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 MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
197 unsigned Depth, const Query &Q);
199 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
200 unsigned Depth, AssumptionCache *AC,
201 const Instruction *CxtI, const DominatorTree *DT) {
202 return ::MaskedValueIsZero(V, Mask, DL, Depth,
203 Query(AC, safeCxtI(V, CxtI), DT));
206 static unsigned ComputeNumSignBits(Value *V, const DataLayout &DL,
207 unsigned Depth, const Query &Q);
209 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL,
210 unsigned Depth, AssumptionCache *AC,
211 const Instruction *CxtI,
212 const DominatorTree *DT) {
213 return ::ComputeNumSignBits(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
216 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
217 APInt &KnownZero, APInt &KnownOne,
218 APInt &KnownZero2, APInt &KnownOne2,
219 const DataLayout &DL, unsigned Depth,
222 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
223 // We know that the top bits of C-X are clear if X contains less bits
224 // than C (i.e. no wrap-around can happen). For example, 20-X is
225 // positive if we can prove that X is >= 0 and < 16.
226 if (!CLHS->getValue().isNegative()) {
227 unsigned BitWidth = KnownZero.getBitWidth();
228 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
229 // NLZ can't be BitWidth with no sign bit
230 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
231 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
233 // If all of the MaskV bits are known to be zero, then we know the
234 // output top bits are zero, because we now know that the output is
236 if ((KnownZero2 & MaskV) == MaskV) {
237 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
238 // Top bits known zero.
239 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
245 unsigned BitWidth = KnownZero.getBitWidth();
247 // If an initial sequence of bits in the result is not needed, the
248 // corresponding bits in the operands are not needed.
249 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
250 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, DL, Depth + 1, Q);
251 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
253 // Carry in a 1 for a subtract, rather than a 0.
254 APInt CarryIn(BitWidth, 0);
256 // Sum = LHS + ~RHS + 1
257 std::swap(KnownZero2, KnownOne2);
261 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
262 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
264 // Compute known bits of the carry.
265 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
266 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
268 // Compute set of known bits (where all three relevant bits are known).
269 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
270 APInt RHSKnown = KnownZero2 | KnownOne2;
271 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
272 APInt Known = LHSKnown & RHSKnown & CarryKnown;
274 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
275 "known bits of sum differ");
277 // Compute known bits of the result.
278 KnownZero = ~PossibleSumOne & Known;
279 KnownOne = PossibleSumOne & Known;
281 // Are we still trying to solve for the sign bit?
282 if (!Known.isNegative()) {
284 // Adding two non-negative numbers, or subtracting a negative number from
285 // a non-negative one, can't wrap into negative.
286 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
287 KnownZero |= APInt::getSignBit(BitWidth);
288 // Adding two negative numbers, or subtracting a non-negative number from
289 // a negative one, can't wrap into non-negative.
290 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
291 KnownOne |= APInt::getSignBit(BitWidth);
296 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
297 APInt &KnownZero, APInt &KnownOne,
298 APInt &KnownZero2, APInt &KnownOne2,
299 const DataLayout &DL, unsigned Depth,
301 unsigned BitWidth = KnownZero.getBitWidth();
302 computeKnownBits(Op1, KnownZero, KnownOne, DL, Depth + 1, Q);
303 computeKnownBits(Op0, KnownZero2, KnownOne2, DL, Depth + 1, Q);
305 bool isKnownNegative = false;
306 bool isKnownNonNegative = false;
307 // If the multiplication is known not to overflow, compute the sign bit.
310 // The product of a number with itself is non-negative.
311 isKnownNonNegative = true;
313 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
314 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
315 bool isKnownNegativeOp1 = KnownOne.isNegative();
316 bool isKnownNegativeOp0 = KnownOne2.isNegative();
317 // The product of two numbers with the same sign is non-negative.
318 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
319 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
320 // The product of a negative number and a non-negative number is either
322 if (!isKnownNonNegative)
323 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
324 isKnownNonZero(Op0, DL, Depth, Q)) ||
325 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
326 isKnownNonZero(Op1, DL, Depth, Q));
330 // If low bits are zero in either operand, output low known-0 bits.
331 // Also compute a conservative estimate for high known-0 bits.
332 // More trickiness is possible, but this is sufficient for the
333 // interesting case of alignment computation.
334 KnownOne.clearAllBits();
335 unsigned TrailZ = KnownZero.countTrailingOnes() +
336 KnownZero2.countTrailingOnes();
337 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
338 KnownZero2.countLeadingOnes(),
339 BitWidth) - BitWidth;
341 TrailZ = std::min(TrailZ, BitWidth);
342 LeadZ = std::min(LeadZ, BitWidth);
343 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
344 APInt::getHighBitsSet(BitWidth, LeadZ);
346 // Only make use of no-wrap flags if we failed to compute the sign bit
347 // directly. This matters if the multiplication always overflows, in
348 // which case we prefer to follow the result of the direct computation,
349 // though as the program is invoking undefined behaviour we can choose
350 // whatever we like here.
351 if (isKnownNonNegative && !KnownOne.isNegative())
352 KnownZero.setBit(BitWidth - 1);
353 else if (isKnownNegative && !KnownZero.isNegative())
354 KnownOne.setBit(BitWidth - 1);
357 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
359 unsigned BitWidth = KnownZero.getBitWidth();
360 unsigned NumRanges = Ranges.getNumOperands() / 2;
361 assert(NumRanges >= 1);
363 // Use the high end of the ranges to find leading zeros.
364 unsigned MinLeadingZeros = BitWidth;
365 for (unsigned i = 0; i < NumRanges; ++i) {
367 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
369 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
370 ConstantRange Range(Lower->getValue(), Upper->getValue());
371 if (Range.isWrappedSet())
372 MinLeadingZeros = 0; // -1 has no zeros
373 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
374 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
377 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
380 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
381 SmallVector<const Value *, 16> WorkSet(1, I);
382 SmallPtrSet<const Value *, 32> Visited;
383 SmallPtrSet<const Value *, 16> EphValues;
385 while (!WorkSet.empty()) {
386 const Value *V = WorkSet.pop_back_val();
387 if (!Visited.insert(V).second)
390 // If all uses of this value are ephemeral, then so is this value.
391 bool FoundNEUse = false;
392 for (const User *I : V->users())
393 if (!EphValues.count(I)) {
403 if (const User *U = dyn_cast<User>(V))
404 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
406 if (isSafeToSpeculativelyExecute(*J))
407 WorkSet.push_back(*J);
415 // Is this an intrinsic that cannot be speculated but also cannot trap?
416 static bool isAssumeLikeIntrinsic(const Instruction *I) {
417 if (const CallInst *CI = dyn_cast<CallInst>(I))
418 if (Function *F = CI->getCalledFunction())
419 switch (F->getIntrinsicID()) {
421 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
422 case Intrinsic::assume:
423 case Intrinsic::dbg_declare:
424 case Intrinsic::dbg_value:
425 case Intrinsic::invariant_start:
426 case Intrinsic::invariant_end:
427 case Intrinsic::lifetime_start:
428 case Intrinsic::lifetime_end:
429 case Intrinsic::objectsize:
430 case Intrinsic::ptr_annotation:
431 case Intrinsic::var_annotation:
438 static bool isValidAssumeForContext(Value *V, const Query &Q) {
439 Instruction *Inv = cast<Instruction>(V);
441 // There are two restrictions on the use of an assume:
442 // 1. The assume must dominate the context (or the control flow must
443 // reach the assume whenever it reaches the context).
444 // 2. The context must not be in the assume's set of ephemeral values
445 // (otherwise we will use the assume to prove that the condition
446 // feeding the assume is trivially true, thus causing the removal of
450 if (Q.DT->dominates(Inv, Q.CxtI)) {
452 } else if (Inv->getParent() == Q.CxtI->getParent()) {
453 // The context comes first, but they're both in the same block. Make sure
454 // there is nothing in between that might interrupt the control flow.
455 for (BasicBlock::const_iterator I =
456 std::next(BasicBlock::const_iterator(Q.CxtI)),
457 IE(Inv); I != IE; ++I)
458 if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I))
461 return !isEphemeralValueOf(Inv, Q.CxtI);
467 // When we don't have a DT, we do a limited search...
468 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
470 } else if (Inv->getParent() == Q.CxtI->getParent()) {
471 // Search forward from the assume until we reach the context (or the end
472 // of the block); the common case is that the assume will come first.
473 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
474 IE = Inv->getParent()->end(); I != IE; ++I)
478 // The context must come first...
479 for (BasicBlock::const_iterator I =
480 std::next(BasicBlock::const_iterator(Q.CxtI)),
481 IE(Inv); I != IE; ++I)
482 if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I))
485 return !isEphemeralValueOf(Inv, Q.CxtI);
491 bool llvm::isValidAssumeForContext(const Instruction *I,
492 const Instruction *CxtI,
493 const DominatorTree *DT) {
494 return ::isValidAssumeForContext(const_cast<Instruction *>(I),
495 Query(nullptr, CxtI, DT));
498 template<typename LHS, typename RHS>
499 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
500 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
501 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
502 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
505 template<typename LHS, typename RHS>
506 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
507 BinaryOp_match<RHS, LHS, Instruction::And>>
508 m_c_And(const LHS &L, const RHS &R) {
509 return m_CombineOr(m_And(L, R), m_And(R, L));
512 template<typename LHS, typename RHS>
513 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
514 BinaryOp_match<RHS, LHS, Instruction::Or>>
515 m_c_Or(const LHS &L, const RHS &R) {
516 return m_CombineOr(m_Or(L, R), m_Or(R, L));
519 template<typename LHS, typename RHS>
520 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
521 BinaryOp_match<RHS, LHS, Instruction::Xor>>
522 m_c_Xor(const LHS &L, const RHS &R) {
523 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
526 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
527 /// true (at the context instruction.) This is mostly a utility function for
528 /// the prototype dominating conditions reasoning below.
529 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
532 const DataLayout &DL,
533 unsigned Depth, const Query &Q) {
534 Value *LHS = Cmp->getOperand(0);
535 Value *RHS = Cmp->getOperand(1);
536 // TODO: We could potentially be more aggressive here. This would be worth
537 // evaluating. If we can, explore commoning this code with the assume
539 if (LHS != V && RHS != V)
542 const unsigned BitWidth = KnownZero.getBitWidth();
544 switch (Cmp->getPredicate()) {
546 // We know nothing from this condition
548 // TODO: implement unsigned bound from below (known one bits)
549 // TODO: common condition check implementations with assumes
550 // TODO: implement other patterns from assume (e.g. V & B == A)
551 case ICmpInst::ICMP_SGT:
553 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
554 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
555 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
556 // We know that the sign bit is zero.
557 KnownZero |= APInt::getSignBit(BitWidth);
561 case ICmpInst::ICMP_EQ:
563 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
565 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
567 computeKnownBits(LHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
569 llvm_unreachable("missing use?");
570 KnownZero |= KnownZeroTemp;
571 KnownOne |= KnownOneTemp;
574 case ICmpInst::ICMP_ULE:
576 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
577 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
578 // The known zero bits carry over
579 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
580 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
583 case ICmpInst::ICMP_ULT:
585 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
586 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
587 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
588 // power of 2, then one more).
589 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
590 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
592 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
598 /// Compute known bits in 'V' from conditions which are known to be true along
599 /// all paths leading to the context instruction. In particular, look for
600 /// cases where one branch of an interesting condition dominates the context
601 /// instruction. This does not do general dataflow.
602 /// NOTE: This code is EXPERIMENTAL and currently off by default.
603 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
605 const DataLayout &DL,
608 // Need both the dominator tree and the query location to do anything useful
609 if (!Q.DT || !Q.CxtI)
611 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
612 // The context instruction might be in a statically unreachable block. If
613 // so, asking dominator queries may yield suprising results. (e.g. the block
614 // may not have a dom tree node)
615 if (!Q.DT->isReachableFromEntry(Cxt->getParent()))
618 // Avoid useless work
619 if (auto VI = dyn_cast<Instruction>(V))
620 if (VI->getParent() == Cxt->getParent())
623 // Note: We currently implement two options. It's not clear which of these
624 // will survive long term, we need data for that.
625 // Option 1 - Try walking the dominator tree looking for conditions which
626 // might apply. This works well for local conditions (loop guards, etc..),
627 // but not as well for things far from the context instruction (presuming a
628 // low max blocks explored). If we can set an high enough limit, this would
630 // Option 2 - We restrict out search to those conditions which are uses of
631 // the value we're interested in. This is independent of dom structure,
632 // but is slightly less powerful without looking through lots of use chains.
633 // It does handle conditions far from the context instruction (e.g. early
634 // function exits on entry) really well though.
636 // Option 1 - Search the dom tree
637 unsigned NumBlocksExplored = 0;
638 BasicBlock *Current = Cxt->getParent();
640 // Stop searching if we've gone too far up the chain
641 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
645 if (!Q.DT->getNode(Current)->getIDom())
647 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
649 // found function entry
652 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
653 if (!BI || BI->isUnconditional())
655 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
659 // We're looking for conditions that are guaranteed to hold at the context
660 // instruction. Finding a condition where one path dominates the context
661 // isn't enough because both the true and false cases could merge before
662 // the context instruction we're actually interested in. Instead, we need
663 // to ensure that the taken *edge* dominates the context instruction. We
664 // know that the edge must be reachable since we started from a reachable
666 BasicBlock *BB0 = BI->getSuccessor(0);
667 BasicBlockEdge Edge(BI->getParent(), BB0);
668 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
671 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
675 // Option 2 - Search the other uses of V
676 unsigned NumUsesExplored = 0;
677 for (auto U : V->users()) {
678 // Avoid massive lists
679 if (NumUsesExplored >= DomConditionsMaxUses)
682 // Consider only compare instructions uniquely controlling a branch
683 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
687 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
690 for (auto *CmpU : Cmp->users()) {
691 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
692 if (!BI || BI->isUnconditional())
694 // We're looking for conditions that are guaranteed to hold at the
695 // context instruction. Finding a condition where one path dominates
696 // the context isn't enough because both the true and false cases could
697 // merge before the context instruction we're actually interested in.
698 // Instead, we need to ensure that the taken *edge* dominates the context
700 BasicBlock *BB0 = BI->getSuccessor(0);
701 BasicBlockEdge Edge(BI->getParent(), BB0);
702 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
705 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
711 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
712 APInt &KnownOne, const DataLayout &DL,
713 unsigned Depth, const Query &Q) {
714 // Use of assumptions is context-sensitive. If we don't have a context, we
716 if (!Q.AC || !Q.CxtI)
719 unsigned BitWidth = KnownZero.getBitWidth();
721 for (auto &AssumeVH : Q.AC->assumptions()) {
724 CallInst *I = cast<CallInst>(AssumeVH);
725 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
726 "Got assumption for the wrong function!");
727 if (Q.ExclInvs.count(I))
730 // Warning: This loop can end up being somewhat performance sensetive.
731 // We're running this loop for once for each value queried resulting in a
732 // runtime of ~O(#assumes * #values).
734 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
735 "must be an assume intrinsic");
737 Value *Arg = I->getArgOperand(0);
739 if (Arg == V && isValidAssumeForContext(I, Q)) {
740 assert(BitWidth == 1 && "assume operand is not i1?");
741 KnownZero.clearAllBits();
742 KnownOne.setAllBits();
746 // The remaining tests are all recursive, so bail out if we hit the limit.
747 if (Depth == MaxDepth)
751 auto m_V = m_CombineOr(m_Specific(V),
752 m_CombineOr(m_PtrToInt(m_Specific(V)),
753 m_BitCast(m_Specific(V))));
755 CmpInst::Predicate Pred;
758 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
759 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
760 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
761 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
762 KnownZero |= RHSKnownZero;
763 KnownOne |= RHSKnownOne;
765 } else if (match(Arg,
766 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
767 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
768 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
769 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
770 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
771 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
773 // For those bits in the mask that are known to be one, we can propagate
774 // known bits from the RHS to V.
775 KnownZero |= RHSKnownZero & MaskKnownOne;
776 KnownOne |= RHSKnownOne & MaskKnownOne;
777 // assume(~(v & b) = a)
778 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
780 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
781 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
782 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
783 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
784 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
786 // For those bits in the mask that are known to be one, we can propagate
787 // inverted known bits from the RHS to V.
788 KnownZero |= RHSKnownOne & MaskKnownOne;
789 KnownOne |= RHSKnownZero & MaskKnownOne;
791 } else if (match(Arg,
792 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
793 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
794 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
795 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
796 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
797 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
799 // For those bits in B that are known to be zero, we can propagate known
800 // bits from the RHS to V.
801 KnownZero |= RHSKnownZero & BKnownZero;
802 KnownOne |= RHSKnownOne & BKnownZero;
803 // assume(~(v | b) = a)
804 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
806 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
807 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
808 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
809 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
810 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
812 // For those bits in B that are known to be zero, we can propagate
813 // inverted known bits from the RHS to V.
814 KnownZero |= RHSKnownOne & BKnownZero;
815 KnownOne |= RHSKnownZero & BKnownZero;
817 } else if (match(Arg,
818 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
819 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
820 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
821 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
822 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
823 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
825 // For those bits in B that are known to be zero, we can propagate known
826 // bits from the RHS to V. For those bits in B that are known to be one,
827 // we can propagate inverted known bits from the RHS to V.
828 KnownZero |= RHSKnownZero & BKnownZero;
829 KnownOne |= RHSKnownOne & BKnownZero;
830 KnownZero |= RHSKnownOne & BKnownOne;
831 KnownOne |= RHSKnownZero & BKnownOne;
832 // assume(~(v ^ b) = a)
833 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
835 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
836 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
837 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
838 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
839 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
841 // For those bits in B that are known to be zero, we can propagate
842 // inverted known bits from the RHS to V. For those bits in B that are
843 // known to be one, we can propagate known bits from the RHS to V.
844 KnownZero |= RHSKnownOne & BKnownZero;
845 KnownOne |= RHSKnownZero & BKnownZero;
846 KnownZero |= RHSKnownZero & BKnownOne;
847 KnownOne |= RHSKnownOne & BKnownOne;
848 // assume(v << c = a)
849 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
851 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
852 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
853 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
854 // For those bits in RHS that are known, we can propagate them to known
855 // bits in V shifted to the right by C.
856 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
857 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
858 // assume(~(v << c) = a)
859 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
861 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
862 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
863 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
864 // For those bits in RHS that are known, we can propagate them inverted
865 // to known bits in V shifted to the right by C.
866 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
867 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
868 // assume(v >> c = a)
869 } else if (match(Arg,
870 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
871 m_AShr(m_V, m_ConstantInt(C))),
873 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
874 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
875 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
876 // For those bits in RHS that are known, we can propagate them to known
877 // bits in V shifted to the right by C.
878 KnownZero |= RHSKnownZero << C->getZExtValue();
879 KnownOne |= RHSKnownOne << C->getZExtValue();
880 // assume(~(v >> c) = a)
881 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
882 m_LShr(m_V, m_ConstantInt(C)),
883 m_AShr(m_V, m_ConstantInt(C)))),
885 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
886 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
887 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
888 // For those bits in RHS that are known, we can propagate them inverted
889 // to known bits in V shifted to the right by C.
890 KnownZero |= RHSKnownOne << C->getZExtValue();
891 KnownOne |= RHSKnownZero << C->getZExtValue();
892 // assume(v >=_s c) where c is non-negative
893 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
894 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
895 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
896 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
898 if (RHSKnownZero.isNegative()) {
899 // We know that the sign bit is zero.
900 KnownZero |= APInt::getSignBit(BitWidth);
902 // assume(v >_s c) where c is at least -1.
903 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
904 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
905 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
906 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
908 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
909 // We know that the sign bit is zero.
910 KnownZero |= APInt::getSignBit(BitWidth);
912 // assume(v <=_s c) where c is negative
913 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
914 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
915 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
916 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
918 if (RHSKnownOne.isNegative()) {
919 // We know that the sign bit is one.
920 KnownOne |= APInt::getSignBit(BitWidth);
922 // assume(v <_s c) where c is non-positive
923 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
924 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
925 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
926 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
928 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
929 // We know that the sign bit is one.
930 KnownOne |= APInt::getSignBit(BitWidth);
933 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
934 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
935 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
936 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
938 // Whatever high bits in c are zero are known to be zero.
940 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
942 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
943 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
944 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
945 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
947 // Whatever high bits in c are zero are known to be zero (if c is a power
948 // of 2, then one more).
949 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
951 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
954 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
959 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
960 APInt &KnownOne, const DataLayout &DL,
961 unsigned Depth, const Query &Q) {
962 unsigned BitWidth = KnownZero.getBitWidth();
964 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
965 switch (I->getOpcode()) {
967 case Instruction::Load:
968 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
969 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
971 case Instruction::And: {
972 // If either the LHS or the RHS are Zero, the result is zero.
973 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
974 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
976 // Output known-1 bits are only known if set in both the LHS & RHS.
977 KnownOne &= KnownOne2;
978 // Output known-0 are known to be clear if zero in either the LHS | RHS.
979 KnownZero |= KnownZero2;
982 case Instruction::Or: {
983 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
984 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
986 // Output known-0 bits are only known if clear in both the LHS & RHS.
987 KnownZero &= KnownZero2;
988 // Output known-1 are known to be set if set in either the LHS | RHS.
989 KnownOne |= KnownOne2;
992 case Instruction::Xor: {
993 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
994 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
996 // Output known-0 bits are known if clear or set in both the LHS & RHS.
997 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
998 // Output known-1 are known to be set if set in only one of the LHS, RHS.
999 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1000 KnownZero = KnownZeroOut;
1003 case Instruction::Mul: {
1004 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1005 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1006 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1009 case Instruction::UDiv: {
1010 // For the purposes of computing leading zeros we can conservatively
1011 // treat a udiv as a logical right shift by the power of 2 known to
1012 // be less than the denominator.
1013 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1014 unsigned LeadZ = KnownZero2.countLeadingOnes();
1016 KnownOne2.clearAllBits();
1017 KnownZero2.clearAllBits();
1018 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1019 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1020 if (RHSUnknownLeadingOnes != BitWidth)
1021 LeadZ = std::min(BitWidth,
1022 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1024 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1027 case Instruction::Select:
1028 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1029 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1031 // Only known if known in both the LHS and RHS.
1032 KnownOne &= KnownOne2;
1033 KnownZero &= KnownZero2;
1035 case Instruction::FPTrunc:
1036 case Instruction::FPExt:
1037 case Instruction::FPToUI:
1038 case Instruction::FPToSI:
1039 case Instruction::SIToFP:
1040 case Instruction::UIToFP:
1041 break; // Can't work with floating point.
1042 case Instruction::PtrToInt:
1043 case Instruction::IntToPtr:
1044 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1045 // FALL THROUGH and handle them the same as zext/trunc.
1046 case Instruction::ZExt:
1047 case Instruction::Trunc: {
1048 Type *SrcTy = I->getOperand(0)->getType();
1050 unsigned SrcBitWidth;
1051 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1052 // which fall through here.
1053 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1055 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1056 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1057 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1058 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1059 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1060 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1061 // Any top bits are known to be zero.
1062 if (BitWidth > SrcBitWidth)
1063 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1066 case Instruction::BitCast: {
1067 Type *SrcTy = I->getOperand(0)->getType();
1068 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
1069 // TODO: For now, not handling conversions like:
1070 // (bitcast i64 %x to <2 x i32>)
1071 !I->getType()->isVectorTy()) {
1072 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1077 case Instruction::SExt: {
1078 // Compute the bits in the result that are not present in the input.
1079 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1081 KnownZero = KnownZero.trunc(SrcBitWidth);
1082 KnownOne = KnownOne.trunc(SrcBitWidth);
1083 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1084 KnownZero = KnownZero.zext(BitWidth);
1085 KnownOne = KnownOne.zext(BitWidth);
1087 // If the sign bit of the input is known set or clear, then we know the
1088 // top bits of the result.
1089 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1090 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1091 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1092 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1095 case Instruction::Shl:
1096 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1097 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1098 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1099 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1100 KnownZero <<= ShiftAmt;
1101 KnownOne <<= ShiftAmt;
1102 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1105 case Instruction::LShr:
1106 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1107 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1108 // Compute the new bits that are at the top now.
1109 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1111 // Unsigned shift right.
1112 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1113 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1114 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1115 // high bits known zero.
1116 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1119 case Instruction::AShr:
1120 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1121 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1122 // Compute the new bits that are at the top now.
1123 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1125 // Signed shift right.
1126 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1127 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1128 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1130 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1131 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1132 KnownZero |= HighBits;
1133 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1134 KnownOne |= HighBits;
1137 case Instruction::Sub: {
1138 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1139 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1140 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1144 case Instruction::Add: {
1145 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1146 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1147 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1151 case Instruction::SRem:
1152 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1153 APInt RA = Rem->getValue().abs();
1154 if (RA.isPowerOf2()) {
1155 APInt LowBits = RA - 1;
1156 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1159 // The low bits of the first operand are unchanged by the srem.
1160 KnownZero = KnownZero2 & LowBits;
1161 KnownOne = KnownOne2 & LowBits;
1163 // If the first operand is non-negative or has all low bits zero, then
1164 // the upper bits are all zero.
1165 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1166 KnownZero |= ~LowBits;
1168 // If the first operand is negative and not all low bits are zero, then
1169 // the upper bits are all one.
1170 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1171 KnownOne |= ~LowBits;
1173 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1177 // The sign bit is the LHS's sign bit, except when the result of the
1178 // remainder is zero.
1179 if (KnownZero.isNonNegative()) {
1180 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1181 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1183 // If it's known zero, our sign bit is also zero.
1184 if (LHSKnownZero.isNegative())
1185 KnownZero.setBit(BitWidth - 1);
1189 case Instruction::URem: {
1190 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1191 APInt RA = Rem->getValue();
1192 if (RA.isPowerOf2()) {
1193 APInt LowBits = (RA - 1);
1194 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1196 KnownZero |= ~LowBits;
1197 KnownOne &= LowBits;
1202 // Since the result is less than or equal to either operand, any leading
1203 // zero bits in either operand must also exist in the result.
1204 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1205 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1207 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1208 KnownZero2.countLeadingOnes());
1209 KnownOne.clearAllBits();
1210 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1214 case Instruction::Alloca: {
1215 AllocaInst *AI = cast<AllocaInst>(I);
1216 unsigned Align = AI->getAlignment();
1218 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1221 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1224 case Instruction::GetElementPtr: {
1225 // Analyze all of the subscripts of this getelementptr instruction
1226 // to determine if we can prove known low zero bits.
1227 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1228 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1230 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1232 gep_type_iterator GTI = gep_type_begin(I);
1233 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1234 Value *Index = I->getOperand(i);
1235 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1236 // Handle struct member offset arithmetic.
1238 // Handle case when index is vector zeroinitializer
1239 Constant *CIndex = cast<Constant>(Index);
1240 if (CIndex->isZeroValue())
1243 if (CIndex->getType()->isVectorTy())
1244 Index = CIndex->getSplatValue();
1246 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1247 const StructLayout *SL = DL.getStructLayout(STy);
1248 uint64_t Offset = SL->getElementOffset(Idx);
1249 TrailZ = std::min<unsigned>(TrailZ,
1250 countTrailingZeros(Offset));
1252 // Handle array index arithmetic.
1253 Type *IndexedTy = GTI.getIndexedType();
1254 if (!IndexedTy->isSized()) {
1258 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1259 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1260 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1261 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1263 TrailZ = std::min(TrailZ,
1264 unsigned(countTrailingZeros(TypeSize) +
1265 LocalKnownZero.countTrailingOnes()));
1269 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1272 case Instruction::PHI: {
1273 PHINode *P = cast<PHINode>(I);
1274 // Handle the case of a simple two-predecessor recurrence PHI.
1275 // There's a lot more that could theoretically be done here, but
1276 // this is sufficient to catch some interesting cases.
1277 if (P->getNumIncomingValues() == 2) {
1278 for (unsigned i = 0; i != 2; ++i) {
1279 Value *L = P->getIncomingValue(i);
1280 Value *R = P->getIncomingValue(!i);
1281 Operator *LU = dyn_cast<Operator>(L);
1284 unsigned Opcode = LU->getOpcode();
1285 // Check for operations that have the property that if
1286 // both their operands have low zero bits, the result
1287 // will have low zero bits.
1288 if (Opcode == Instruction::Add ||
1289 Opcode == Instruction::Sub ||
1290 Opcode == Instruction::And ||
1291 Opcode == Instruction::Or ||
1292 Opcode == Instruction::Mul) {
1293 Value *LL = LU->getOperand(0);
1294 Value *LR = LU->getOperand(1);
1295 // Find a recurrence.
1302 // Ok, we have a PHI of the form L op= R. Check for low
1304 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1306 // We need to take the minimum number of known bits
1307 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1308 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1310 KnownZero = APInt::getLowBitsSet(BitWidth,
1311 std::min(KnownZero2.countTrailingOnes(),
1312 KnownZero3.countTrailingOnes()));
1318 // Unreachable blocks may have zero-operand PHI nodes.
1319 if (P->getNumIncomingValues() == 0)
1322 // Otherwise take the unions of the known bit sets of the operands,
1323 // taking conservative care to avoid excessive recursion.
1324 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1325 // Skip if every incoming value references to ourself.
1326 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1329 KnownZero = APInt::getAllOnesValue(BitWidth);
1330 KnownOne = APInt::getAllOnesValue(BitWidth);
1331 for (Value *IncValue : P->incoming_values()) {
1332 // Skip direct self references.
1333 if (IncValue == P) continue;
1335 KnownZero2 = APInt(BitWidth, 0);
1336 KnownOne2 = APInt(BitWidth, 0);
1337 // Recurse, but cap the recursion to one level, because we don't
1338 // want to waste time spinning around in loops.
1339 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1341 KnownZero &= KnownZero2;
1342 KnownOne &= KnownOne2;
1343 // If all bits have been ruled out, there's no need to check
1345 if (!KnownZero && !KnownOne)
1351 case Instruction::Call:
1352 case Instruction::Invoke:
1353 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1354 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1355 // If a range metadata is attached to this IntrinsicInst, intersect the
1356 // explicit range specified by the metadata and the implicit range of
1358 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1359 switch (II->getIntrinsicID()) {
1361 case Intrinsic::bswap:
1362 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1364 KnownZero |= KnownZero2.byteSwap();
1365 KnownOne |= KnownOne2.byteSwap();
1367 case Intrinsic::ctlz:
1368 case Intrinsic::cttz: {
1369 unsigned LowBits = Log2_32(BitWidth)+1;
1370 // If this call is undefined for 0, the result will be less than 2^n.
1371 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1373 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1376 case Intrinsic::ctpop: {
1377 unsigned LowBits = Log2_32(BitWidth)+1;
1378 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1381 case Intrinsic::x86_sse42_crc32_64_64:
1382 KnownZero |= APInt::getHighBitsSet(64, 32);
1387 case Instruction::ExtractValue:
1388 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1389 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1390 if (EVI->getNumIndices() != 1) break;
1391 if (EVI->getIndices()[0] == 0) {
1392 switch (II->getIntrinsicID()) {
1394 case Intrinsic::uadd_with_overflow:
1395 case Intrinsic::sadd_with_overflow:
1396 computeKnownBitsAddSub(true, II->getArgOperand(0),
1397 II->getArgOperand(1), false, KnownZero,
1398 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1400 case Intrinsic::usub_with_overflow:
1401 case Intrinsic::ssub_with_overflow:
1402 computeKnownBitsAddSub(false, II->getArgOperand(0),
1403 II->getArgOperand(1), false, KnownZero,
1404 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1406 case Intrinsic::umul_with_overflow:
1407 case Intrinsic::smul_with_overflow:
1408 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1409 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1418 static unsigned getAlignment(Value *V, const DataLayout &DL) {
1420 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1421 Align = GO->getAlignment();
1423 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1424 Type *ObjectType = GVar->getType()->getElementType();
1425 if (ObjectType->isSized()) {
1426 // If the object is defined in the current Module, we'll be giving
1427 // it the preferred alignment. Otherwise, we have to assume that it
1428 // may only have the minimum ABI alignment.
1429 if (GVar->isStrongDefinitionForLinker())
1430 Align = DL.getPreferredAlignment(GVar);
1432 Align = DL.getABITypeAlignment(ObjectType);
1436 } else if (Argument *A = dyn_cast<Argument>(V)) {
1437 Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1439 if (!Align && A->hasStructRetAttr()) {
1440 // An sret parameter has at least the ABI alignment of the return type.
1441 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1442 if (EltTy->isSized())
1443 Align = DL.getABITypeAlignment(EltTy);
1449 /// Determine which bits of V are known to be either zero or one and return
1450 /// them in the KnownZero/KnownOne bit sets.
1452 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1453 /// we cannot optimize based on the assumption that it is zero without changing
1454 /// it to be an explicit zero. If we don't change it to zero, other code could
1455 /// optimized based on the contradictory assumption that it is non-zero.
1456 /// Because instcombine aggressively folds operations with undef args anyway,
1457 /// this won't lose us code quality.
1459 /// This function is defined on values with integer type, values with pointer
1460 /// type, and vectors of integers. In the case
1461 /// where V is a vector, known zero, and known one values are the
1462 /// same width as the vector element, and the bit is set only if it is true
1463 /// for all of the elements in the vector.
1464 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1465 const DataLayout &DL, unsigned Depth, const Query &Q) {
1466 assert(V && "No Value?");
1467 assert(Depth <= MaxDepth && "Limit Search Depth");
1468 unsigned BitWidth = KnownZero.getBitWidth();
1470 assert((V->getType()->isIntOrIntVectorTy() ||
1471 V->getType()->getScalarType()->isPointerTy()) &&
1472 "Not integer or pointer type!");
1473 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1474 (!V->getType()->isIntOrIntVectorTy() ||
1475 V->getType()->getScalarSizeInBits() == BitWidth) &&
1476 KnownZero.getBitWidth() == BitWidth &&
1477 KnownOne.getBitWidth() == BitWidth &&
1478 "V, KnownOne and KnownZero should have same BitWidth");
1480 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1481 // We know all of the bits for a constant!
1482 KnownOne = CI->getValue();
1483 KnownZero = ~KnownOne;
1486 // Null and aggregate-zero are all-zeros.
1487 if (isa<ConstantPointerNull>(V) ||
1488 isa<ConstantAggregateZero>(V)) {
1489 KnownOne.clearAllBits();
1490 KnownZero = APInt::getAllOnesValue(BitWidth);
1493 // Handle a constant vector by taking the intersection of the known bits of
1494 // each element. There is no real need to handle ConstantVector here, because
1495 // we don't handle undef in any particularly useful way.
1496 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1497 // We know that CDS must be a vector of integers. Take the intersection of
1499 KnownZero.setAllBits(); KnownOne.setAllBits();
1500 APInt Elt(KnownZero.getBitWidth(), 0);
1501 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1502 Elt = CDS->getElementAsInteger(i);
1509 // Start out not knowing anything.
1510 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1512 // Limit search depth.
1513 // All recursive calls that increase depth must come after this.
1514 if (Depth == MaxDepth)
1517 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1518 // the bits of its aliasee.
1519 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1520 if (!GA->mayBeOverridden())
1521 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1525 if (Operator *I = dyn_cast<Operator>(V))
1526 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1528 // Aligned pointers have trailing zeros - refine KnownZero set
1529 if (V->getType()->isPointerTy()) {
1530 unsigned Align = getAlignment(V, DL);
1532 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1535 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1536 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1537 // computeKnownBitsFromOperator.
1539 // Check whether a nearby assume intrinsic can determine some known bits.
1540 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1542 // Check whether there's a dominating condition which implies something about
1543 // this value at the given context.
1544 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1545 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1548 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1551 /// Determine whether the sign bit is known to be zero or one.
1552 /// Convenience wrapper around computeKnownBits.
1553 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1554 const DataLayout &DL, unsigned Depth, const Query &Q) {
1555 unsigned BitWidth = getBitWidth(V->getType(), DL);
1561 APInt ZeroBits(BitWidth, 0);
1562 APInt OneBits(BitWidth, 0);
1563 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1564 KnownOne = OneBits[BitWidth - 1];
1565 KnownZero = ZeroBits[BitWidth - 1];
1568 /// Return true if the given value is known to have exactly one
1569 /// bit set when defined. For vectors return true if every element is known to
1570 /// be a power of two when defined. Supports values with integer or pointer
1571 /// types and vectors of integers.
1572 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1573 const Query &Q, const DataLayout &DL) {
1574 if (Constant *C = dyn_cast<Constant>(V)) {
1575 if (C->isNullValue())
1577 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1578 return CI->getValue().isPowerOf2();
1579 // TODO: Handle vector constants.
1582 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1583 // it is shifted off the end then the result is undefined.
1584 if (match(V, m_Shl(m_One(), m_Value())))
1587 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1588 // bottom. If it is shifted off the bottom then the result is undefined.
1589 if (match(V, m_LShr(m_SignBit(), m_Value())))
1592 // The remaining tests are all recursive, so bail out if we hit the limit.
1593 if (Depth++ == MaxDepth)
1596 Value *X = nullptr, *Y = nullptr;
1597 // A shift of a power of two is a power of two or zero.
1598 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1599 match(V, m_Shr(m_Value(X), m_Value()))))
1600 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1602 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1603 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1605 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1606 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1607 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1609 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1610 // A power of two and'd with anything is a power of two or zero.
1611 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1612 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1614 // X & (-X) is always a power of two or zero.
1615 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1620 // Adding a power-of-two or zero to the same power-of-two or zero yields
1621 // either the original power-of-two, a larger power-of-two or zero.
1622 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1623 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1624 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1625 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1626 match(X, m_And(m_Value(), m_Specific(Y))))
1627 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1629 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1630 match(Y, m_And(m_Value(), m_Specific(X))))
1631 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1634 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1635 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1636 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1638 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1639 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1640 // If i8 V is a power of two or zero:
1641 // ZeroBits: 1 1 1 0 1 1 1 1
1642 // ~ZeroBits: 0 0 0 1 0 0 0 0
1643 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1644 // If OrZero isn't set, we cannot give back a zero result.
1645 // Make sure either the LHS or RHS has a bit set.
1646 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1651 // An exact divide or right shift can only shift off zero bits, so the result
1652 // is a power of two only if the first operand is a power of two and not
1653 // copying a sign bit (sdiv int_min, 2).
1654 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1655 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1656 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1663 /// \brief Test whether a GEP's result is known to be non-null.
1665 /// Uses properties inherent in a GEP to try to determine whether it is known
1668 /// Currently this routine does not support vector GEPs.
1669 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1670 unsigned Depth, const Query &Q) {
1671 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1674 // FIXME: Support vector-GEPs.
1675 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1677 // If the base pointer is non-null, we cannot walk to a null address with an
1678 // inbounds GEP in address space zero.
1679 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1682 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1683 // If so, then the GEP cannot produce a null pointer, as doing so would
1684 // inherently violate the inbounds contract within address space zero.
1685 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1686 GTI != GTE; ++GTI) {
1687 // Struct types are easy -- they must always be indexed by a constant.
1688 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1689 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1690 unsigned ElementIdx = OpC->getZExtValue();
1691 const StructLayout *SL = DL.getStructLayout(STy);
1692 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1693 if (ElementOffset > 0)
1698 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1699 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1702 // Fast path the constant operand case both for efficiency and so we don't
1703 // increment Depth when just zipping down an all-constant GEP.
1704 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1710 // We post-increment Depth here because while isKnownNonZero increments it
1711 // as well, when we pop back up that increment won't persist. We don't want
1712 // to recurse 10k times just because we have 10k GEP operands. We don't
1713 // bail completely out because we want to handle constant GEPs regardless
1715 if (Depth++ >= MaxDepth)
1718 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1725 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1726 /// ensure that the value it's attached to is never Value? 'RangeType' is
1727 /// is the type of the value described by the range.
1728 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1729 const APInt& Value) {
1730 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1731 assert(NumRanges >= 1);
1732 for (unsigned i = 0; i < NumRanges; ++i) {
1733 ConstantInt *Lower =
1734 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1735 ConstantInt *Upper =
1736 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1737 ConstantRange Range(Lower->getValue(), Upper->getValue());
1738 if (Range.contains(Value))
1744 /// Return true if the given value is known to be non-zero when defined.
1745 /// For vectors return true if every element is known to be non-zero when
1746 /// defined. Supports values with integer or pointer type and vectors of
1748 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1750 if (Constant *C = dyn_cast<Constant>(V)) {
1751 if (C->isNullValue())
1753 if (isa<ConstantInt>(C))
1754 // Must be non-zero due to null test above.
1756 // TODO: Handle vectors
1760 if (Instruction* I = dyn_cast<Instruction>(V)) {
1761 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1762 // If the possible ranges don't contain zero, then the value is
1763 // definitely non-zero.
1764 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1765 const APInt ZeroValue(Ty->getBitWidth(), 0);
1766 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1772 // The remaining tests are all recursive, so bail out if we hit the limit.
1773 if (Depth++ >= MaxDepth)
1776 // Check for pointer simplifications.
1777 if (V->getType()->isPointerTy()) {
1778 if (isKnownNonNull(V))
1780 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1781 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1785 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1787 // X | Y != 0 if X != 0 or Y != 0.
1788 Value *X = nullptr, *Y = nullptr;
1789 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1790 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1792 // ext X != 0 if X != 0.
1793 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1794 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1796 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1797 // if the lowest bit is shifted off the end.
1798 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1799 // shl nuw can't remove any non-zero bits.
1800 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1801 if (BO->hasNoUnsignedWrap())
1802 return isKnownNonZero(X, DL, Depth, Q);
1804 APInt KnownZero(BitWidth, 0);
1805 APInt KnownOne(BitWidth, 0);
1806 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1810 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1811 // defined if the sign bit is shifted off the end.
1812 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1813 // shr exact can only shift out zero bits.
1814 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1816 return isKnownNonZero(X, DL, Depth, Q);
1818 bool XKnownNonNegative, XKnownNegative;
1819 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1823 // If the shifter operand is a constant, and all of the bits shifted
1824 // out are known to be zero, and X is known non-zero then at least one
1825 // non-zero bit must remain.
1826 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1827 APInt KnownZero(BitWidth, 0);
1828 APInt KnownOne(BitWidth, 0);
1829 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1831 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1832 // Is there a known one in the portion not shifted out?
1833 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1835 // Are all the bits to be shifted out known zero?
1836 if (KnownZero.countTrailingOnes() >= ShiftVal)
1837 return isKnownNonZero(X, DL, Depth, Q);
1840 // div exact can only produce a zero if the dividend is zero.
1841 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1842 return isKnownNonZero(X, DL, Depth, Q);
1845 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1846 bool XKnownNonNegative, XKnownNegative;
1847 bool YKnownNonNegative, YKnownNegative;
1848 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1849 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1851 // If X and Y are both non-negative (as signed values) then their sum is not
1852 // zero unless both X and Y are zero.
1853 if (XKnownNonNegative && YKnownNonNegative)
1854 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1857 // If X and Y are both negative (as signed values) then their sum is not
1858 // zero unless both X and Y equal INT_MIN.
1859 if (BitWidth && XKnownNegative && YKnownNegative) {
1860 APInt KnownZero(BitWidth, 0);
1861 APInt KnownOne(BitWidth, 0);
1862 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1863 // The sign bit of X is set. If some other bit is set then X is not equal
1865 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1866 if ((KnownOne & Mask) != 0)
1868 // The sign bit of Y is set. If some other bit is set then Y is not equal
1870 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1871 if ((KnownOne & Mask) != 0)
1875 // The sum of a non-negative number and a power of two is not zero.
1876 if (XKnownNonNegative &&
1877 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1879 if (YKnownNonNegative &&
1880 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1884 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1885 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1886 // If X and Y are non-zero then so is X * Y as long as the multiplication
1887 // does not overflow.
1888 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1889 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
1892 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1893 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1894 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
1895 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
1899 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1900 // Try and detect a recurrence that monotonically increases from a
1901 // starting value, as these are common as induction variables.
1902 if (PN->getNumIncomingValues() == 2) {
1903 Value *Start = PN->getIncomingValue(0);
1904 Value *Induction = PN->getIncomingValue(1);
1905 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
1906 std::swap(Start, Induction);
1907 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
1908 if (!C->isZero() && !C->isNegative()) {
1910 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
1911 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
1919 if (!BitWidth) return false;
1920 APInt KnownZero(BitWidth, 0);
1921 APInt KnownOne(BitWidth, 0);
1922 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1923 return KnownOne != 0;
1926 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1927 /// simplify operations downstream. Mask is known to be zero for bits that V
1930 /// This function is defined on values with integer type, values with pointer
1931 /// type, and vectors of integers. In the case
1932 /// where V is a vector, the mask, known zero, and known one values are the
1933 /// same width as the vector element, and the bit is set only if it is true
1934 /// for all of the elements in the vector.
1935 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
1936 unsigned Depth, const Query &Q) {
1937 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1938 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1939 return (KnownZero & Mask) == Mask;
1944 /// Return the number of times the sign bit of the register is replicated into
1945 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1946 /// (itself), but other cases can give us information. For example, immediately
1947 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1948 /// other, so we return 3.
1950 /// 'Op' must have a scalar integer type.
1952 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
1954 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
1956 unsigned FirstAnswer = 1;
1958 // Note that ConstantInt is handled by the general computeKnownBits case
1962 return 1; // Limit search depth.
1964 Operator *U = dyn_cast<Operator>(V);
1965 switch (Operator::getOpcode(V)) {
1967 case Instruction::SExt:
1968 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1969 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
1971 case Instruction::SDiv: {
1972 const APInt *Denominator;
1973 // sdiv X, C -> adds log(C) sign bits.
1974 if (match(U->getOperand(1), m_APInt(Denominator))) {
1976 // Ignore non-positive denominator.
1977 if (!Denominator->isStrictlyPositive())
1980 // Calculate the incoming numerator bits.
1981 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1983 // Add floor(log(C)) bits to the numerator bits.
1984 return std::min(TyBits, NumBits + Denominator->logBase2());
1989 case Instruction::SRem: {
1990 const APInt *Denominator;
1991 // srem X, C -> we know that the result is within [-C+1,C) when C is a
1992 // positive constant. This let us put a lower bound on the number of sign
1994 if (match(U->getOperand(1), m_APInt(Denominator))) {
1996 // Ignore non-positive denominator.
1997 if (!Denominator->isStrictlyPositive())
2000 // Calculate the incoming numerator bits. SRem by a positive constant
2001 // can't lower the number of sign bits.
2003 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2005 // Calculate the leading sign bit constraints by examining the
2006 // denominator. Given that the denominator is positive, there are two
2009 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2010 // (1 << ceilLogBase2(C)).
2012 // 2. the numerator is negative. Then the result range is (-C,0] and
2013 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2015 // Thus a lower bound on the number of sign bits is `TyBits -
2016 // ceilLogBase2(C)`.
2018 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2019 return std::max(NumrBits, ResBits);
2024 case Instruction::AShr: {
2025 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2026 // ashr X, C -> adds C sign bits. Vectors too.
2028 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2029 Tmp += ShAmt->getZExtValue();
2030 if (Tmp > TyBits) Tmp = TyBits;
2034 case Instruction::Shl: {
2036 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2037 // shl destroys sign bits.
2038 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2039 Tmp2 = ShAmt->getZExtValue();
2040 if (Tmp2 >= TyBits || // Bad shift.
2041 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2046 case Instruction::And:
2047 case Instruction::Or:
2048 case Instruction::Xor: // NOT is handled here.
2049 // Logical binary ops preserve the number of sign bits at the worst.
2050 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2052 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2053 FirstAnswer = std::min(Tmp, Tmp2);
2054 // We computed what we know about the sign bits as our first
2055 // answer. Now proceed to the generic code that uses
2056 // computeKnownBits, and pick whichever answer is better.
2060 case Instruction::Select:
2061 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2062 if (Tmp == 1) return 1; // Early out.
2063 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2064 return std::min(Tmp, Tmp2);
2066 case Instruction::Add:
2067 // Add can have at most one carry bit. Thus we know that the output
2068 // is, at worst, one more bit than the inputs.
2069 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2070 if (Tmp == 1) return 1; // Early out.
2072 // Special case decrementing a value (ADD X, -1):
2073 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2074 if (CRHS->isAllOnesValue()) {
2075 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2076 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2079 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2081 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2084 // If we are subtracting one from a positive number, there is no carry
2085 // out of the result.
2086 if (KnownZero.isNegative())
2090 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2091 if (Tmp2 == 1) return 1;
2092 return std::min(Tmp, Tmp2)-1;
2094 case Instruction::Sub:
2095 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2096 if (Tmp2 == 1) return 1;
2099 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2100 if (CLHS->isNullValue()) {
2101 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2102 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2104 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2106 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2109 // If the input is known to be positive (the sign bit is known clear),
2110 // the output of the NEG has the same number of sign bits as the input.
2111 if (KnownZero.isNegative())
2114 // Otherwise, we treat this like a SUB.
2117 // Sub can have at most one carry bit. Thus we know that the output
2118 // is, at worst, one more bit than the inputs.
2119 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2120 if (Tmp == 1) return 1; // Early out.
2121 return std::min(Tmp, Tmp2)-1;
2123 case Instruction::PHI: {
2124 PHINode *PN = cast<PHINode>(U);
2125 unsigned NumIncomingValues = PN->getNumIncomingValues();
2126 // Don't analyze large in-degree PHIs.
2127 if (NumIncomingValues > 4) break;
2128 // Unreachable blocks may have zero-operand PHI nodes.
2129 if (NumIncomingValues == 0) break;
2131 // Take the minimum of all incoming values. This can't infinitely loop
2132 // because of our depth threshold.
2133 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2134 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2135 if (Tmp == 1) return Tmp;
2137 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2142 case Instruction::Trunc:
2143 // FIXME: it's tricky to do anything useful for this, but it is an important
2144 // case for targets like X86.
2148 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2149 // use this information.
2150 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2152 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2154 if (KnownZero.isNegative()) { // sign bit is 0
2156 } else if (KnownOne.isNegative()) { // sign bit is 1;
2163 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2164 // the number of identical bits in the top of the input value.
2166 Mask <<= Mask.getBitWidth()-TyBits;
2167 // Return # leading zeros. We use 'min' here in case Val was zero before
2168 // shifting. We don't want to return '64' as for an i32 "0".
2169 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2172 /// This function computes the integer multiple of Base that equals V.
2173 /// If successful, it returns true and returns the multiple in
2174 /// Multiple. If unsuccessful, it returns false. It looks
2175 /// through SExt instructions only if LookThroughSExt is true.
2176 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2177 bool LookThroughSExt, unsigned Depth) {
2178 const unsigned MaxDepth = 6;
2180 assert(V && "No Value?");
2181 assert(Depth <= MaxDepth && "Limit Search Depth");
2182 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2184 Type *T = V->getType();
2186 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2196 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2197 Constant *BaseVal = ConstantInt::get(T, Base);
2198 if (CO && CO == BaseVal) {
2200 Multiple = ConstantInt::get(T, 1);
2204 if (CI && CI->getZExtValue() % Base == 0) {
2205 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2209 if (Depth == MaxDepth) return false; // Limit search depth.
2211 Operator *I = dyn_cast<Operator>(V);
2212 if (!I) return false;
2214 switch (I->getOpcode()) {
2216 case Instruction::SExt:
2217 if (!LookThroughSExt) return false;
2218 // otherwise fall through to ZExt
2219 case Instruction::ZExt:
2220 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2221 LookThroughSExt, Depth+1);
2222 case Instruction::Shl:
2223 case Instruction::Mul: {
2224 Value *Op0 = I->getOperand(0);
2225 Value *Op1 = I->getOperand(1);
2227 if (I->getOpcode() == Instruction::Shl) {
2228 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2229 if (!Op1CI) return false;
2230 // Turn Op0 << Op1 into Op0 * 2^Op1
2231 APInt Op1Int = Op1CI->getValue();
2232 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2233 APInt API(Op1Int.getBitWidth(), 0);
2234 API.setBit(BitToSet);
2235 Op1 = ConstantInt::get(V->getContext(), API);
2238 Value *Mul0 = nullptr;
2239 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2240 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2241 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2242 if (Op1C->getType()->getPrimitiveSizeInBits() <
2243 MulC->getType()->getPrimitiveSizeInBits())
2244 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2245 if (Op1C->getType()->getPrimitiveSizeInBits() >
2246 MulC->getType()->getPrimitiveSizeInBits())
2247 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2249 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2250 Multiple = ConstantExpr::getMul(MulC, Op1C);
2254 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2255 if (Mul0CI->getValue() == 1) {
2256 // V == Base * Op1, so return Op1
2262 Value *Mul1 = nullptr;
2263 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2264 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2265 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2266 if (Op0C->getType()->getPrimitiveSizeInBits() <
2267 MulC->getType()->getPrimitiveSizeInBits())
2268 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2269 if (Op0C->getType()->getPrimitiveSizeInBits() >
2270 MulC->getType()->getPrimitiveSizeInBits())
2271 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2273 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2274 Multiple = ConstantExpr::getMul(MulC, Op0C);
2278 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2279 if (Mul1CI->getValue() == 1) {
2280 // V == Base * Op0, so return Op0
2288 // We could not determine if V is a multiple of Base.
2292 /// Return true if we can prove that the specified FP value is never equal to
2295 /// NOTE: this function will need to be revisited when we support non-default
2298 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2299 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2300 return !CFP->getValueAPF().isNegZero();
2302 // FIXME: Magic number! At the least, this should be given a name because it's
2303 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2304 // expose it as a parameter, so it can be used for testing / experimenting.
2306 return false; // Limit search depth.
2308 const Operator *I = dyn_cast<Operator>(V);
2309 if (!I) return false;
2311 // Check if the nsz fast-math flag is set
2312 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2313 if (FPO->hasNoSignedZeros())
2316 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2317 if (I->getOpcode() == Instruction::FAdd)
2318 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2319 if (CFP->isNullValue())
2322 // sitofp and uitofp turn into +0.0 for zero.
2323 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2326 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2327 // sqrt(-0.0) = -0.0, no other negative results are possible.
2328 if (II->getIntrinsicID() == Intrinsic::sqrt)
2329 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2331 if (const CallInst *CI = dyn_cast<CallInst>(I))
2332 if (const Function *F = CI->getCalledFunction()) {
2333 if (F->isDeclaration()) {
2335 if (F->getName() == "abs") return true;
2336 // fabs[lf](x) != -0.0
2337 if (F->getName() == "fabs") return true;
2338 if (F->getName() == "fabsf") return true;
2339 if (F->getName() == "fabsl") return true;
2340 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2341 F->getName() == "sqrtl")
2342 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2349 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2350 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2351 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2353 // FIXME: Magic number! At the least, this should be given a name because it's
2354 // used similarly in CannotBeNegativeZero(). A better fix may be to
2355 // expose it as a parameter, so it can be used for testing / experimenting.
2357 return false; // Limit search depth.
2359 const Operator *I = dyn_cast<Operator>(V);
2360 if (!I) return false;
2362 switch (I->getOpcode()) {
2364 case Instruction::FMul:
2365 // x*x is always non-negative or a NaN.
2366 if (I->getOperand(0) == I->getOperand(1))
2369 case Instruction::FAdd:
2370 case Instruction::FDiv:
2371 case Instruction::FRem:
2372 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2373 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2374 case Instruction::FPExt:
2375 case Instruction::FPTrunc:
2376 // Widening/narrowing never change sign.
2377 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2378 case Instruction::Call:
2379 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2380 switch (II->getIntrinsicID()) {
2382 case Intrinsic::exp:
2383 case Intrinsic::exp2:
2384 case Intrinsic::fabs:
2385 case Intrinsic::sqrt:
2387 case Intrinsic::powi:
2388 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2389 // powi(x,n) is non-negative if n is even.
2390 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2393 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2394 case Intrinsic::fma:
2395 case Intrinsic::fmuladd:
2396 // x*x+y is non-negative if y is non-negative.
2397 return I->getOperand(0) == I->getOperand(1) &&
2398 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2405 /// If the specified value can be set by repeating the same byte in memory,
2406 /// return the i8 value that it is represented with. This is
2407 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2408 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2409 /// byte store (e.g. i16 0x1234), return null.
2410 Value *llvm::isBytewiseValue(Value *V) {
2411 // All byte-wide stores are splatable, even of arbitrary variables.
2412 if (V->getType()->isIntegerTy(8)) return V;
2414 // Handle 'null' ConstantArrayZero etc.
2415 if (Constant *C = dyn_cast<Constant>(V))
2416 if (C->isNullValue())
2417 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2419 // Constant float and double values can be handled as integer values if the
2420 // corresponding integer value is "byteable". An important case is 0.0.
2421 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2422 if (CFP->getType()->isFloatTy())
2423 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2424 if (CFP->getType()->isDoubleTy())
2425 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2426 // Don't handle long double formats, which have strange constraints.
2429 // We can handle constant integers that are multiple of 8 bits.
2430 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2431 if (CI->getBitWidth() % 8 == 0) {
2432 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2434 if (!CI->getValue().isSplat(8))
2436 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2440 // A ConstantDataArray/Vector is splatable if all its members are equal and
2442 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2443 Value *Elt = CA->getElementAsConstant(0);
2444 Value *Val = isBytewiseValue(Elt);
2448 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2449 if (CA->getElementAsConstant(I) != Elt)
2455 // Conceptually, we could handle things like:
2456 // %a = zext i8 %X to i16
2457 // %b = shl i16 %a, 8
2458 // %c = or i16 %a, %b
2459 // but until there is an example that actually needs this, it doesn't seem
2460 // worth worrying about.
2465 // This is the recursive version of BuildSubAggregate. It takes a few different
2466 // arguments. Idxs is the index within the nested struct From that we are
2467 // looking at now (which is of type IndexedType). IdxSkip is the number of
2468 // indices from Idxs that should be left out when inserting into the resulting
2469 // struct. To is the result struct built so far, new insertvalue instructions
2471 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2472 SmallVectorImpl<unsigned> &Idxs,
2474 Instruction *InsertBefore) {
2475 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2477 // Save the original To argument so we can modify it
2479 // General case, the type indexed by Idxs is a struct
2480 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2481 // Process each struct element recursively
2484 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2488 // Couldn't find any inserted value for this index? Cleanup
2489 while (PrevTo != OrigTo) {
2490 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2491 PrevTo = Del->getAggregateOperand();
2492 Del->eraseFromParent();
2494 // Stop processing elements
2498 // If we successfully found a value for each of our subaggregates
2502 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2503 // the struct's elements had a value that was inserted directly. In the latter
2504 // case, perhaps we can't determine each of the subelements individually, but
2505 // we might be able to find the complete struct somewhere.
2507 // Find the value that is at that particular spot
2508 Value *V = FindInsertedValue(From, Idxs);
2513 // Insert the value in the new (sub) aggregrate
2514 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2515 "tmp", InsertBefore);
2518 // This helper takes a nested struct and extracts a part of it (which is again a
2519 // struct) into a new value. For example, given the struct:
2520 // { a, { b, { c, d }, e } }
2521 // and the indices "1, 1" this returns
2524 // It does this by inserting an insertvalue for each element in the resulting
2525 // struct, as opposed to just inserting a single struct. This will only work if
2526 // each of the elements of the substruct are known (ie, inserted into From by an
2527 // insertvalue instruction somewhere).
2529 // All inserted insertvalue instructions are inserted before InsertBefore
2530 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2531 Instruction *InsertBefore) {
2532 assert(InsertBefore && "Must have someplace to insert!");
2533 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2535 Value *To = UndefValue::get(IndexedType);
2536 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2537 unsigned IdxSkip = Idxs.size();
2539 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2542 /// Given an aggregrate and an sequence of indices, see if
2543 /// the scalar value indexed is already around as a register, for example if it
2544 /// were inserted directly into the aggregrate.
2546 /// If InsertBefore is not null, this function will duplicate (modified)
2547 /// insertvalues when a part of a nested struct is extracted.
2548 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2549 Instruction *InsertBefore) {
2550 // Nothing to index? Just return V then (this is useful at the end of our
2552 if (idx_range.empty())
2554 // We have indices, so V should have an indexable type.
2555 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2556 "Not looking at a struct or array?");
2557 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2558 "Invalid indices for type?");
2560 if (Constant *C = dyn_cast<Constant>(V)) {
2561 C = C->getAggregateElement(idx_range[0]);
2562 if (!C) return nullptr;
2563 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2566 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2567 // Loop the indices for the insertvalue instruction in parallel with the
2568 // requested indices
2569 const unsigned *req_idx = idx_range.begin();
2570 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2571 i != e; ++i, ++req_idx) {
2572 if (req_idx == idx_range.end()) {
2573 // We can't handle this without inserting insertvalues
2577 // The requested index identifies a part of a nested aggregate. Handle
2578 // this specially. For example,
2579 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2580 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2581 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2582 // This can be changed into
2583 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2584 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2585 // which allows the unused 0,0 element from the nested struct to be
2587 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2591 // This insert value inserts something else than what we are looking for.
2592 // See if the (aggregate) value inserted into has the value we are
2593 // looking for, then.
2595 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2598 // If we end up here, the indices of the insertvalue match with those
2599 // requested (though possibly only partially). Now we recursively look at
2600 // the inserted value, passing any remaining indices.
2601 return FindInsertedValue(I->getInsertedValueOperand(),
2602 makeArrayRef(req_idx, idx_range.end()),
2606 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2607 // If we're extracting a value from an aggregate that was extracted from
2608 // something else, we can extract from that something else directly instead.
2609 // However, we will need to chain I's indices with the requested indices.
2611 // Calculate the number of indices required
2612 unsigned size = I->getNumIndices() + idx_range.size();
2613 // Allocate some space to put the new indices in
2614 SmallVector<unsigned, 5> Idxs;
2616 // Add indices from the extract value instruction
2617 Idxs.append(I->idx_begin(), I->idx_end());
2619 // Add requested indices
2620 Idxs.append(idx_range.begin(), idx_range.end());
2622 assert(Idxs.size() == size
2623 && "Number of indices added not correct?");
2625 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2627 // Otherwise, we don't know (such as, extracting from a function return value
2628 // or load instruction)
2632 /// Analyze the specified pointer to see if it can be expressed as a base
2633 /// pointer plus a constant offset. Return the base and offset to the caller.
2634 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2635 const DataLayout &DL) {
2636 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2637 APInt ByteOffset(BitWidth, 0);
2639 if (Ptr->getType()->isVectorTy())
2642 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2643 APInt GEPOffset(BitWidth, 0);
2644 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2647 ByteOffset += GEPOffset;
2649 Ptr = GEP->getPointerOperand();
2650 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2651 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2652 Ptr = cast<Operator>(Ptr)->getOperand(0);
2653 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2654 if (GA->mayBeOverridden())
2656 Ptr = GA->getAliasee();
2661 Offset = ByteOffset.getSExtValue();
2666 /// This function computes the length of a null-terminated C string pointed to
2667 /// by V. If successful, it returns true and returns the string in Str.
2668 /// If unsuccessful, it returns false.
2669 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2670 uint64_t Offset, bool TrimAtNul) {
2673 // Look through bitcast instructions and geps.
2674 V = V->stripPointerCasts();
2676 // If the value is a GEP instruction or constant expression, treat it as an
2678 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2679 // Make sure the GEP has exactly three arguments.
2680 if (GEP->getNumOperands() != 3)
2683 // Make sure the index-ee is a pointer to array of i8.
2684 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2685 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2686 if (!AT || !AT->getElementType()->isIntegerTy(8))
2689 // Check to make sure that the first operand of the GEP is an integer and
2690 // has value 0 so that we are sure we're indexing into the initializer.
2691 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2692 if (!FirstIdx || !FirstIdx->isZero())
2695 // If the second index isn't a ConstantInt, then this is a variable index
2696 // into the array. If this occurs, we can't say anything meaningful about
2698 uint64_t StartIdx = 0;
2699 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2700 StartIdx = CI->getZExtValue();
2703 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2707 // The GEP instruction, constant or instruction, must reference a global
2708 // variable that is a constant and is initialized. The referenced constant
2709 // initializer is the array that we'll use for optimization.
2710 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2711 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2714 // Handle the all-zeros case
2715 if (GV->getInitializer()->isNullValue()) {
2716 // This is a degenerate case. The initializer is constant zero so the
2717 // length of the string must be zero.
2722 // Must be a Constant Array
2723 const ConstantDataArray *Array =
2724 dyn_cast<ConstantDataArray>(GV->getInitializer());
2725 if (!Array || !Array->isString())
2728 // Get the number of elements in the array
2729 uint64_t NumElts = Array->getType()->getArrayNumElements();
2731 // Start out with the entire array in the StringRef.
2732 Str = Array->getAsString();
2734 if (Offset > NumElts)
2737 // Skip over 'offset' bytes.
2738 Str = Str.substr(Offset);
2741 // Trim off the \0 and anything after it. If the array is not nul
2742 // terminated, we just return the whole end of string. The client may know
2743 // some other way that the string is length-bound.
2744 Str = Str.substr(0, Str.find('\0'));
2749 // These next two are very similar to the above, but also look through PHI
2751 // TODO: See if we can integrate these two together.
2753 /// If we can compute the length of the string pointed to by
2754 /// the specified pointer, return 'len+1'. If we can't, return 0.
2755 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2756 // Look through noop bitcast instructions.
2757 V = V->stripPointerCasts();
2759 // If this is a PHI node, there are two cases: either we have already seen it
2761 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2762 if (!PHIs.insert(PN).second)
2763 return ~0ULL; // already in the set.
2765 // If it was new, see if all the input strings are the same length.
2766 uint64_t LenSoFar = ~0ULL;
2767 for (Value *IncValue : PN->incoming_values()) {
2768 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2769 if (Len == 0) return 0; // Unknown length -> unknown.
2771 if (Len == ~0ULL) continue;
2773 if (Len != LenSoFar && LenSoFar != ~0ULL)
2774 return 0; // Disagree -> unknown.
2778 // Success, all agree.
2782 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2783 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2784 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2785 if (Len1 == 0) return 0;
2786 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2787 if (Len2 == 0) return 0;
2788 if (Len1 == ~0ULL) return Len2;
2789 if (Len2 == ~0ULL) return Len1;
2790 if (Len1 != Len2) return 0;
2794 // Otherwise, see if we can read the string.
2796 if (!getConstantStringInfo(V, StrData))
2799 return StrData.size()+1;
2802 /// If we can compute the length of the string pointed to by
2803 /// the specified pointer, return 'len+1'. If we can't, return 0.
2804 uint64_t llvm::GetStringLength(Value *V) {
2805 if (!V->getType()->isPointerTy()) return 0;
2807 SmallPtrSet<PHINode*, 32> PHIs;
2808 uint64_t Len = GetStringLengthH(V, PHIs);
2809 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2810 // an empty string as a length.
2811 return Len == ~0ULL ? 1 : Len;
2814 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2815 /// previous iteration of the loop was referring to the same object as \p PN.
2816 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2817 // Find the loop-defined value.
2818 Loop *L = LI->getLoopFor(PN->getParent());
2819 if (PN->getNumIncomingValues() != 2)
2822 // Find the value from previous iteration.
2823 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2824 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2825 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2826 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2829 // If a new pointer is loaded in the loop, the pointer references a different
2830 // object in every iteration. E.g.:
2834 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2835 if (!L->isLoopInvariant(Load->getPointerOperand()))
2840 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2841 unsigned MaxLookup) {
2842 if (!V->getType()->isPointerTy())
2844 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2845 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2846 V = GEP->getPointerOperand();
2847 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2848 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2849 V = cast<Operator>(V)->getOperand(0);
2850 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2851 if (GA->mayBeOverridden())
2853 V = GA->getAliasee();
2855 // See if InstructionSimplify knows any relevant tricks.
2856 if (Instruction *I = dyn_cast<Instruction>(V))
2857 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2858 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
2865 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2870 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
2871 const DataLayout &DL, LoopInfo *LI,
2872 unsigned MaxLookup) {
2873 SmallPtrSet<Value *, 4> Visited;
2874 SmallVector<Value *, 4> Worklist;
2875 Worklist.push_back(V);
2877 Value *P = Worklist.pop_back_val();
2878 P = GetUnderlyingObject(P, DL, MaxLookup);
2880 if (!Visited.insert(P).second)
2883 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2884 Worklist.push_back(SI->getTrueValue());
2885 Worklist.push_back(SI->getFalseValue());
2889 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2890 // If this PHI changes the underlying object in every iteration of the
2891 // loop, don't look through it. Consider:
2894 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
2898 // Prev is tracking Curr one iteration behind so they refer to different
2899 // underlying objects.
2900 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
2901 isSameUnderlyingObjectInLoop(PN, LI))
2902 for (Value *IncValue : PN->incoming_values())
2903 Worklist.push_back(IncValue);
2907 Objects.push_back(P);
2908 } while (!Worklist.empty());
2911 /// Return true if the only users of this pointer are lifetime markers.
2912 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2913 for (const User *U : V->users()) {
2914 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2915 if (!II) return false;
2917 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2918 II->getIntrinsicID() != Intrinsic::lifetime_end)
2924 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
2925 Type *Ty, const DataLayout &DL,
2926 const Instruction *CtxI,
2927 const DominatorTree *DT,
2928 const TargetLibraryInfo *TLI) {
2929 assert(Offset.isNonNegative() && "offset can't be negative");
2930 assert(Ty->isSized() && "must be sized");
2932 APInt DerefBytes(Offset.getBitWidth(), 0);
2933 bool CheckForNonNull = false;
2934 if (const Argument *A = dyn_cast<Argument>(BV)) {
2935 DerefBytes = A->getDereferenceableBytes();
2936 if (!DerefBytes.getBoolValue()) {
2937 DerefBytes = A->getDereferenceableOrNullBytes();
2938 CheckForNonNull = true;
2940 } else if (auto CS = ImmutableCallSite(BV)) {
2941 DerefBytes = CS.getDereferenceableBytes(0);
2942 if (!DerefBytes.getBoolValue()) {
2943 DerefBytes = CS.getDereferenceableOrNullBytes(0);
2944 CheckForNonNull = true;
2946 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
2947 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
2948 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2949 DerefBytes = CI->getLimitedValue();
2951 if (!DerefBytes.getBoolValue()) {
2953 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
2954 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2955 DerefBytes = CI->getLimitedValue();
2957 CheckForNonNull = true;
2961 if (DerefBytes.getBoolValue())
2962 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
2963 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
2969 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
2970 const Instruction *CtxI,
2971 const DominatorTree *DT,
2972 const TargetLibraryInfo *TLI) {
2973 Type *VTy = V->getType();
2974 Type *Ty = VTy->getPointerElementType();
2978 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
2979 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
2982 static bool isAligned(const Value *Base, APInt Offset, unsigned Align,
2983 const DataLayout &DL) {
2984 APInt BaseAlign(Offset.getBitWidth(), 0);
2985 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Base))
2986 BaseAlign = AI->getAlignment();
2987 else if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Base))
2988 BaseAlign = GV->getAlignment();
2989 else if (const Argument *A = dyn_cast<Argument>(Base))
2990 BaseAlign = A->getParamAlignment();
2991 else if (auto CS = ImmutableCallSite(Base))
2992 BaseAlign = CS.getAttributes().getParamAlignment(AttributeSet::ReturnIndex);
2993 else if (const LoadInst *LI = dyn_cast<LoadInst>(Base))
2994 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_align)) {
2995 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2996 BaseAlign = CI->getLimitedValue();
3000 Type *Ty = Base->getType()->getPointerElementType();
3001 BaseAlign = DL.getABITypeAlignment(Ty);
3004 APInt Alignment(Offset.getBitWidth(), Align);
3006 assert(Alignment.isPowerOf2() && "must be a power of 2!");
3007 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1));
3010 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) {
3011 APInt Offset(DL.getTypeStoreSizeInBits(Base->getType()), 0);
3012 return isAligned(Base, Offset, Align, DL);
3015 /// Test if V is always a pointer to allocated and suitably aligned memory for
3016 /// a simple load or store.
3017 static bool isDereferenceableAndAlignedPointer(
3018 const Value *V, unsigned Align, const DataLayout &DL,
3019 const Instruction *CtxI, const DominatorTree *DT,
3020 const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited) {
3021 // Note that it is not safe to speculate into a malloc'd region because
3022 // malloc may return null.
3024 // These are obviously ok if aligned.
3025 if (isa<AllocaInst>(V))
3026 return isAligned(V, Align, DL);
3028 // It's not always safe to follow a bitcast, for example:
3029 // bitcast i8* (alloca i8) to i32*
3030 // would result in a 4-byte load from a 1-byte alloca. However,
3031 // if we're casting from a pointer from a type of larger size
3032 // to a type of smaller size (or the same size), and the alignment
3033 // is at least as large as for the resulting pointer type, then
3034 // we can look through the bitcast.
3035 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
3036 Type *STy = BC->getSrcTy()->getPointerElementType(),
3037 *DTy = BC->getDestTy()->getPointerElementType();
3038 if (STy->isSized() && DTy->isSized() &&
3039 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
3040 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
3041 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, DL,
3042 CtxI, DT, TLI, Visited);
3045 // Global variables which can't collapse to null are ok.
3046 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
3047 if (!GV->hasExternalWeakLinkage())
3048 return isAligned(V, Align, DL);
3050 // byval arguments are okay.
3051 if (const Argument *A = dyn_cast<Argument>(V))
3052 if (A->hasByValAttr())
3053 return isAligned(V, Align, DL);
3055 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
3056 return isAligned(V, Align, DL);
3058 // For GEPs, determine if the indexing lands within the allocated object.
3059 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3060 Type *VTy = GEP->getType();
3061 Type *Ty = VTy->getPointerElementType();
3062 const Value *Base = GEP->getPointerOperand();
3064 // Conservatively require that the base pointer be fully dereferenceable
3066 if (!Visited.insert(Base).second)
3068 if (!isDereferenceableAndAlignedPointer(Base, Align, DL, CtxI, DT, TLI,
3072 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
3073 if (!GEP->accumulateConstantOffset(DL, Offset))
3076 // Check if the load is within the bounds of the underlying object
3077 // and offset is aligned.
3078 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3079 Type *BaseType = Base->getType()->getPointerElementType();
3080 assert(isPowerOf2_32(Align) && "must be a power of 2!");
3081 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType)) &&
3082 !(Offset & APInt(Offset.getBitWidth(), Align-1));
3085 // For gc.relocate, look through relocations
3086 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3087 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3088 GCRelocateOperands RelocateInst(I);
3089 return isDereferenceableAndAlignedPointer(
3090 RelocateInst.getDerivedPtr(), Align, DL, CtxI, DT, TLI, Visited);
3093 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3094 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, DL,
3095 CtxI, DT, TLI, Visited);
3097 // If we don't know, assume the worst.
3101 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align,
3102 const DataLayout &DL,
3103 const Instruction *CtxI,
3104 const DominatorTree *DT,
3105 const TargetLibraryInfo *TLI) {
3106 // When dereferenceability information is provided by a dereferenceable
3107 // attribute, we know exactly how many bytes are dereferenceable. If we can
3108 // determine the exact offset to the attributed variable, we can use that
3109 // information here.
3110 Type *VTy = V->getType();
3111 Type *Ty = VTy->getPointerElementType();
3113 // Require ABI alignment for loads without alignment specification
3115 Align = DL.getABITypeAlignment(Ty);
3117 if (Ty->isSized()) {
3118 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3119 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3121 if (Offset.isNonNegative())
3122 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL, CtxI, DT, TLI) &&
3123 isAligned(BV, Offset, Align, DL))
3127 SmallPtrSet<const Value *, 32> Visited;
3128 return ::isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT, TLI,
3132 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3133 const Instruction *CtxI,
3134 const DominatorTree *DT,
3135 const TargetLibraryInfo *TLI) {
3136 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT, TLI);
3139 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3140 const Instruction *CtxI,
3141 const DominatorTree *DT,
3142 const TargetLibraryInfo *TLI) {
3143 const Operator *Inst = dyn_cast<Operator>(V);
3147 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3148 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3152 switch (Inst->getOpcode()) {
3155 case Instruction::UDiv:
3156 case Instruction::URem: {
3157 // x / y is undefined if y == 0.
3159 if (match(Inst->getOperand(1), m_APInt(V)))
3163 case Instruction::SDiv:
3164 case Instruction::SRem: {
3165 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3166 const APInt *Numerator, *Denominator;
3167 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3169 // We cannot hoist this division if the denominator is 0.
3170 if (*Denominator == 0)
3172 // It's safe to hoist if the denominator is not 0 or -1.
3173 if (*Denominator != -1)
3175 // At this point we know that the denominator is -1. It is safe to hoist as
3176 // long we know that the numerator is not INT_MIN.
3177 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3178 return !Numerator->isMinSignedValue();
3179 // The numerator *might* be MinSignedValue.
3182 case Instruction::Load: {
3183 const LoadInst *LI = cast<LoadInst>(Inst);
3184 if (!LI->isUnordered() ||
3185 // Speculative load may create a race that did not exist in the source.
3186 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
3188 const DataLayout &DL = LI->getModule()->getDataLayout();
3189 return isDereferenceableAndAlignedPointer(
3190 LI->getPointerOperand(), LI->getAlignment(), DL, CtxI, DT, TLI);
3192 case Instruction::Call: {
3193 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3194 switch (II->getIntrinsicID()) {
3195 // These synthetic intrinsics have no side-effects and just mark
3196 // information about their operands.
3197 // FIXME: There are other no-op synthetic instructions that potentially
3198 // should be considered at least *safe* to speculate...
3199 case Intrinsic::dbg_declare:
3200 case Intrinsic::dbg_value:
3203 case Intrinsic::bswap:
3204 case Intrinsic::ctlz:
3205 case Intrinsic::ctpop:
3206 case Intrinsic::cttz:
3207 case Intrinsic::objectsize:
3208 case Intrinsic::sadd_with_overflow:
3209 case Intrinsic::smul_with_overflow:
3210 case Intrinsic::ssub_with_overflow:
3211 case Intrinsic::uadd_with_overflow:
3212 case Intrinsic::umul_with_overflow:
3213 case Intrinsic::usub_with_overflow:
3215 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3216 // errno like libm sqrt would.
3217 case Intrinsic::sqrt:
3218 case Intrinsic::fma:
3219 case Intrinsic::fmuladd:
3220 case Intrinsic::fabs:
3221 case Intrinsic::minnum:
3222 case Intrinsic::maxnum:
3224 // TODO: some fp intrinsics are marked as having the same error handling
3225 // as libm. They're safe to speculate when they won't error.
3226 // TODO: are convert_{from,to}_fp16 safe?
3227 // TODO: can we list target-specific intrinsics here?
3231 return false; // The called function could have undefined behavior or
3232 // side-effects, even if marked readnone nounwind.
3234 case Instruction::VAArg:
3235 case Instruction::Alloca:
3236 case Instruction::Invoke:
3237 case Instruction::PHI:
3238 case Instruction::Store:
3239 case Instruction::Ret:
3240 case Instruction::Br:
3241 case Instruction::IndirectBr:
3242 case Instruction::Switch:
3243 case Instruction::Unreachable:
3244 case Instruction::Fence:
3245 case Instruction::AtomicRMW:
3246 case Instruction::AtomicCmpXchg:
3247 case Instruction::LandingPad:
3248 case Instruction::Resume:
3249 case Instruction::CatchPad:
3250 case Instruction::CatchEndPad:
3251 case Instruction::CatchRet:
3252 case Instruction::CleanupPad:
3253 case Instruction::CleanupEndPad:
3254 case Instruction::CleanupRet:
3255 case Instruction::TerminatePad:
3256 return false; // Misc instructions which have effects
3260 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3261 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3264 /// Return true if we know that the specified value is never null.
3265 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3266 assert(V->getType()->isPointerTy() && "V must be pointer type");
3268 // Alloca never returns null, malloc might.
3269 if (isa<AllocaInst>(V)) return true;
3271 // A byval, inalloca, or nonnull argument is never null.
3272 if (const Argument *A = dyn_cast<Argument>(V))
3273 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3275 // A global variable in address space 0 is non null unless extern weak.
3276 // Other address spaces may have null as a valid address for a global,
3277 // so we can't assume anything.
3278 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3279 return !GV->hasExternalWeakLinkage() &&
3280 GV->getType()->getAddressSpace() == 0;
3282 // A Load tagged w/nonnull metadata is never null.
3283 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3284 return LI->getMetadata(LLVMContext::MD_nonnull);
3286 if (auto CS = ImmutableCallSite(V))
3287 if (CS.isReturnNonNull())
3290 // operator new never returns null.
3291 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3297 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3298 const Instruction *CtxI,
3299 const DominatorTree *DT) {
3300 assert(V->getType()->isPointerTy() && "V must be pointer type");
3302 unsigned NumUsesExplored = 0;
3303 for (auto U : V->users()) {
3304 // Avoid massive lists
3305 if (NumUsesExplored >= DomConditionsMaxUses)
3308 // Consider only compare instructions uniquely controlling a branch
3309 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3313 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3316 for (auto *CmpU : Cmp->users()) {
3317 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3321 assert(BI->isConditional() && "uses a comparison!");
3323 BasicBlock *NonNullSuccessor = nullptr;
3324 CmpInst::Predicate Pred;
3326 if (match(const_cast<ICmpInst*>(Cmp),
3327 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3328 if (Pred == ICmpInst::ICMP_EQ)
3329 NonNullSuccessor = BI->getSuccessor(1);
3330 else if (Pred == ICmpInst::ICMP_NE)
3331 NonNullSuccessor = BI->getSuccessor(0);
3334 if (NonNullSuccessor) {
3335 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3336 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3345 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3346 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3347 if (isKnownNonNull(V, TLI))
3350 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3353 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3354 const DataLayout &DL,
3355 AssumptionCache *AC,
3356 const Instruction *CxtI,
3357 const DominatorTree *DT) {
3358 // Multiplying n * m significant bits yields a result of n + m significant
3359 // bits. If the total number of significant bits does not exceed the
3360 // result bit width (minus 1), there is no overflow.
3361 // This means if we have enough leading zero bits in the operands
3362 // we can guarantee that the result does not overflow.
3363 // Ref: "Hacker's Delight" by Henry Warren
3364 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3365 APInt LHSKnownZero(BitWidth, 0);
3366 APInt LHSKnownOne(BitWidth, 0);
3367 APInt RHSKnownZero(BitWidth, 0);
3368 APInt RHSKnownOne(BitWidth, 0);
3369 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3371 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3373 // Note that underestimating the number of zero bits gives a more
3374 // conservative answer.
3375 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3376 RHSKnownZero.countLeadingOnes();
3377 // First handle the easy case: if we have enough zero bits there's
3378 // definitely no overflow.
3379 if (ZeroBits >= BitWidth)
3380 return OverflowResult::NeverOverflows;
3382 // Get the largest possible values for each operand.
3383 APInt LHSMax = ~LHSKnownZero;
3384 APInt RHSMax = ~RHSKnownZero;
3386 // We know the multiply operation doesn't overflow if the maximum values for
3387 // each operand will not overflow after we multiply them together.
3389 LHSMax.umul_ov(RHSMax, MaxOverflow);
3391 return OverflowResult::NeverOverflows;
3393 // We know it always overflows if multiplying the smallest possible values for
3394 // the operands also results in overflow.
3396 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3398 return OverflowResult::AlwaysOverflows;
3400 return OverflowResult::MayOverflow;
3403 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3404 const DataLayout &DL,
3405 AssumptionCache *AC,
3406 const Instruction *CxtI,
3407 const DominatorTree *DT) {
3408 bool LHSKnownNonNegative, LHSKnownNegative;
3409 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3411 if (LHSKnownNonNegative || LHSKnownNegative) {
3412 bool RHSKnownNonNegative, RHSKnownNegative;
3413 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3416 if (LHSKnownNegative && RHSKnownNegative) {
3417 // The sign bit is set in both cases: this MUST overflow.
3418 // Create a simple add instruction, and insert it into the struct.
3419 return OverflowResult::AlwaysOverflows;
3422 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3423 // The sign bit is clear in both cases: this CANNOT overflow.
3424 // Create a simple add instruction, and insert it into the struct.
3425 return OverflowResult::NeverOverflows;
3429 return OverflowResult::MayOverflow;
3432 static OverflowResult computeOverflowForSignedAdd(
3433 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3434 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3435 if (Add && Add->hasNoSignedWrap()) {
3436 return OverflowResult::NeverOverflows;
3439 bool LHSKnownNonNegative, LHSKnownNegative;
3440 bool RHSKnownNonNegative, RHSKnownNegative;
3441 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3443 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3446 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3447 (LHSKnownNegative && RHSKnownNonNegative)) {
3448 // The sign bits are opposite: this CANNOT overflow.
3449 return OverflowResult::NeverOverflows;
3452 // The remaining code needs Add to be available. Early returns if not so.
3454 return OverflowResult::MayOverflow;
3456 // If the sign of Add is the same as at least one of the operands, this add
3457 // CANNOT overflow. This is particularly useful when the sum is
3458 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3460 bool LHSOrRHSKnownNonNegative =
3461 (LHSKnownNonNegative || RHSKnownNonNegative);
3462 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3463 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3464 bool AddKnownNonNegative, AddKnownNegative;
3465 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3466 /*Depth=*/0, AC, CxtI, DT);
3467 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3468 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3469 return OverflowResult::NeverOverflows;
3473 return OverflowResult::MayOverflow;
3476 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3477 const DataLayout &DL,
3478 AssumptionCache *AC,
3479 const Instruction *CxtI,
3480 const DominatorTree *DT) {
3481 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3482 Add, DL, AC, CxtI, DT);
3485 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3486 const DataLayout &DL,
3487 AssumptionCache *AC,
3488 const Instruction *CxtI,
3489 const DominatorTree *DT) {
3490 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3493 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3494 // FIXME: This conservative implementation can be relaxed. E.g. most
3495 // atomic operations are guaranteed to terminate on most platforms
3496 // and most functions terminate.
3498 return !I->isAtomic() && // atomics may never succeed on some platforms
3499 !isa<CallInst>(I) && // could throw and might not terminate
3500 !isa<InvokeInst>(I) && // might not terminate and could throw to
3501 // non-successor (see bug 24185 for details).
3502 !isa<ResumeInst>(I) && // has no successors
3503 !isa<ReturnInst>(I); // has no successors
3506 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3508 // The loop header is guaranteed to be executed for every iteration.
3510 // FIXME: Relax this constraint to cover all basic blocks that are
3511 // guaranteed to be executed at every iteration.
3512 if (I->getParent() != L->getHeader()) return false;
3514 for (const Instruction &LI : *L->getHeader()) {
3515 if (&LI == I) return true;
3516 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3518 llvm_unreachable("Instruction not contained in its own parent basic block.");
3521 bool llvm::propagatesFullPoison(const Instruction *I) {
3522 switch (I->getOpcode()) {
3523 case Instruction::Add:
3524 case Instruction::Sub:
3525 case Instruction::Xor:
3526 case Instruction::Trunc:
3527 case Instruction::BitCast:
3528 case Instruction::AddrSpaceCast:
3529 // These operations all propagate poison unconditionally. Note that poison
3530 // is not any particular value, so xor or subtraction of poison with
3531 // itself still yields poison, not zero.
3534 case Instruction::AShr:
3535 case Instruction::SExt:
3536 // For these operations, one bit of the input is replicated across
3537 // multiple output bits. A replicated poison bit is still poison.
3540 case Instruction::Shl: {
3541 // Left shift *by* a poison value is poison. The number of
3542 // positions to shift is unsigned, so no negative values are
3543 // possible there. Left shift by zero places preserves poison. So
3544 // it only remains to consider left shift of poison by a positive
3545 // number of places.
3547 // A left shift by a positive number of places leaves the lowest order bit
3548 // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3549 // make the poison operand violate that flag, yielding a fresh full-poison
3551 auto *OBO = cast<OverflowingBinaryOperator>(I);
3552 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3555 case Instruction::Mul: {
3556 // A multiplication by zero yields a non-poison zero result, so we need to
3557 // rule out zero as an operand. Conservatively, multiplication by a
3558 // non-zero constant is not multiplication by zero.
3560 // Multiplication by a non-zero constant can leave some bits
3561 // non-poisoned. For example, a multiplication by 2 leaves the lowest
3562 // order bit unpoisoned. So we need to consider that.
3564 // Multiplication by 1 preserves poison. If the multiplication has a
3565 // no-wrap flag, then we can make the poison operand violate that flag
3566 // when multiplied by any integer other than 0 and 1.
3567 auto *OBO = cast<OverflowingBinaryOperator>(I);
3568 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3569 for (Value *V : OBO->operands()) {
3570 if (auto *CI = dyn_cast<ConstantInt>(V)) {
3571 // A ConstantInt cannot yield poison, so we can assume that it is
3572 // the other operand that is poison.
3573 return !CI->isZero();
3580 case Instruction::GetElementPtr:
3581 // A GEP implicitly represents a sequence of additions, subtractions,
3582 // truncations, sign extensions and multiplications. The multiplications
3583 // are by the non-zero sizes of some set of types, so we do not have to be
3584 // concerned with multiplication by zero. If the GEP is in-bounds, then
3585 // these operations are implicitly no-signed-wrap so poison is propagated
3586 // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3587 return cast<GEPOperator>(I)->isInBounds();
3594 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3595 switch (I->getOpcode()) {
3596 case Instruction::Store:
3597 return cast<StoreInst>(I)->getPointerOperand();
3599 case Instruction::Load:
3600 return cast<LoadInst>(I)->getPointerOperand();
3602 case Instruction::AtomicCmpXchg:
3603 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3605 case Instruction::AtomicRMW:
3606 return cast<AtomicRMWInst>(I)->getPointerOperand();
3608 case Instruction::UDiv:
3609 case Instruction::SDiv:
3610 case Instruction::URem:
3611 case Instruction::SRem:
3612 return I->getOperand(1);
3619 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3620 // We currently only look for uses of poison values within the same basic
3621 // block, as that makes it easier to guarantee that the uses will be
3622 // executed given that PoisonI is executed.
3624 // FIXME: Expand this to consider uses beyond the same basic block. To do
3625 // this, look out for the distinction between post-dominance and strong
3627 const BasicBlock *BB = PoisonI->getParent();
3629 // Set of instructions that we have proved will yield poison if PoisonI
3631 SmallSet<const Value *, 16> YieldsPoison;
3632 YieldsPoison.insert(PoisonI);
3634 for (const Instruction *I = PoisonI, *E = BB->end(); I != E;
3635 I = I->getNextNode()) {
3637 const Value *NotPoison = getGuaranteedNonFullPoisonOp(I);
3638 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) return true;
3639 if (!isGuaranteedToTransferExecutionToSuccessor(I)) return false;
3642 // Mark poison that propagates from I through uses of I.
3643 if (YieldsPoison.count(I)) {
3644 for (const User *User : I->users()) {
3645 const Instruction *UserI = cast<Instruction>(User);
3646 if (UserI->getParent() == BB && propagatesFullPoison(UserI))
3647 YieldsPoison.insert(User);
3654 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3658 if (auto *C = dyn_cast<ConstantFP>(V))
3663 static bool isKnownNonZero(Value *V) {
3664 if (auto *C = dyn_cast<ConstantFP>(V))
3665 return !C->isZero();
3669 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3671 Value *CmpLHS, Value *CmpRHS,
3672 Value *TrueVal, Value *FalseVal,
3673 Value *&LHS, Value *&RHS) {
3677 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3678 // return inconsistent results between implementations.
3679 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3680 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3681 // Therefore we behave conservatively and only proceed if at least one of the
3682 // operands is known to not be zero, or if we don't care about signed zeroes.
3685 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3686 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3687 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3688 !isKnownNonZero(CmpRHS))
3689 return {SPF_UNKNOWN, SPNB_NA, false};
3692 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3693 bool Ordered = false;
3695 // When given one NaN and one non-NaN input:
3696 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3697 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3698 // ordered comparison fails), which could be NaN or non-NaN.
3699 // so here we discover exactly what NaN behavior is required/accepted.
3700 if (CmpInst::isFPPredicate(Pred)) {
3701 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3702 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3704 if (LHSSafe && RHSSafe) {
3705 // Both operands are known non-NaN.
3706 NaNBehavior = SPNB_RETURNS_ANY;
3707 } else if (CmpInst::isOrdered(Pred)) {
3708 // An ordered comparison will return false when given a NaN, so it
3712 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3713 NaNBehavior = SPNB_RETURNS_NAN;
3715 NaNBehavior = SPNB_RETURNS_OTHER;
3717 // Completely unsafe.
3718 return {SPF_UNKNOWN, SPNB_NA, false};
3721 // An unordered comparison will return true when given a NaN, so it
3724 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3725 NaNBehavior = SPNB_RETURNS_OTHER;
3727 NaNBehavior = SPNB_RETURNS_NAN;
3729 // Completely unsafe.
3730 return {SPF_UNKNOWN, SPNB_NA, false};
3734 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3735 std::swap(CmpLHS, CmpRHS);
3736 Pred = CmpInst::getSwappedPredicate(Pred);
3737 if (NaNBehavior == SPNB_RETURNS_NAN)
3738 NaNBehavior = SPNB_RETURNS_OTHER;
3739 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3740 NaNBehavior = SPNB_RETURNS_NAN;
3744 // ([if]cmp X, Y) ? X : Y
3745 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3747 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3748 case ICmpInst::ICMP_UGT:
3749 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3750 case ICmpInst::ICMP_SGT:
3751 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3752 case ICmpInst::ICMP_ULT:
3753 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3754 case ICmpInst::ICMP_SLT:
3755 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3756 case FCmpInst::FCMP_UGT:
3757 case FCmpInst::FCMP_UGE:
3758 case FCmpInst::FCMP_OGT:
3759 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3760 case FCmpInst::FCMP_ULT:
3761 case FCmpInst::FCMP_ULE:
3762 case FCmpInst::FCMP_OLT:
3763 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3767 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3768 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3769 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3771 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3772 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3773 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3774 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3777 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3778 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3779 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3780 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3784 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3785 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3786 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3787 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3788 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3791 return {SPF_SMIN, SPNB_NA, false};
3796 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3798 return {SPF_UNKNOWN, SPNB_NA, false};
3801 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
3802 Instruction::CastOps *CastOp) {
3803 CastInst *CI = dyn_cast<CastInst>(V1);
3804 Constant *C = dyn_cast<Constant>(V2);
3805 CastInst *CI2 = dyn_cast<CastInst>(V2);
3808 *CastOp = CI->getOpcode();
3811 // If V1 and V2 are both the same cast from the same type, we can look
3813 if (CI2->getOpcode() == CI->getOpcode() &&
3814 CI2->getSrcTy() == CI->getSrcTy())
3815 return CI2->getOperand(0);
3821 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3822 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3823 // This is only valid if the truncated value can be sign-extended
3824 // back to the original value.
3825 if (ConstantExpr::getSExt(T, C->getType()) == C)
3829 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3830 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3832 if (isa<TruncInst>(CI))
3833 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3835 if (isa<FPToUIInst>(CI))
3836 return ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
3838 if (isa<FPToSIInst>(CI))
3839 return ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
3841 if (isa<UIToFPInst>(CI))
3842 return ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
3844 if (isa<SIToFPInst>(CI))
3845 return ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
3847 if (isa<FPTruncInst>(CI))
3848 return ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
3850 if (isa<FPExtInst>(CI))
3851 return ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
3856 SelectPatternResult llvm::matchSelectPattern(Value *V,
3857 Value *&LHS, Value *&RHS,
3858 Instruction::CastOps *CastOp) {
3859 SelectInst *SI = dyn_cast<SelectInst>(V);
3860 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
3862 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
3863 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
3865 CmpInst::Predicate Pred = CmpI->getPredicate();
3866 Value *CmpLHS = CmpI->getOperand(0);
3867 Value *CmpRHS = CmpI->getOperand(1);
3868 Value *TrueVal = SI->getTrueValue();
3869 Value *FalseVal = SI->getFalseValue();
3871 if (isa<FPMathOperator>(CmpI))
3872 FMF = CmpI->getFastMathFlags();
3875 if (CmpI->isEquality())
3876 return {SPF_UNKNOWN, SPNB_NA, false};
3878 // Deal with type mismatches.
3879 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
3880 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
3881 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3882 cast<CastInst>(TrueVal)->getOperand(0), C,
3884 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
3885 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3886 C, cast<CastInst>(FalseVal)->getOperand(0),
3889 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,