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 SrcTy->isFloatingPointTy()) &&
1070 // TODO: For now, not handling conversions like:
1071 // (bitcast i64 %x to <2 x i32>)
1072 !I->getType()->isVectorTy()) {
1073 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1078 case Instruction::SExt: {
1079 // Compute the bits in the result that are not present in the input.
1080 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1082 KnownZero = KnownZero.trunc(SrcBitWidth);
1083 KnownOne = KnownOne.trunc(SrcBitWidth);
1084 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1085 KnownZero = KnownZero.zext(BitWidth);
1086 KnownOne = KnownOne.zext(BitWidth);
1088 // If the sign bit of the input is known set or clear, then we know the
1089 // top bits of the result.
1090 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1091 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1092 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1093 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1096 case Instruction::Shl:
1097 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1098 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1099 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1100 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1101 KnownZero <<= ShiftAmt;
1102 KnownOne <<= ShiftAmt;
1103 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1106 case Instruction::LShr:
1107 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1108 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1109 // Compute the new bits that are at the top now.
1110 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1112 // Unsigned shift right.
1113 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1114 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1115 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1116 // high bits known zero.
1117 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1120 case Instruction::AShr:
1121 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1122 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1123 // Compute the new bits that are at the top now.
1124 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1126 // Signed shift right.
1127 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1128 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1129 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1131 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1132 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1133 KnownZero |= HighBits;
1134 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1135 KnownOne |= HighBits;
1138 case Instruction::Sub: {
1139 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1140 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1141 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1145 case Instruction::Add: {
1146 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1147 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1148 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1152 case Instruction::SRem:
1153 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1154 APInt RA = Rem->getValue().abs();
1155 if (RA.isPowerOf2()) {
1156 APInt LowBits = RA - 1;
1157 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1160 // The low bits of the first operand are unchanged by the srem.
1161 KnownZero = KnownZero2 & LowBits;
1162 KnownOne = KnownOne2 & LowBits;
1164 // If the first operand is non-negative or has all low bits zero, then
1165 // the upper bits are all zero.
1166 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1167 KnownZero |= ~LowBits;
1169 // If the first operand is negative and not all low bits are zero, then
1170 // the upper bits are all one.
1171 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1172 KnownOne |= ~LowBits;
1174 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1178 // The sign bit is the LHS's sign bit, except when the result of the
1179 // remainder is zero.
1180 if (KnownZero.isNonNegative()) {
1181 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1182 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1184 // If it's known zero, our sign bit is also zero.
1185 if (LHSKnownZero.isNegative())
1186 KnownZero.setBit(BitWidth - 1);
1190 case Instruction::URem: {
1191 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1192 APInt RA = Rem->getValue();
1193 if (RA.isPowerOf2()) {
1194 APInt LowBits = (RA - 1);
1195 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1197 KnownZero |= ~LowBits;
1198 KnownOne &= LowBits;
1203 // Since the result is less than or equal to either operand, any leading
1204 // zero bits in either operand must also exist in the result.
1205 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1206 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1208 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1209 KnownZero2.countLeadingOnes());
1210 KnownOne.clearAllBits();
1211 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1215 case Instruction::Alloca: {
1216 AllocaInst *AI = cast<AllocaInst>(I);
1217 unsigned Align = AI->getAlignment();
1219 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1222 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1225 case Instruction::GetElementPtr: {
1226 // Analyze all of the subscripts of this getelementptr instruction
1227 // to determine if we can prove known low zero bits.
1228 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1229 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1231 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1233 gep_type_iterator GTI = gep_type_begin(I);
1234 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1235 Value *Index = I->getOperand(i);
1236 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1237 // Handle struct member offset arithmetic.
1239 // Handle case when index is vector zeroinitializer
1240 Constant *CIndex = cast<Constant>(Index);
1241 if (CIndex->isZeroValue())
1244 if (CIndex->getType()->isVectorTy())
1245 Index = CIndex->getSplatValue();
1247 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1248 const StructLayout *SL = DL.getStructLayout(STy);
1249 uint64_t Offset = SL->getElementOffset(Idx);
1250 TrailZ = std::min<unsigned>(TrailZ,
1251 countTrailingZeros(Offset));
1253 // Handle array index arithmetic.
1254 Type *IndexedTy = GTI.getIndexedType();
1255 if (!IndexedTy->isSized()) {
1259 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1260 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1261 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1262 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1264 TrailZ = std::min(TrailZ,
1265 unsigned(countTrailingZeros(TypeSize) +
1266 LocalKnownZero.countTrailingOnes()));
1270 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1273 case Instruction::PHI: {
1274 PHINode *P = cast<PHINode>(I);
1275 // Handle the case of a simple two-predecessor recurrence PHI.
1276 // There's a lot more that could theoretically be done here, but
1277 // this is sufficient to catch some interesting cases.
1278 if (P->getNumIncomingValues() == 2) {
1279 for (unsigned i = 0; i != 2; ++i) {
1280 Value *L = P->getIncomingValue(i);
1281 Value *R = P->getIncomingValue(!i);
1282 Operator *LU = dyn_cast<Operator>(L);
1285 unsigned Opcode = LU->getOpcode();
1286 // Check for operations that have the property that if
1287 // both their operands have low zero bits, the result
1288 // will have low zero bits.
1289 if (Opcode == Instruction::Add ||
1290 Opcode == Instruction::Sub ||
1291 Opcode == Instruction::And ||
1292 Opcode == Instruction::Or ||
1293 Opcode == Instruction::Mul) {
1294 Value *LL = LU->getOperand(0);
1295 Value *LR = LU->getOperand(1);
1296 // Find a recurrence.
1303 // Ok, we have a PHI of the form L op= R. Check for low
1305 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1307 // We need to take the minimum number of known bits
1308 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1309 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1311 KnownZero = APInt::getLowBitsSet(BitWidth,
1312 std::min(KnownZero2.countTrailingOnes(),
1313 KnownZero3.countTrailingOnes()));
1319 // Unreachable blocks may have zero-operand PHI nodes.
1320 if (P->getNumIncomingValues() == 0)
1323 // Otherwise take the unions of the known bit sets of the operands,
1324 // taking conservative care to avoid excessive recursion.
1325 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1326 // Skip if every incoming value references to ourself.
1327 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1330 KnownZero = APInt::getAllOnesValue(BitWidth);
1331 KnownOne = APInt::getAllOnesValue(BitWidth);
1332 for (Value *IncValue : P->incoming_values()) {
1333 // Skip direct self references.
1334 if (IncValue == P) continue;
1336 KnownZero2 = APInt(BitWidth, 0);
1337 KnownOne2 = APInt(BitWidth, 0);
1338 // Recurse, but cap the recursion to one level, because we don't
1339 // want to waste time spinning around in loops.
1340 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1342 KnownZero &= KnownZero2;
1343 KnownOne &= KnownOne2;
1344 // If all bits have been ruled out, there's no need to check
1346 if (!KnownZero && !KnownOne)
1352 case Instruction::Call:
1353 case Instruction::Invoke:
1354 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1355 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1356 // If a range metadata is attached to this IntrinsicInst, intersect the
1357 // explicit range specified by the metadata and the implicit range of
1359 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1360 switch (II->getIntrinsicID()) {
1362 case Intrinsic::bswap:
1363 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1365 KnownZero |= KnownZero2.byteSwap();
1366 KnownOne |= KnownOne2.byteSwap();
1368 case Intrinsic::ctlz:
1369 case Intrinsic::cttz: {
1370 unsigned LowBits = Log2_32(BitWidth)+1;
1371 // If this call is undefined for 0, the result will be less than 2^n.
1372 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1374 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1377 case Intrinsic::ctpop: {
1378 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1380 // We can bound the space the count needs. Also, bits known to be zero
1381 // can't contribute to the population.
1382 unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation();
1383 unsigned LeadingZeros =
1384 APInt(BitWidth, BitsPossiblySet).countLeadingZeros();
1385 assert(LeadingZeros >= 0 && LeadingZeros <= BitWidth);
1386 KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros);
1387 KnownOne &= ~KnownZero;
1388 // TODO: we could bound KnownOne using the lower bound on the number
1389 // of bits which might be set provided by popcnt KnownOne2.
1392 case Intrinsic::fabs: {
1393 Type *Ty = II->getType();
1394 APInt SignBit = APInt::getSignBit(Ty->getScalarSizeInBits());
1395 KnownZero |= APInt::getSplat(Ty->getPrimitiveSizeInBits(), SignBit);
1398 case Intrinsic::x86_sse42_crc32_64_64:
1399 KnownZero |= APInt::getHighBitsSet(64, 32);
1404 case Instruction::ExtractValue:
1405 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1406 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1407 if (EVI->getNumIndices() != 1) break;
1408 if (EVI->getIndices()[0] == 0) {
1409 switch (II->getIntrinsicID()) {
1411 case Intrinsic::uadd_with_overflow:
1412 case Intrinsic::sadd_with_overflow:
1413 computeKnownBitsAddSub(true, II->getArgOperand(0),
1414 II->getArgOperand(1), false, KnownZero,
1415 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1417 case Intrinsic::usub_with_overflow:
1418 case Intrinsic::ssub_with_overflow:
1419 computeKnownBitsAddSub(false, II->getArgOperand(0),
1420 II->getArgOperand(1), false, KnownZero,
1421 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1423 case Intrinsic::umul_with_overflow:
1424 case Intrinsic::smul_with_overflow:
1425 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1426 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1435 static unsigned getAlignment(const Value *V, const DataLayout &DL) {
1437 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1438 Align = GO->getAlignment();
1440 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1441 Type *ObjectType = GVar->getType()->getElementType();
1442 if (ObjectType->isSized()) {
1443 // If the object is defined in the current Module, we'll be giving
1444 // it the preferred alignment. Otherwise, we have to assume that it
1445 // may only have the minimum ABI alignment.
1446 if (GVar->isStrongDefinitionForLinker())
1447 Align = DL.getPreferredAlignment(GVar);
1449 Align = DL.getABITypeAlignment(ObjectType);
1453 } else if (const Argument *A = dyn_cast<Argument>(V)) {
1454 Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1456 if (!Align && A->hasStructRetAttr()) {
1457 // An sret parameter has at least the ABI alignment of the return type.
1458 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1459 if (EltTy->isSized())
1460 Align = DL.getABITypeAlignment(EltTy);
1462 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
1463 Align = AI->getAlignment();
1464 else if (auto CS = ImmutableCallSite(V))
1465 Align = CS.getAttributes().getParamAlignment(AttributeSet::ReturnIndex);
1466 else if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1467 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_align)) {
1468 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
1469 Align = CI->getLimitedValue();
1475 /// Determine which bits of V are known to be either zero or one and return
1476 /// them in the KnownZero/KnownOne bit sets.
1478 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1479 /// we cannot optimize based on the assumption that it is zero without changing
1480 /// it to be an explicit zero. If we don't change it to zero, other code could
1481 /// optimized based on the contradictory assumption that it is non-zero.
1482 /// Because instcombine aggressively folds operations with undef args anyway,
1483 /// this won't lose us code quality.
1485 /// This function is defined on values with integer type, values with pointer
1486 /// type, and vectors of integers. In the case
1487 /// where V is a vector, known zero, and known one values are the
1488 /// same width as the vector element, and the bit is set only if it is true
1489 /// for all of the elements in the vector.
1490 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1491 const DataLayout &DL, unsigned Depth, const Query &Q) {
1492 assert(V && "No Value?");
1493 assert(Depth <= MaxDepth && "Limit Search Depth");
1494 unsigned BitWidth = KnownZero.getBitWidth();
1496 assert((V->getType()->isIntOrIntVectorTy() ||
1497 V->getType()->isFPOrFPVectorTy() ||
1498 V->getType()->getScalarType()->isPointerTy()) &&
1499 "Not integer, floating point, or pointer type!");
1500 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1501 (!V->getType()->isIntOrIntVectorTy() ||
1502 V->getType()->getScalarSizeInBits() == BitWidth) &&
1503 KnownZero.getBitWidth() == BitWidth &&
1504 KnownOne.getBitWidth() == BitWidth &&
1505 "V, KnownOne and KnownZero should have same BitWidth");
1507 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1508 // We know all of the bits for a constant!
1509 KnownOne = CI->getValue();
1510 KnownZero = ~KnownOne;
1513 // Null and aggregate-zero are all-zeros.
1514 if (isa<ConstantPointerNull>(V) ||
1515 isa<ConstantAggregateZero>(V)) {
1516 KnownOne.clearAllBits();
1517 KnownZero = APInt::getAllOnesValue(BitWidth);
1520 // Handle a constant vector by taking the intersection of the known bits of
1521 // each element. There is no real need to handle ConstantVector here, because
1522 // we don't handle undef in any particularly useful way.
1523 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1524 // We know that CDS must be a vector of integers. Take the intersection of
1526 KnownZero.setAllBits(); KnownOne.setAllBits();
1527 APInt Elt(KnownZero.getBitWidth(), 0);
1528 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1529 Elt = CDS->getElementAsInteger(i);
1536 // Start out not knowing anything.
1537 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1539 // Limit search depth.
1540 // All recursive calls that increase depth must come after this.
1541 if (Depth == MaxDepth)
1544 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1545 // the bits of its aliasee.
1546 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1547 if (!GA->mayBeOverridden())
1548 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1552 if (Operator *I = dyn_cast<Operator>(V))
1553 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1555 // Aligned pointers have trailing zeros - refine KnownZero set
1556 if (V->getType()->isPointerTy()) {
1557 unsigned Align = getAlignment(V, DL);
1559 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1562 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1563 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1564 // computeKnownBitsFromOperator.
1566 // Check whether a nearby assume intrinsic can determine some known bits.
1567 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1569 // Check whether there's a dominating condition which implies something about
1570 // this value at the given context.
1571 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1572 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1575 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1578 /// Determine whether the sign bit is known to be zero or one.
1579 /// Convenience wrapper around computeKnownBits.
1580 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1581 const DataLayout &DL, unsigned Depth, const Query &Q) {
1582 unsigned BitWidth = getBitWidth(V->getType(), DL);
1588 APInt ZeroBits(BitWidth, 0);
1589 APInt OneBits(BitWidth, 0);
1590 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1591 KnownOne = OneBits[BitWidth - 1];
1592 KnownZero = ZeroBits[BitWidth - 1];
1595 /// Return true if the given value is known to have exactly one
1596 /// bit set when defined. For vectors return true if every element is known to
1597 /// be a power of two when defined. Supports values with integer or pointer
1598 /// types and vectors of integers.
1599 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1600 const Query &Q, const DataLayout &DL) {
1601 if (Constant *C = dyn_cast<Constant>(V)) {
1602 if (C->isNullValue())
1604 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1605 return CI->getValue().isPowerOf2();
1606 // TODO: Handle vector constants.
1609 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1610 // it is shifted off the end then the result is undefined.
1611 if (match(V, m_Shl(m_One(), m_Value())))
1614 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1615 // bottom. If it is shifted off the bottom then the result is undefined.
1616 if (match(V, m_LShr(m_SignBit(), m_Value())))
1619 // The remaining tests are all recursive, so bail out if we hit the limit.
1620 if (Depth++ == MaxDepth)
1623 Value *X = nullptr, *Y = nullptr;
1624 // A shift of a power of two is a power of two or zero.
1625 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1626 match(V, m_Shr(m_Value(X), m_Value()))))
1627 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1629 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1630 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1632 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1633 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1634 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1636 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1637 // A power of two and'd with anything is a power of two or zero.
1638 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1639 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1641 // X & (-X) is always a power of two or zero.
1642 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1647 // Adding a power-of-two or zero to the same power-of-two or zero yields
1648 // either the original power-of-two, a larger power-of-two or zero.
1649 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1650 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1651 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1652 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1653 match(X, m_And(m_Value(), m_Specific(Y))))
1654 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1656 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1657 match(Y, m_And(m_Value(), m_Specific(X))))
1658 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1661 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1662 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1663 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1665 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1666 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1667 // If i8 V is a power of two or zero:
1668 // ZeroBits: 1 1 1 0 1 1 1 1
1669 // ~ZeroBits: 0 0 0 1 0 0 0 0
1670 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1671 // If OrZero isn't set, we cannot give back a zero result.
1672 // Make sure either the LHS or RHS has a bit set.
1673 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1678 // An exact divide or right shift can only shift off zero bits, so the result
1679 // is a power of two only if the first operand is a power of two and not
1680 // copying a sign bit (sdiv int_min, 2).
1681 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1682 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1683 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1690 /// \brief Test whether a GEP's result is known to be non-null.
1692 /// Uses properties inherent in a GEP to try to determine whether it is known
1695 /// Currently this routine does not support vector GEPs.
1696 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1697 unsigned Depth, const Query &Q) {
1698 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1701 // FIXME: Support vector-GEPs.
1702 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1704 // If the base pointer is non-null, we cannot walk to a null address with an
1705 // inbounds GEP in address space zero.
1706 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1709 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1710 // If so, then the GEP cannot produce a null pointer, as doing so would
1711 // inherently violate the inbounds contract within address space zero.
1712 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1713 GTI != GTE; ++GTI) {
1714 // Struct types are easy -- they must always be indexed by a constant.
1715 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1716 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1717 unsigned ElementIdx = OpC->getZExtValue();
1718 const StructLayout *SL = DL.getStructLayout(STy);
1719 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1720 if (ElementOffset > 0)
1725 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1726 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1729 // Fast path the constant operand case both for efficiency and so we don't
1730 // increment Depth when just zipping down an all-constant GEP.
1731 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1737 // We post-increment Depth here because while isKnownNonZero increments it
1738 // as well, when we pop back up that increment won't persist. We don't want
1739 // to recurse 10k times just because we have 10k GEP operands. We don't
1740 // bail completely out because we want to handle constant GEPs regardless
1742 if (Depth++ >= MaxDepth)
1745 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1752 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1753 /// ensure that the value it's attached to is never Value? 'RangeType' is
1754 /// is the type of the value described by the range.
1755 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1756 const APInt& Value) {
1757 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1758 assert(NumRanges >= 1);
1759 for (unsigned i = 0; i < NumRanges; ++i) {
1760 ConstantInt *Lower =
1761 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1762 ConstantInt *Upper =
1763 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1764 ConstantRange Range(Lower->getValue(), Upper->getValue());
1765 if (Range.contains(Value))
1771 /// Return true if the given value is known to be non-zero when defined.
1772 /// For vectors return true if every element is known to be non-zero when
1773 /// defined. Supports values with integer or pointer type and vectors of
1775 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1777 if (Constant *C = dyn_cast<Constant>(V)) {
1778 if (C->isNullValue())
1780 if (isa<ConstantInt>(C))
1781 // Must be non-zero due to null test above.
1783 // TODO: Handle vectors
1787 if (Instruction* I = dyn_cast<Instruction>(V)) {
1788 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1789 // If the possible ranges don't contain zero, then the value is
1790 // definitely non-zero.
1791 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1792 const APInt ZeroValue(Ty->getBitWidth(), 0);
1793 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1799 // The remaining tests are all recursive, so bail out if we hit the limit.
1800 if (Depth++ >= MaxDepth)
1803 // Check for pointer simplifications.
1804 if (V->getType()->isPointerTy()) {
1805 if (isKnownNonNull(V))
1807 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1808 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1812 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1814 // X | Y != 0 if X != 0 or Y != 0.
1815 Value *X = nullptr, *Y = nullptr;
1816 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1817 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1819 // ext X != 0 if X != 0.
1820 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1821 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1823 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1824 // if the lowest bit is shifted off the end.
1825 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1826 // shl nuw can't remove any non-zero bits.
1827 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1828 if (BO->hasNoUnsignedWrap())
1829 return isKnownNonZero(X, DL, Depth, Q);
1831 APInt KnownZero(BitWidth, 0);
1832 APInt KnownOne(BitWidth, 0);
1833 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1837 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1838 // defined if the sign bit is shifted off the end.
1839 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1840 // shr exact can only shift out zero bits.
1841 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1843 return isKnownNonZero(X, DL, Depth, Q);
1845 bool XKnownNonNegative, XKnownNegative;
1846 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1850 // If the shifter operand is a constant, and all of the bits shifted
1851 // out are known to be zero, and X is known non-zero then at least one
1852 // non-zero bit must remain.
1853 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1854 APInt KnownZero(BitWidth, 0);
1855 APInt KnownOne(BitWidth, 0);
1856 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1858 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1859 // Is there a known one in the portion not shifted out?
1860 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1862 // Are all the bits to be shifted out known zero?
1863 if (KnownZero.countTrailingOnes() >= ShiftVal)
1864 return isKnownNonZero(X, DL, Depth, Q);
1867 // div exact can only produce a zero if the dividend is zero.
1868 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1869 return isKnownNonZero(X, DL, Depth, Q);
1872 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1873 bool XKnownNonNegative, XKnownNegative;
1874 bool YKnownNonNegative, YKnownNegative;
1875 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1876 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1878 // If X and Y are both non-negative (as signed values) then their sum is not
1879 // zero unless both X and Y are zero.
1880 if (XKnownNonNegative && YKnownNonNegative)
1881 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1884 // If X and Y are both negative (as signed values) then their sum is not
1885 // zero unless both X and Y equal INT_MIN.
1886 if (BitWidth && XKnownNegative && YKnownNegative) {
1887 APInt KnownZero(BitWidth, 0);
1888 APInt KnownOne(BitWidth, 0);
1889 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1890 // The sign bit of X is set. If some other bit is set then X is not equal
1892 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1893 if ((KnownOne & Mask) != 0)
1895 // The sign bit of Y is set. If some other bit is set then Y is not equal
1897 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1898 if ((KnownOne & Mask) != 0)
1902 // The sum of a non-negative number and a power of two is not zero.
1903 if (XKnownNonNegative &&
1904 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1906 if (YKnownNonNegative &&
1907 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1911 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1912 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1913 // If X and Y are non-zero then so is X * Y as long as the multiplication
1914 // does not overflow.
1915 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1916 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
1919 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1920 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1921 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
1922 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
1926 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1927 // Try and detect a recurrence that monotonically increases from a
1928 // starting value, as these are common as induction variables.
1929 if (PN->getNumIncomingValues() == 2) {
1930 Value *Start = PN->getIncomingValue(0);
1931 Value *Induction = PN->getIncomingValue(1);
1932 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
1933 std::swap(Start, Induction);
1934 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
1935 if (!C->isZero() && !C->isNegative()) {
1937 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
1938 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
1946 if (!BitWidth) return false;
1947 APInt KnownZero(BitWidth, 0);
1948 APInt KnownOne(BitWidth, 0);
1949 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1950 return KnownOne != 0;
1953 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1954 /// simplify operations downstream. Mask is known to be zero for bits that V
1957 /// This function is defined on values with integer type, values with pointer
1958 /// type, and vectors of integers. In the case
1959 /// where V is a vector, the mask, known zero, and known one values are the
1960 /// same width as the vector element, and the bit is set only if it is true
1961 /// for all of the elements in the vector.
1962 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
1963 unsigned Depth, const Query &Q) {
1964 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1965 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1966 return (KnownZero & Mask) == Mask;
1971 /// Return the number of times the sign bit of the register is replicated into
1972 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1973 /// (itself), but other cases can give us information. For example, immediately
1974 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1975 /// other, so we return 3.
1977 /// 'Op' must have a scalar integer type.
1979 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
1981 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
1983 unsigned FirstAnswer = 1;
1985 // Note that ConstantInt is handled by the general computeKnownBits case
1989 return 1; // Limit search depth.
1991 Operator *U = dyn_cast<Operator>(V);
1992 switch (Operator::getOpcode(V)) {
1994 case Instruction::SExt:
1995 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1996 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
1998 case Instruction::SDiv: {
1999 const APInt *Denominator;
2000 // sdiv X, C -> adds log(C) sign bits.
2001 if (match(U->getOperand(1), m_APInt(Denominator))) {
2003 // Ignore non-positive denominator.
2004 if (!Denominator->isStrictlyPositive())
2007 // Calculate the incoming numerator bits.
2008 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2010 // Add floor(log(C)) bits to the numerator bits.
2011 return std::min(TyBits, NumBits + Denominator->logBase2());
2016 case Instruction::SRem: {
2017 const APInt *Denominator;
2018 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2019 // positive constant. This let us put a lower bound on the number of sign
2021 if (match(U->getOperand(1), m_APInt(Denominator))) {
2023 // Ignore non-positive denominator.
2024 if (!Denominator->isStrictlyPositive())
2027 // Calculate the incoming numerator bits. SRem by a positive constant
2028 // can't lower the number of sign bits.
2030 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2032 // Calculate the leading sign bit constraints by examining the
2033 // denominator. Given that the denominator is positive, there are two
2036 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2037 // (1 << ceilLogBase2(C)).
2039 // 2. the numerator is negative. Then the result range is (-C,0] and
2040 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2042 // Thus a lower bound on the number of sign bits is `TyBits -
2043 // ceilLogBase2(C)`.
2045 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2046 return std::max(NumrBits, ResBits);
2051 case Instruction::AShr: {
2052 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2053 // ashr X, C -> adds C sign bits. Vectors too.
2055 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2056 Tmp += ShAmt->getZExtValue();
2057 if (Tmp > TyBits) Tmp = TyBits;
2061 case Instruction::Shl: {
2063 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2064 // shl destroys sign bits.
2065 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2066 Tmp2 = ShAmt->getZExtValue();
2067 if (Tmp2 >= TyBits || // Bad shift.
2068 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2073 case Instruction::And:
2074 case Instruction::Or:
2075 case Instruction::Xor: // NOT is handled here.
2076 // Logical binary ops preserve the number of sign bits at the worst.
2077 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2079 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2080 FirstAnswer = std::min(Tmp, Tmp2);
2081 // We computed what we know about the sign bits as our first
2082 // answer. Now proceed to the generic code that uses
2083 // computeKnownBits, and pick whichever answer is better.
2087 case Instruction::Select:
2088 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2089 if (Tmp == 1) return 1; // Early out.
2090 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2091 return std::min(Tmp, Tmp2);
2093 case Instruction::Add:
2094 // Add can have at most one carry bit. Thus we know that the output
2095 // is, at worst, one more bit than the inputs.
2096 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2097 if (Tmp == 1) return 1; // Early out.
2099 // Special case decrementing a value (ADD X, -1):
2100 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2101 if (CRHS->isAllOnesValue()) {
2102 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2103 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2106 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2108 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2111 // If we are subtracting one from a positive number, there is no carry
2112 // out of the result.
2113 if (KnownZero.isNegative())
2117 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2118 if (Tmp2 == 1) return 1;
2119 return std::min(Tmp, Tmp2)-1;
2121 case Instruction::Sub:
2122 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2123 if (Tmp2 == 1) return 1;
2126 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2127 if (CLHS->isNullValue()) {
2128 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2129 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2131 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2133 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2136 // If the input is known to be positive (the sign bit is known clear),
2137 // the output of the NEG has the same number of sign bits as the input.
2138 if (KnownZero.isNegative())
2141 // Otherwise, we treat this like a SUB.
2144 // Sub can have at most one carry bit. Thus we know that the output
2145 // is, at worst, one more bit than the inputs.
2146 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2147 if (Tmp == 1) return 1; // Early out.
2148 return std::min(Tmp, Tmp2)-1;
2150 case Instruction::PHI: {
2151 PHINode *PN = cast<PHINode>(U);
2152 unsigned NumIncomingValues = PN->getNumIncomingValues();
2153 // Don't analyze large in-degree PHIs.
2154 if (NumIncomingValues > 4) break;
2155 // Unreachable blocks may have zero-operand PHI nodes.
2156 if (NumIncomingValues == 0) break;
2158 // Take the minimum of all incoming values. This can't infinitely loop
2159 // because of our depth threshold.
2160 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2161 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2162 if (Tmp == 1) return Tmp;
2164 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2169 case Instruction::Trunc:
2170 // FIXME: it's tricky to do anything useful for this, but it is an important
2171 // case for targets like X86.
2175 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2176 // use this information.
2177 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2179 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2181 if (KnownZero.isNegative()) { // sign bit is 0
2183 } else if (KnownOne.isNegative()) { // sign bit is 1;
2190 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2191 // the number of identical bits in the top of the input value.
2193 Mask <<= Mask.getBitWidth()-TyBits;
2194 // Return # leading zeros. We use 'min' here in case Val was zero before
2195 // shifting. We don't want to return '64' as for an i32 "0".
2196 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2199 /// This function computes the integer multiple of Base that equals V.
2200 /// If successful, it returns true and returns the multiple in
2201 /// Multiple. If unsuccessful, it returns false. It looks
2202 /// through SExt instructions only if LookThroughSExt is true.
2203 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2204 bool LookThroughSExt, unsigned Depth) {
2205 const unsigned MaxDepth = 6;
2207 assert(V && "No Value?");
2208 assert(Depth <= MaxDepth && "Limit Search Depth");
2209 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2211 Type *T = V->getType();
2213 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2223 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2224 Constant *BaseVal = ConstantInt::get(T, Base);
2225 if (CO && CO == BaseVal) {
2227 Multiple = ConstantInt::get(T, 1);
2231 if (CI && CI->getZExtValue() % Base == 0) {
2232 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2236 if (Depth == MaxDepth) return false; // Limit search depth.
2238 Operator *I = dyn_cast<Operator>(V);
2239 if (!I) return false;
2241 switch (I->getOpcode()) {
2243 case Instruction::SExt:
2244 if (!LookThroughSExt) return false;
2245 // otherwise fall through to ZExt
2246 case Instruction::ZExt:
2247 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2248 LookThroughSExt, Depth+1);
2249 case Instruction::Shl:
2250 case Instruction::Mul: {
2251 Value *Op0 = I->getOperand(0);
2252 Value *Op1 = I->getOperand(1);
2254 if (I->getOpcode() == Instruction::Shl) {
2255 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2256 if (!Op1CI) return false;
2257 // Turn Op0 << Op1 into Op0 * 2^Op1
2258 APInt Op1Int = Op1CI->getValue();
2259 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2260 APInt API(Op1Int.getBitWidth(), 0);
2261 API.setBit(BitToSet);
2262 Op1 = ConstantInt::get(V->getContext(), API);
2265 Value *Mul0 = nullptr;
2266 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2267 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2268 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2269 if (Op1C->getType()->getPrimitiveSizeInBits() <
2270 MulC->getType()->getPrimitiveSizeInBits())
2271 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2272 if (Op1C->getType()->getPrimitiveSizeInBits() >
2273 MulC->getType()->getPrimitiveSizeInBits())
2274 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2276 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2277 Multiple = ConstantExpr::getMul(MulC, Op1C);
2281 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2282 if (Mul0CI->getValue() == 1) {
2283 // V == Base * Op1, so return Op1
2289 Value *Mul1 = nullptr;
2290 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2291 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2292 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2293 if (Op0C->getType()->getPrimitiveSizeInBits() <
2294 MulC->getType()->getPrimitiveSizeInBits())
2295 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2296 if (Op0C->getType()->getPrimitiveSizeInBits() >
2297 MulC->getType()->getPrimitiveSizeInBits())
2298 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2300 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2301 Multiple = ConstantExpr::getMul(MulC, Op0C);
2305 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2306 if (Mul1CI->getValue() == 1) {
2307 // V == Base * Op0, so return Op0
2315 // We could not determine if V is a multiple of Base.
2319 /// Return true if we can prove that the specified FP value is never equal to
2322 /// NOTE: this function will need to be revisited when we support non-default
2325 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2326 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2327 return !CFP->getValueAPF().isNegZero();
2329 // FIXME: Magic number! At the least, this should be given a name because it's
2330 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2331 // expose it as a parameter, so it can be used for testing / experimenting.
2333 return false; // Limit search depth.
2335 const Operator *I = dyn_cast<Operator>(V);
2336 if (!I) return false;
2338 // Check if the nsz fast-math flag is set
2339 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2340 if (FPO->hasNoSignedZeros())
2343 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2344 if (I->getOpcode() == Instruction::FAdd)
2345 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2346 if (CFP->isNullValue())
2349 // sitofp and uitofp turn into +0.0 for zero.
2350 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2353 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2354 // sqrt(-0.0) = -0.0, no other negative results are possible.
2355 if (II->getIntrinsicID() == Intrinsic::sqrt)
2356 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2358 if (const CallInst *CI = dyn_cast<CallInst>(I))
2359 if (const Function *F = CI->getCalledFunction()) {
2360 if (F->isDeclaration()) {
2362 if (F->getName() == "abs") return true;
2363 // fabs[lf](x) != -0.0
2364 if (F->getName() == "fabs") return true;
2365 if (F->getName() == "fabsf") return true;
2366 if (F->getName() == "fabsl") return true;
2367 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2368 F->getName() == "sqrtl")
2369 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2376 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2377 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2378 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2380 // FIXME: Magic number! At the least, this should be given a name because it's
2381 // used similarly in CannotBeNegativeZero(). A better fix may be to
2382 // expose it as a parameter, so it can be used for testing / experimenting.
2384 return false; // Limit search depth.
2386 const Operator *I = dyn_cast<Operator>(V);
2387 if (!I) return false;
2389 switch (I->getOpcode()) {
2391 case Instruction::FMul:
2392 // x*x is always non-negative or a NaN.
2393 if (I->getOperand(0) == I->getOperand(1))
2396 case Instruction::FAdd:
2397 case Instruction::FDiv:
2398 case Instruction::FRem:
2399 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2400 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2401 case Instruction::FPExt:
2402 case Instruction::FPTrunc:
2403 // Widening/narrowing never change sign.
2404 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2405 case Instruction::Call:
2406 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2407 switch (II->getIntrinsicID()) {
2409 case Intrinsic::exp:
2410 case Intrinsic::exp2:
2411 case Intrinsic::fabs:
2412 case Intrinsic::sqrt:
2414 case Intrinsic::powi:
2415 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2416 // powi(x,n) is non-negative if n is even.
2417 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2420 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2421 case Intrinsic::fma:
2422 case Intrinsic::fmuladd:
2423 // x*x+y is non-negative if y is non-negative.
2424 return I->getOperand(0) == I->getOperand(1) &&
2425 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2432 /// If the specified value can be set by repeating the same byte in memory,
2433 /// return the i8 value that it is represented with. This is
2434 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2435 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2436 /// byte store (e.g. i16 0x1234), return null.
2437 Value *llvm::isBytewiseValue(Value *V) {
2438 // All byte-wide stores are splatable, even of arbitrary variables.
2439 if (V->getType()->isIntegerTy(8)) return V;
2441 // Handle 'null' ConstantArrayZero etc.
2442 if (Constant *C = dyn_cast<Constant>(V))
2443 if (C->isNullValue())
2444 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2446 // Constant float and double values can be handled as integer values if the
2447 // corresponding integer value is "byteable". An important case is 0.0.
2448 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2449 if (CFP->getType()->isFloatTy())
2450 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2451 if (CFP->getType()->isDoubleTy())
2452 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2453 // Don't handle long double formats, which have strange constraints.
2456 // We can handle constant integers that are multiple of 8 bits.
2457 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2458 if (CI->getBitWidth() % 8 == 0) {
2459 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2461 if (!CI->getValue().isSplat(8))
2463 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2467 // A ConstantDataArray/Vector is splatable if all its members are equal and
2469 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2470 Value *Elt = CA->getElementAsConstant(0);
2471 Value *Val = isBytewiseValue(Elt);
2475 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2476 if (CA->getElementAsConstant(I) != Elt)
2482 // Conceptually, we could handle things like:
2483 // %a = zext i8 %X to i16
2484 // %b = shl i16 %a, 8
2485 // %c = or i16 %a, %b
2486 // but until there is an example that actually needs this, it doesn't seem
2487 // worth worrying about.
2492 // This is the recursive version of BuildSubAggregate. It takes a few different
2493 // arguments. Idxs is the index within the nested struct From that we are
2494 // looking at now (which is of type IndexedType). IdxSkip is the number of
2495 // indices from Idxs that should be left out when inserting into the resulting
2496 // struct. To is the result struct built so far, new insertvalue instructions
2498 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2499 SmallVectorImpl<unsigned> &Idxs,
2501 Instruction *InsertBefore) {
2502 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2504 // Save the original To argument so we can modify it
2506 // General case, the type indexed by Idxs is a struct
2507 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2508 // Process each struct element recursively
2511 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2515 // Couldn't find any inserted value for this index? Cleanup
2516 while (PrevTo != OrigTo) {
2517 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2518 PrevTo = Del->getAggregateOperand();
2519 Del->eraseFromParent();
2521 // Stop processing elements
2525 // If we successfully found a value for each of our subaggregates
2529 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2530 // the struct's elements had a value that was inserted directly. In the latter
2531 // case, perhaps we can't determine each of the subelements individually, but
2532 // we might be able to find the complete struct somewhere.
2534 // Find the value that is at that particular spot
2535 Value *V = FindInsertedValue(From, Idxs);
2540 // Insert the value in the new (sub) aggregrate
2541 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2542 "tmp", InsertBefore);
2545 // This helper takes a nested struct and extracts a part of it (which is again a
2546 // struct) into a new value. For example, given the struct:
2547 // { a, { b, { c, d }, e } }
2548 // and the indices "1, 1" this returns
2551 // It does this by inserting an insertvalue for each element in the resulting
2552 // struct, as opposed to just inserting a single struct. This will only work if
2553 // each of the elements of the substruct are known (ie, inserted into From by an
2554 // insertvalue instruction somewhere).
2556 // All inserted insertvalue instructions are inserted before InsertBefore
2557 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2558 Instruction *InsertBefore) {
2559 assert(InsertBefore && "Must have someplace to insert!");
2560 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2562 Value *To = UndefValue::get(IndexedType);
2563 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2564 unsigned IdxSkip = Idxs.size();
2566 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2569 /// Given an aggregrate and an sequence of indices, see if
2570 /// the scalar value indexed is already around as a register, for example if it
2571 /// were inserted directly into the aggregrate.
2573 /// If InsertBefore is not null, this function will duplicate (modified)
2574 /// insertvalues when a part of a nested struct is extracted.
2575 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2576 Instruction *InsertBefore) {
2577 // Nothing to index? Just return V then (this is useful at the end of our
2579 if (idx_range.empty())
2581 // We have indices, so V should have an indexable type.
2582 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2583 "Not looking at a struct or array?");
2584 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2585 "Invalid indices for type?");
2587 if (Constant *C = dyn_cast<Constant>(V)) {
2588 C = C->getAggregateElement(idx_range[0]);
2589 if (!C) return nullptr;
2590 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2593 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2594 // Loop the indices for the insertvalue instruction in parallel with the
2595 // requested indices
2596 const unsigned *req_idx = idx_range.begin();
2597 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2598 i != e; ++i, ++req_idx) {
2599 if (req_idx == idx_range.end()) {
2600 // We can't handle this without inserting insertvalues
2604 // The requested index identifies a part of a nested aggregate. Handle
2605 // this specially. For example,
2606 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2607 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2608 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2609 // This can be changed into
2610 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2611 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2612 // which allows the unused 0,0 element from the nested struct to be
2614 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2618 // This insert value inserts something else than what we are looking for.
2619 // See if the (aggregate) value inserted into has the value we are
2620 // looking for, then.
2622 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2625 // If we end up here, the indices of the insertvalue match with those
2626 // requested (though possibly only partially). Now we recursively look at
2627 // the inserted value, passing any remaining indices.
2628 return FindInsertedValue(I->getInsertedValueOperand(),
2629 makeArrayRef(req_idx, idx_range.end()),
2633 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2634 // If we're extracting a value from an aggregate that was extracted from
2635 // something else, we can extract from that something else directly instead.
2636 // However, we will need to chain I's indices with the requested indices.
2638 // Calculate the number of indices required
2639 unsigned size = I->getNumIndices() + idx_range.size();
2640 // Allocate some space to put the new indices in
2641 SmallVector<unsigned, 5> Idxs;
2643 // Add indices from the extract value instruction
2644 Idxs.append(I->idx_begin(), I->idx_end());
2646 // Add requested indices
2647 Idxs.append(idx_range.begin(), idx_range.end());
2649 assert(Idxs.size() == size
2650 && "Number of indices added not correct?");
2652 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2654 // Otherwise, we don't know (such as, extracting from a function return value
2655 // or load instruction)
2659 /// Analyze the specified pointer to see if it can be expressed as a base
2660 /// pointer plus a constant offset. Return the base and offset to the caller.
2661 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2662 const DataLayout &DL) {
2663 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2664 APInt ByteOffset(BitWidth, 0);
2666 if (Ptr->getType()->isVectorTy())
2669 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2670 APInt GEPOffset(BitWidth, 0);
2671 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2674 ByteOffset += GEPOffset;
2676 Ptr = GEP->getPointerOperand();
2677 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2678 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2679 Ptr = cast<Operator>(Ptr)->getOperand(0);
2680 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2681 if (GA->mayBeOverridden())
2683 Ptr = GA->getAliasee();
2688 Offset = ByteOffset.getSExtValue();
2693 /// This function computes the length of a null-terminated C string pointed to
2694 /// by V. If successful, it returns true and returns the string in Str.
2695 /// If unsuccessful, it returns false.
2696 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2697 uint64_t Offset, bool TrimAtNul) {
2700 // Look through bitcast instructions and geps.
2701 V = V->stripPointerCasts();
2703 // If the value is a GEP instruction or constant expression, treat it as an
2705 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2706 // Make sure the GEP has exactly three arguments.
2707 if (GEP->getNumOperands() != 3)
2710 // Make sure the index-ee is a pointer to array of i8.
2711 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2712 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2713 if (!AT || !AT->getElementType()->isIntegerTy(8))
2716 // Check to make sure that the first operand of the GEP is an integer and
2717 // has value 0 so that we are sure we're indexing into the initializer.
2718 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2719 if (!FirstIdx || !FirstIdx->isZero())
2722 // If the second index isn't a ConstantInt, then this is a variable index
2723 // into the array. If this occurs, we can't say anything meaningful about
2725 uint64_t StartIdx = 0;
2726 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2727 StartIdx = CI->getZExtValue();
2730 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2734 // The GEP instruction, constant or instruction, must reference a global
2735 // variable that is a constant and is initialized. The referenced constant
2736 // initializer is the array that we'll use for optimization.
2737 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2738 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2741 // Handle the all-zeros case
2742 if (GV->getInitializer()->isNullValue()) {
2743 // This is a degenerate case. The initializer is constant zero so the
2744 // length of the string must be zero.
2749 // Must be a Constant Array
2750 const ConstantDataArray *Array =
2751 dyn_cast<ConstantDataArray>(GV->getInitializer());
2752 if (!Array || !Array->isString())
2755 // Get the number of elements in the array
2756 uint64_t NumElts = Array->getType()->getArrayNumElements();
2758 // Start out with the entire array in the StringRef.
2759 Str = Array->getAsString();
2761 if (Offset > NumElts)
2764 // Skip over 'offset' bytes.
2765 Str = Str.substr(Offset);
2768 // Trim off the \0 and anything after it. If the array is not nul
2769 // terminated, we just return the whole end of string. The client may know
2770 // some other way that the string is length-bound.
2771 Str = Str.substr(0, Str.find('\0'));
2776 // These next two are very similar to the above, but also look through PHI
2778 // TODO: See if we can integrate these two together.
2780 /// If we can compute the length of the string pointed to by
2781 /// the specified pointer, return 'len+1'. If we can't, return 0.
2782 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2783 // Look through noop bitcast instructions.
2784 V = V->stripPointerCasts();
2786 // If this is a PHI node, there are two cases: either we have already seen it
2788 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2789 if (!PHIs.insert(PN).second)
2790 return ~0ULL; // already in the set.
2792 // If it was new, see if all the input strings are the same length.
2793 uint64_t LenSoFar = ~0ULL;
2794 for (Value *IncValue : PN->incoming_values()) {
2795 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2796 if (Len == 0) return 0; // Unknown length -> unknown.
2798 if (Len == ~0ULL) continue;
2800 if (Len != LenSoFar && LenSoFar != ~0ULL)
2801 return 0; // Disagree -> unknown.
2805 // Success, all agree.
2809 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2810 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2811 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2812 if (Len1 == 0) return 0;
2813 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2814 if (Len2 == 0) return 0;
2815 if (Len1 == ~0ULL) return Len2;
2816 if (Len2 == ~0ULL) return Len1;
2817 if (Len1 != Len2) return 0;
2821 // Otherwise, see if we can read the string.
2823 if (!getConstantStringInfo(V, StrData))
2826 return StrData.size()+1;
2829 /// If we can compute the length of the string pointed to by
2830 /// the specified pointer, return 'len+1'. If we can't, return 0.
2831 uint64_t llvm::GetStringLength(Value *V) {
2832 if (!V->getType()->isPointerTy()) return 0;
2834 SmallPtrSet<PHINode*, 32> PHIs;
2835 uint64_t Len = GetStringLengthH(V, PHIs);
2836 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2837 // an empty string as a length.
2838 return Len == ~0ULL ? 1 : Len;
2841 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2842 /// previous iteration of the loop was referring to the same object as \p PN.
2843 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2844 // Find the loop-defined value.
2845 Loop *L = LI->getLoopFor(PN->getParent());
2846 if (PN->getNumIncomingValues() != 2)
2849 // Find the value from previous iteration.
2850 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2851 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2852 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2853 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2856 // If a new pointer is loaded in the loop, the pointer references a different
2857 // object in every iteration. E.g.:
2861 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2862 if (!L->isLoopInvariant(Load->getPointerOperand()))
2867 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2868 unsigned MaxLookup) {
2869 if (!V->getType()->isPointerTy())
2871 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2872 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2873 V = GEP->getPointerOperand();
2874 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2875 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2876 V = cast<Operator>(V)->getOperand(0);
2877 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2878 if (GA->mayBeOverridden())
2880 V = GA->getAliasee();
2882 // See if InstructionSimplify knows any relevant tricks.
2883 if (Instruction *I = dyn_cast<Instruction>(V))
2884 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2885 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
2892 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2897 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
2898 const DataLayout &DL, LoopInfo *LI,
2899 unsigned MaxLookup) {
2900 SmallPtrSet<Value *, 4> Visited;
2901 SmallVector<Value *, 4> Worklist;
2902 Worklist.push_back(V);
2904 Value *P = Worklist.pop_back_val();
2905 P = GetUnderlyingObject(P, DL, MaxLookup);
2907 if (!Visited.insert(P).second)
2910 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2911 Worklist.push_back(SI->getTrueValue());
2912 Worklist.push_back(SI->getFalseValue());
2916 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2917 // If this PHI changes the underlying object in every iteration of the
2918 // loop, don't look through it. Consider:
2921 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
2925 // Prev is tracking Curr one iteration behind so they refer to different
2926 // underlying objects.
2927 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
2928 isSameUnderlyingObjectInLoop(PN, LI))
2929 for (Value *IncValue : PN->incoming_values())
2930 Worklist.push_back(IncValue);
2934 Objects.push_back(P);
2935 } while (!Worklist.empty());
2938 /// Return true if the only users of this pointer are lifetime markers.
2939 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2940 for (const User *U : V->users()) {
2941 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2942 if (!II) return false;
2944 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2945 II->getIntrinsicID() != Intrinsic::lifetime_end)
2951 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
2952 Type *Ty, const DataLayout &DL,
2953 const Instruction *CtxI,
2954 const DominatorTree *DT,
2955 const TargetLibraryInfo *TLI) {
2956 assert(Offset.isNonNegative() && "offset can't be negative");
2957 assert(Ty->isSized() && "must be sized");
2959 APInt DerefBytes(Offset.getBitWidth(), 0);
2960 bool CheckForNonNull = false;
2961 if (const Argument *A = dyn_cast<Argument>(BV)) {
2962 DerefBytes = A->getDereferenceableBytes();
2963 if (!DerefBytes.getBoolValue()) {
2964 DerefBytes = A->getDereferenceableOrNullBytes();
2965 CheckForNonNull = true;
2967 } else if (auto CS = ImmutableCallSite(BV)) {
2968 DerefBytes = CS.getDereferenceableBytes(0);
2969 if (!DerefBytes.getBoolValue()) {
2970 DerefBytes = CS.getDereferenceableOrNullBytes(0);
2971 CheckForNonNull = true;
2973 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
2974 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
2975 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2976 DerefBytes = CI->getLimitedValue();
2978 if (!DerefBytes.getBoolValue()) {
2980 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
2981 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2982 DerefBytes = CI->getLimitedValue();
2984 CheckForNonNull = true;
2988 if (DerefBytes.getBoolValue())
2989 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
2990 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
2996 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
2997 const Instruction *CtxI,
2998 const DominatorTree *DT,
2999 const TargetLibraryInfo *TLI) {
3000 Type *VTy = V->getType();
3001 Type *Ty = VTy->getPointerElementType();
3005 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3006 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
3009 static bool isAligned(const Value *Base, APInt Offset, unsigned Align,
3010 const DataLayout &DL) {
3011 APInt BaseAlign(Offset.getBitWidth(), getAlignment(Base, DL));
3014 Type *Ty = Base->getType()->getPointerElementType();
3015 BaseAlign = DL.getABITypeAlignment(Ty);
3018 APInt Alignment(Offset.getBitWidth(), Align);
3020 assert(Alignment.isPowerOf2() && "must be a power of 2!");
3021 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1));
3024 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) {
3025 APInt Offset(DL.getTypeStoreSizeInBits(Base->getType()), 0);
3026 return isAligned(Base, Offset, Align, DL);
3029 /// Test if V is always a pointer to allocated and suitably aligned memory for
3030 /// a simple load or store.
3031 static bool isDereferenceableAndAlignedPointer(
3032 const Value *V, unsigned Align, const DataLayout &DL,
3033 const Instruction *CtxI, const DominatorTree *DT,
3034 const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited) {
3035 // Note that it is not safe to speculate into a malloc'd region because
3036 // malloc may return null.
3038 // These are obviously ok if aligned.
3039 if (isa<AllocaInst>(V))
3040 return isAligned(V, Align, DL);
3042 // It's not always safe to follow a bitcast, for example:
3043 // bitcast i8* (alloca i8) to i32*
3044 // would result in a 4-byte load from a 1-byte alloca. However,
3045 // if we're casting from a pointer from a type of larger size
3046 // to a type of smaller size (or the same size), and the alignment
3047 // is at least as large as for the resulting pointer type, then
3048 // we can look through the bitcast.
3049 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
3050 Type *STy = BC->getSrcTy()->getPointerElementType(),
3051 *DTy = BC->getDestTy()->getPointerElementType();
3052 if (STy->isSized() && DTy->isSized() &&
3053 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
3054 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
3055 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, DL,
3056 CtxI, DT, TLI, Visited);
3059 // Global variables which can't collapse to null are ok.
3060 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
3061 if (!GV->hasExternalWeakLinkage())
3062 return isAligned(V, Align, DL);
3064 // byval arguments are okay.
3065 if (const Argument *A = dyn_cast<Argument>(V))
3066 if (A->hasByValAttr())
3067 return isAligned(V, Align, DL);
3069 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
3070 return isAligned(V, Align, DL);
3072 // For GEPs, determine if the indexing lands within the allocated object.
3073 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3074 Type *VTy = GEP->getType();
3075 Type *Ty = VTy->getPointerElementType();
3076 const Value *Base = GEP->getPointerOperand();
3078 // Conservatively require that the base pointer be fully dereferenceable
3080 if (!Visited.insert(Base).second)
3082 if (!isDereferenceableAndAlignedPointer(Base, Align, DL, CtxI, DT, TLI,
3086 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
3087 if (!GEP->accumulateConstantOffset(DL, Offset))
3090 // Check if the load is within the bounds of the underlying object
3091 // and offset is aligned.
3092 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3093 Type *BaseType = Base->getType()->getPointerElementType();
3094 assert(isPowerOf2_32(Align) && "must be a power of 2!");
3095 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType)) &&
3096 !(Offset & APInt(Offset.getBitWidth(), Align-1));
3099 // For gc.relocate, look through relocations
3100 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3101 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3102 GCRelocateOperands RelocateInst(I);
3103 return isDereferenceableAndAlignedPointer(
3104 RelocateInst.getDerivedPtr(), Align, DL, CtxI, DT, TLI, Visited);
3107 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3108 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, DL,
3109 CtxI, DT, TLI, Visited);
3111 // If we don't know, assume the worst.
3115 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align,
3116 const DataLayout &DL,
3117 const Instruction *CtxI,
3118 const DominatorTree *DT,
3119 const TargetLibraryInfo *TLI) {
3120 // When dereferenceability information is provided by a dereferenceable
3121 // attribute, we know exactly how many bytes are dereferenceable. If we can
3122 // determine the exact offset to the attributed variable, we can use that
3123 // information here.
3124 Type *VTy = V->getType();
3125 Type *Ty = VTy->getPointerElementType();
3127 // Require ABI alignment for loads without alignment specification
3129 Align = DL.getABITypeAlignment(Ty);
3131 if (Ty->isSized()) {
3132 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3133 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3135 if (Offset.isNonNegative())
3136 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL, CtxI, DT, TLI) &&
3137 isAligned(BV, Offset, Align, DL))
3141 SmallPtrSet<const Value *, 32> Visited;
3142 return ::isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT, TLI,
3146 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3147 const Instruction *CtxI,
3148 const DominatorTree *DT,
3149 const TargetLibraryInfo *TLI) {
3150 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT, TLI);
3153 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3154 const Instruction *CtxI,
3155 const DominatorTree *DT,
3156 const TargetLibraryInfo *TLI) {
3157 const Operator *Inst = dyn_cast<Operator>(V);
3161 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3162 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3166 switch (Inst->getOpcode()) {
3169 case Instruction::UDiv:
3170 case Instruction::URem: {
3171 // x / y is undefined if y == 0.
3173 if (match(Inst->getOperand(1), m_APInt(V)))
3177 case Instruction::SDiv:
3178 case Instruction::SRem: {
3179 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3180 const APInt *Numerator, *Denominator;
3181 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3183 // We cannot hoist this division if the denominator is 0.
3184 if (*Denominator == 0)
3186 // It's safe to hoist if the denominator is not 0 or -1.
3187 if (*Denominator != -1)
3189 // At this point we know that the denominator is -1. It is safe to hoist as
3190 // long we know that the numerator is not INT_MIN.
3191 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3192 return !Numerator->isMinSignedValue();
3193 // The numerator *might* be MinSignedValue.
3196 case Instruction::Load: {
3197 const LoadInst *LI = cast<LoadInst>(Inst);
3198 if (!LI->isUnordered() ||
3199 // Speculative load may create a race that did not exist in the source.
3200 LI->getParent()->getParent()->hasFnAttribute(
3201 Attribute::SanitizeThread) ||
3202 // Speculative load may load data from dirty regions.
3203 LI->getParent()->getParent()->hasFnAttribute(
3204 Attribute::SanitizeAddress))
3206 const DataLayout &DL = LI->getModule()->getDataLayout();
3207 return isDereferenceableAndAlignedPointer(
3208 LI->getPointerOperand(), LI->getAlignment(), DL, CtxI, DT, TLI);
3210 case Instruction::Call: {
3211 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3212 switch (II->getIntrinsicID()) {
3213 // These synthetic intrinsics have no side-effects and just mark
3214 // information about their operands.
3215 // FIXME: There are other no-op synthetic instructions that potentially
3216 // should be considered at least *safe* to speculate...
3217 case Intrinsic::dbg_declare:
3218 case Intrinsic::dbg_value:
3221 case Intrinsic::bswap:
3222 case Intrinsic::ctlz:
3223 case Intrinsic::ctpop:
3224 case Intrinsic::cttz:
3225 case Intrinsic::objectsize:
3226 case Intrinsic::sadd_with_overflow:
3227 case Intrinsic::smul_with_overflow:
3228 case Intrinsic::ssub_with_overflow:
3229 case Intrinsic::uadd_with_overflow:
3230 case Intrinsic::umul_with_overflow:
3231 case Intrinsic::usub_with_overflow:
3233 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3234 // errno like libm sqrt would.
3235 case Intrinsic::sqrt:
3236 case Intrinsic::fma:
3237 case Intrinsic::fmuladd:
3238 case Intrinsic::fabs:
3239 case Intrinsic::minnum:
3240 case Intrinsic::maxnum:
3242 // TODO: some fp intrinsics are marked as having the same error handling
3243 // as libm. They're safe to speculate when they won't error.
3244 // TODO: are convert_{from,to}_fp16 safe?
3245 // TODO: can we list target-specific intrinsics here?
3249 return false; // The called function could have undefined behavior or
3250 // side-effects, even if marked readnone nounwind.
3252 case Instruction::VAArg:
3253 case Instruction::Alloca:
3254 case Instruction::Invoke:
3255 case Instruction::PHI:
3256 case Instruction::Store:
3257 case Instruction::Ret:
3258 case Instruction::Br:
3259 case Instruction::IndirectBr:
3260 case Instruction::Switch:
3261 case Instruction::Unreachable:
3262 case Instruction::Fence:
3263 case Instruction::AtomicRMW:
3264 case Instruction::AtomicCmpXchg:
3265 case Instruction::LandingPad:
3266 case Instruction::Resume:
3267 case Instruction::CatchPad:
3268 case Instruction::CatchEndPad:
3269 case Instruction::CatchRet:
3270 case Instruction::CleanupPad:
3271 case Instruction::CleanupEndPad:
3272 case Instruction::CleanupRet:
3273 case Instruction::TerminatePad:
3274 return false; // Misc instructions which have effects
3278 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3279 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3282 /// Return true if we know that the specified value is never null.
3283 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3284 assert(V->getType()->isPointerTy() && "V must be pointer type");
3286 // Alloca never returns null, malloc might.
3287 if (isa<AllocaInst>(V)) return true;
3289 // A byval, inalloca, or nonnull argument is never null.
3290 if (const Argument *A = dyn_cast<Argument>(V))
3291 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3293 // A global variable in address space 0 is non null unless extern weak.
3294 // Other address spaces may have null as a valid address for a global,
3295 // so we can't assume anything.
3296 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3297 return !GV->hasExternalWeakLinkage() &&
3298 GV->getType()->getAddressSpace() == 0;
3300 // A Load tagged w/nonnull metadata is never null.
3301 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3302 return LI->getMetadata(LLVMContext::MD_nonnull);
3304 if (auto CS = ImmutableCallSite(V))
3305 if (CS.isReturnNonNull())
3308 // operator new never returns null.
3309 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3315 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3316 const Instruction *CtxI,
3317 const DominatorTree *DT) {
3318 assert(V->getType()->isPointerTy() && "V must be pointer type");
3320 unsigned NumUsesExplored = 0;
3321 for (auto U : V->users()) {
3322 // Avoid massive lists
3323 if (NumUsesExplored >= DomConditionsMaxUses)
3326 // Consider only compare instructions uniquely controlling a branch
3327 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3331 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3334 for (auto *CmpU : Cmp->users()) {
3335 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3339 assert(BI->isConditional() && "uses a comparison!");
3341 BasicBlock *NonNullSuccessor = nullptr;
3342 CmpInst::Predicate Pred;
3344 if (match(const_cast<ICmpInst*>(Cmp),
3345 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3346 if (Pred == ICmpInst::ICMP_EQ)
3347 NonNullSuccessor = BI->getSuccessor(1);
3348 else if (Pred == ICmpInst::ICMP_NE)
3349 NonNullSuccessor = BI->getSuccessor(0);
3352 if (NonNullSuccessor) {
3353 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3354 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3363 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3364 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3365 if (isKnownNonNull(V, TLI))
3368 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3371 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3372 const DataLayout &DL,
3373 AssumptionCache *AC,
3374 const Instruction *CxtI,
3375 const DominatorTree *DT) {
3376 // Multiplying n * m significant bits yields a result of n + m significant
3377 // bits. If the total number of significant bits does not exceed the
3378 // result bit width (minus 1), there is no overflow.
3379 // This means if we have enough leading zero bits in the operands
3380 // we can guarantee that the result does not overflow.
3381 // Ref: "Hacker's Delight" by Henry Warren
3382 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3383 APInt LHSKnownZero(BitWidth, 0);
3384 APInt LHSKnownOne(BitWidth, 0);
3385 APInt RHSKnownZero(BitWidth, 0);
3386 APInt RHSKnownOne(BitWidth, 0);
3387 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3389 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3391 // Note that underestimating the number of zero bits gives a more
3392 // conservative answer.
3393 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3394 RHSKnownZero.countLeadingOnes();
3395 // First handle the easy case: if we have enough zero bits there's
3396 // definitely no overflow.
3397 if (ZeroBits >= BitWidth)
3398 return OverflowResult::NeverOverflows;
3400 // Get the largest possible values for each operand.
3401 APInt LHSMax = ~LHSKnownZero;
3402 APInt RHSMax = ~RHSKnownZero;
3404 // We know the multiply operation doesn't overflow if the maximum values for
3405 // each operand will not overflow after we multiply them together.
3407 LHSMax.umul_ov(RHSMax, MaxOverflow);
3409 return OverflowResult::NeverOverflows;
3411 // We know it always overflows if multiplying the smallest possible values for
3412 // the operands also results in overflow.
3414 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3416 return OverflowResult::AlwaysOverflows;
3418 return OverflowResult::MayOverflow;
3421 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3422 const DataLayout &DL,
3423 AssumptionCache *AC,
3424 const Instruction *CxtI,
3425 const DominatorTree *DT) {
3426 bool LHSKnownNonNegative, LHSKnownNegative;
3427 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3429 if (LHSKnownNonNegative || LHSKnownNegative) {
3430 bool RHSKnownNonNegative, RHSKnownNegative;
3431 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3434 if (LHSKnownNegative && RHSKnownNegative) {
3435 // The sign bit is set in both cases: this MUST overflow.
3436 // Create a simple add instruction, and insert it into the struct.
3437 return OverflowResult::AlwaysOverflows;
3440 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3441 // The sign bit is clear in both cases: this CANNOT overflow.
3442 // Create a simple add instruction, and insert it into the struct.
3443 return OverflowResult::NeverOverflows;
3447 return OverflowResult::MayOverflow;
3450 static OverflowResult computeOverflowForSignedAdd(
3451 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3452 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3453 if (Add && Add->hasNoSignedWrap()) {
3454 return OverflowResult::NeverOverflows;
3457 bool LHSKnownNonNegative, LHSKnownNegative;
3458 bool RHSKnownNonNegative, RHSKnownNegative;
3459 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3461 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3464 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3465 (LHSKnownNegative && RHSKnownNonNegative)) {
3466 // The sign bits are opposite: this CANNOT overflow.
3467 return OverflowResult::NeverOverflows;
3470 // The remaining code needs Add to be available. Early returns if not so.
3472 return OverflowResult::MayOverflow;
3474 // If the sign of Add is the same as at least one of the operands, this add
3475 // CANNOT overflow. This is particularly useful when the sum is
3476 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3478 bool LHSOrRHSKnownNonNegative =
3479 (LHSKnownNonNegative || RHSKnownNonNegative);
3480 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3481 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3482 bool AddKnownNonNegative, AddKnownNegative;
3483 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3484 /*Depth=*/0, AC, CxtI, DT);
3485 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3486 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3487 return OverflowResult::NeverOverflows;
3491 return OverflowResult::MayOverflow;
3494 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3495 const DataLayout &DL,
3496 AssumptionCache *AC,
3497 const Instruction *CxtI,
3498 const DominatorTree *DT) {
3499 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3500 Add, DL, AC, CxtI, DT);
3503 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3504 const DataLayout &DL,
3505 AssumptionCache *AC,
3506 const Instruction *CxtI,
3507 const DominatorTree *DT) {
3508 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3511 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3512 // FIXME: This conservative implementation can be relaxed. E.g. most
3513 // atomic operations are guaranteed to terminate on most platforms
3514 // and most functions terminate.
3516 return !I->isAtomic() && // atomics may never succeed on some platforms
3517 !isa<CallInst>(I) && // could throw and might not terminate
3518 !isa<InvokeInst>(I) && // might not terminate and could throw to
3519 // non-successor (see bug 24185 for details).
3520 !isa<ResumeInst>(I) && // has no successors
3521 !isa<ReturnInst>(I); // has no successors
3524 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3526 // The loop header is guaranteed to be executed for every iteration.
3528 // FIXME: Relax this constraint to cover all basic blocks that are
3529 // guaranteed to be executed at every iteration.
3530 if (I->getParent() != L->getHeader()) return false;
3532 for (const Instruction &LI : *L->getHeader()) {
3533 if (&LI == I) return true;
3534 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3536 llvm_unreachable("Instruction not contained in its own parent basic block.");
3539 bool llvm::propagatesFullPoison(const Instruction *I) {
3540 switch (I->getOpcode()) {
3541 case Instruction::Add:
3542 case Instruction::Sub:
3543 case Instruction::Xor:
3544 case Instruction::Trunc:
3545 case Instruction::BitCast:
3546 case Instruction::AddrSpaceCast:
3547 // These operations all propagate poison unconditionally. Note that poison
3548 // is not any particular value, so xor or subtraction of poison with
3549 // itself still yields poison, not zero.
3552 case Instruction::AShr:
3553 case Instruction::SExt:
3554 // For these operations, one bit of the input is replicated across
3555 // multiple output bits. A replicated poison bit is still poison.
3558 case Instruction::Shl: {
3559 // Left shift *by* a poison value is poison. The number of
3560 // positions to shift is unsigned, so no negative values are
3561 // possible there. Left shift by zero places preserves poison. So
3562 // it only remains to consider left shift of poison by a positive
3563 // number of places.
3565 // A left shift by a positive number of places leaves the lowest order bit
3566 // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3567 // make the poison operand violate that flag, yielding a fresh full-poison
3569 auto *OBO = cast<OverflowingBinaryOperator>(I);
3570 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3573 case Instruction::Mul: {
3574 // A multiplication by zero yields a non-poison zero result, so we need to
3575 // rule out zero as an operand. Conservatively, multiplication by a
3576 // non-zero constant is not multiplication by zero.
3578 // Multiplication by a non-zero constant can leave some bits
3579 // non-poisoned. For example, a multiplication by 2 leaves the lowest
3580 // order bit unpoisoned. So we need to consider that.
3582 // Multiplication by 1 preserves poison. If the multiplication has a
3583 // no-wrap flag, then we can make the poison operand violate that flag
3584 // when multiplied by any integer other than 0 and 1.
3585 auto *OBO = cast<OverflowingBinaryOperator>(I);
3586 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3587 for (Value *V : OBO->operands()) {
3588 if (auto *CI = dyn_cast<ConstantInt>(V)) {
3589 // A ConstantInt cannot yield poison, so we can assume that it is
3590 // the other operand that is poison.
3591 return !CI->isZero();
3598 case Instruction::GetElementPtr:
3599 // A GEP implicitly represents a sequence of additions, subtractions,
3600 // truncations, sign extensions and multiplications. The multiplications
3601 // are by the non-zero sizes of some set of types, so we do not have to be
3602 // concerned with multiplication by zero. If the GEP is in-bounds, then
3603 // these operations are implicitly no-signed-wrap so poison is propagated
3604 // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3605 return cast<GEPOperator>(I)->isInBounds();
3612 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3613 switch (I->getOpcode()) {
3614 case Instruction::Store:
3615 return cast<StoreInst>(I)->getPointerOperand();
3617 case Instruction::Load:
3618 return cast<LoadInst>(I)->getPointerOperand();
3620 case Instruction::AtomicCmpXchg:
3621 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3623 case Instruction::AtomicRMW:
3624 return cast<AtomicRMWInst>(I)->getPointerOperand();
3626 case Instruction::UDiv:
3627 case Instruction::SDiv:
3628 case Instruction::URem:
3629 case Instruction::SRem:
3630 return I->getOperand(1);
3637 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3638 // We currently only look for uses of poison values within the same basic
3639 // block, as that makes it easier to guarantee that the uses will be
3640 // executed given that PoisonI is executed.
3642 // FIXME: Expand this to consider uses beyond the same basic block. To do
3643 // this, look out for the distinction between post-dominance and strong
3645 const BasicBlock *BB = PoisonI->getParent();
3647 // Set of instructions that we have proved will yield poison if PoisonI
3649 SmallSet<const Value *, 16> YieldsPoison;
3650 YieldsPoison.insert(PoisonI);
3652 for (BasicBlock::const_iterator I = PoisonI->getIterator(), E = BB->end();
3654 if (&*I != PoisonI) {
3655 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&*I);
3656 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) return true;
3657 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
3661 // Mark poison that propagates from I through uses of I.
3662 if (YieldsPoison.count(&*I)) {
3663 for (const User *User : I->users()) {
3664 const Instruction *UserI = cast<Instruction>(User);
3665 if (UserI->getParent() == BB && propagatesFullPoison(UserI))
3666 YieldsPoison.insert(User);
3673 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3677 if (auto *C = dyn_cast<ConstantFP>(V))
3682 static bool isKnownNonZero(Value *V) {
3683 if (auto *C = dyn_cast<ConstantFP>(V))
3684 return !C->isZero();
3688 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3690 Value *CmpLHS, Value *CmpRHS,
3691 Value *TrueVal, Value *FalseVal,
3692 Value *&LHS, Value *&RHS) {
3696 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3697 // return inconsistent results between implementations.
3698 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3699 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3700 // Therefore we behave conservatively and only proceed if at least one of the
3701 // operands is known to not be zero, or if we don't care about signed zeroes.
3704 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3705 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3706 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3707 !isKnownNonZero(CmpRHS))
3708 return {SPF_UNKNOWN, SPNB_NA, false};
3711 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3712 bool Ordered = false;
3714 // When given one NaN and one non-NaN input:
3715 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3716 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3717 // ordered comparison fails), which could be NaN or non-NaN.
3718 // so here we discover exactly what NaN behavior is required/accepted.
3719 if (CmpInst::isFPPredicate(Pred)) {
3720 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3721 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3723 if (LHSSafe && RHSSafe) {
3724 // Both operands are known non-NaN.
3725 NaNBehavior = SPNB_RETURNS_ANY;
3726 } else if (CmpInst::isOrdered(Pred)) {
3727 // An ordered comparison will return false when given a NaN, so it
3731 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3732 NaNBehavior = SPNB_RETURNS_NAN;
3734 NaNBehavior = SPNB_RETURNS_OTHER;
3736 // Completely unsafe.
3737 return {SPF_UNKNOWN, SPNB_NA, false};
3740 // An unordered comparison will return true when given a NaN, so it
3743 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3744 NaNBehavior = SPNB_RETURNS_OTHER;
3746 NaNBehavior = SPNB_RETURNS_NAN;
3748 // Completely unsafe.
3749 return {SPF_UNKNOWN, SPNB_NA, false};
3753 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3754 std::swap(CmpLHS, CmpRHS);
3755 Pred = CmpInst::getSwappedPredicate(Pred);
3756 if (NaNBehavior == SPNB_RETURNS_NAN)
3757 NaNBehavior = SPNB_RETURNS_OTHER;
3758 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3759 NaNBehavior = SPNB_RETURNS_NAN;
3763 // ([if]cmp X, Y) ? X : Y
3764 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3766 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3767 case ICmpInst::ICMP_UGT:
3768 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3769 case ICmpInst::ICMP_SGT:
3770 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3771 case ICmpInst::ICMP_ULT:
3772 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3773 case ICmpInst::ICMP_SLT:
3774 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3775 case FCmpInst::FCMP_UGT:
3776 case FCmpInst::FCMP_UGE:
3777 case FCmpInst::FCMP_OGT:
3778 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3779 case FCmpInst::FCMP_ULT:
3780 case FCmpInst::FCMP_ULE:
3781 case FCmpInst::FCMP_OLT:
3782 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3786 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3787 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3788 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3790 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3791 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3792 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3793 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3796 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3797 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3798 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3799 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3803 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3804 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3805 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3806 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3807 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3810 return {SPF_SMIN, SPNB_NA, false};
3815 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3817 return {SPF_UNKNOWN, SPNB_NA, false};
3820 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
3821 Instruction::CastOps *CastOp) {
3822 CastInst *CI = dyn_cast<CastInst>(V1);
3823 Constant *C = dyn_cast<Constant>(V2);
3824 CastInst *CI2 = dyn_cast<CastInst>(V2);
3827 *CastOp = CI->getOpcode();
3830 // If V1 and V2 are both the same cast from the same type, we can look
3832 if (CI2->getOpcode() == CI->getOpcode() &&
3833 CI2->getSrcTy() == CI->getSrcTy())
3834 return CI2->getOperand(0);
3840 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3841 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3842 // This is only valid if the truncated value can be sign-extended
3843 // back to the original value.
3844 if (ConstantExpr::getSExt(T, C->getType()) == C)
3848 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3849 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3851 if (isa<TruncInst>(CI))
3852 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3854 if (isa<FPToUIInst>(CI))
3855 return ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
3857 if (isa<FPToSIInst>(CI))
3858 return ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
3860 if (isa<UIToFPInst>(CI))
3861 return ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
3863 if (isa<SIToFPInst>(CI))
3864 return ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
3866 if (isa<FPTruncInst>(CI))
3867 return ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
3869 if (isa<FPExtInst>(CI))
3870 return ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
3875 SelectPatternResult llvm::matchSelectPattern(Value *V,
3876 Value *&LHS, Value *&RHS,
3877 Instruction::CastOps *CastOp) {
3878 SelectInst *SI = dyn_cast<SelectInst>(V);
3879 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
3881 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
3882 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
3884 CmpInst::Predicate Pred = CmpI->getPredicate();
3885 Value *CmpLHS = CmpI->getOperand(0);
3886 Value *CmpRHS = CmpI->getOperand(1);
3887 Value *TrueVal = SI->getTrueValue();
3888 Value *FalseVal = SI->getFalseValue();
3890 if (isa<FPMathOperator>(CmpI))
3891 FMF = CmpI->getFastMathFlags();
3894 if (CmpI->isEquality())
3895 return {SPF_UNKNOWN, SPNB_NA, false};
3897 // Deal with type mismatches.
3898 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
3899 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
3900 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3901 cast<CastInst>(TrueVal)->getOperand(0), C,
3903 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
3904 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
3905 C, cast<CastInst>(FalseVal)->getOperand(0),
3908 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,