalseValue());
2271 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2272 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2273 Worklist.push_back(PN->getIncomingValue(i));
2277 Objects.push_back(P);
2278 } while (!Worklist.empty());
2281 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
2282 /// are lifetime markers.
2284 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2285 for (const User *U : V->users()) {
2286 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2287 if (!II) return false;
2289 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2290 II->getIntrinsicID() != Intrinsic::lifetime_end)
2296 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
2297 const DataLayout *TD) {
2298 const Operator *Inst = dyn_cast<Operator>(V);
2302 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2303 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2307 switch (Inst->getOpcode()) {
2310 case Instruction::UDiv:
2311 case Instruction::URem:
2312 // x / y is undefined if y == 0, but calculations like x / 3 are safe.
2313 return isKnownNonZero(Inst->getOperand(1), TD);
2314 case Instruction::SDiv:
2315 case Instruction::SRem: {
2316 Value *Op = Inst->getOperand(1);
2317 // x / y is undefined if y == 0
2318 if (!isKnownNonZero(Op, TD))
2320 // x / y might be undefined if y == -1
2321 unsigned BitWidth = getBitWidth(Op->getType(), TD);
2324 APInt KnownZero(BitWidth, 0);
2325 APInt KnownOne(BitWidth, 0);
2326 computeKnownBits(Op, KnownZero, KnownOne, TD);
2329 case Instruction::Load: {
2330 const LoadInst *LI = cast<LoadInst>(Inst);
2331 if (!LI->isUnordered() ||
2332 // Speculative load may create a race that did not exist in the source.
2333 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2335 return LI->getPointerOperand()->isDereferenceablePointer(TD);
2337 case Instruction::Call: {
2338 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2339 switch (II->getIntrinsicID()) {
2340 // These synthetic intrinsics have no side-effects and just mark
2341 // information about their operands.
2342 // FIXME: There are other no-op synthetic instructions that potentially
2343 // should be considered at least *safe* to speculate...
2344 case Intrinsic::dbg_declare:
2345 case Intrinsic::dbg_value:
2348 case Intrinsic::bswap:
2349 case Intrinsic::ctlz:
2350 case Intrinsic::ctpop:
2351 case Intrinsic::cttz:
2352 case Intrinsic::objectsize:
2353 case Intrinsic::sadd_with_overflow:
2354 case Intrinsic::smul_with_overflow:
2355 case Intrinsic::ssub_with_overflow:
2356 case Intrinsic::uadd_with_overflow:
2357 case Intrinsic::umul_with_overflow:
2358 case Intrinsic::usub_with_overflow:
2360 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2361 // errno like libm sqrt would.
2362 case Intrinsic::sqrt:
2363 case Intrinsic::fma:
2364 case Intrinsic::fmuladd:
2365 case Intrinsic::fabs:
2367 // TODO: some fp intrinsics are marked as having the same error handling
2368 // as libm. They're safe to speculate when they won't error.
2369 // TODO: are convert_{from,to}_fp16 safe?
2370 // TODO: can we list target-specific intrinsics here?
2374 return false; // The called function could have undefined behavior or
2375 // side-effects, even if marked readnone nounwind.
2377 case Instruction::VAArg:
2378 case Instruction::Alloca:
2379 case Instruction::Invoke:
2380 case Instruction::PHI:
2381 case Instruction::Store:
2382 case Instruction::Ret:
2383 case Instruction::Br:
2384 case Instruction::IndirectBr:
2385 case Instruction::Switch:
2386 case Instruction::Unreachable:
2387 case Instruction::Fence:
2388 case Instruction::LandingPad:
2389 case Instruction::AtomicRMW:
2390 case Instruction::AtomicCmpXchg:
2391 case Instruction::Resume:
2392 return false; // Misc instructions which have effects
2396 /// isKnownNonNull - Return true if we know that the specified value is never
2398 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2399 // Alloca never returns null, malloc might.
2400 if (isa<AllocaInst>(V)) return true;
2402 // A byval, inalloca, or nonnull argument is never null.
2403 if (const Argument *A = dyn_cast<Argument>(V))
2404 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2406 // Global values are not null unless extern weak.
2407 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2408 return !GV->hasExternalWeakLinkage();
2410 if (ImmutableCallSite CS = V)
2411 if (CS.isReturnNonNull())
2414 // operator new never returns null.
2415 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))