1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/Analysis/AssumptionCache.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/MemoryBuiltins.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/ConstantRange.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/GetElementPtrTypeIterator.h"
27 #include "llvm/IR/GlobalAlias.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/Metadata.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/Statepoint.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/MathExtras.h"
40 using namespace llvm::PatternMatch;
42 const unsigned MaxDepth = 6;
44 /// Enable an experimental feature to leverage information about dominating
45 /// conditions to compute known bits. The individual options below control how
46 /// hard we search. The defaults are chosen to be fairly aggressive. If you
47 /// run into compile time problems when testing, scale them back and report
49 static cl::opt<bool> EnableDomConditions("value-tracking-dom-conditions",
50 cl::Hidden, cl::init(false));
52 // This is expensive, so we only do it for the top level query value.
53 // (TODO: evaluate cost vs profit, consider higher thresholds)
54 static cl::opt<unsigned> DomConditionsMaxDepth("dom-conditions-max-depth",
55 cl::Hidden, cl::init(1));
57 /// How many dominating blocks should be scanned looking for dominating
59 static cl::opt<unsigned> DomConditionsMaxDomBlocks("dom-conditions-dom-blocks",
63 // Controls the number of uses of the value searched for possible
64 // dominating comparisons.
65 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
66 cl::Hidden, cl::init(20));
68 // If true, don't consider only compares whose only use is a branch.
69 static cl::opt<bool> DomConditionsSingleCmpUse("dom-conditions-single-cmp-use",
70 cl::Hidden, cl::init(false));
72 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
73 /// 0). For vector types, returns the element type's bitwidth.
74 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
75 if (unsigned BitWidth = Ty->getScalarSizeInBits())
78 return DL.getPointerTypeSizeInBits(Ty);
81 // Many of these functions have internal versions that take an assumption
82 // exclusion set. This is because of the potential for mutual recursion to
83 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
84 // classic case of this is assume(x = y), which will attempt to determine
85 // bits in x from bits in y, which will attempt to determine bits in y from
86 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
87 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
88 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
89 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
92 // Simplifying using an assume can only be done in a particular control-flow
93 // context (the context instruction provides that context). If an assume and
94 // the context instruction are not in the same block then the DT helps in
95 // figuring out if we can use it.
99 const Instruction *CxtI;
100 const DominatorTree *DT;
102 Query(AssumptionCache *AC = nullptr, const Instruction *CxtI = nullptr,
103 const DominatorTree *DT = nullptr)
104 : AC(AC), CxtI(CxtI), DT(DT) {}
106 Query(const Query &Q, const Value *NewExcl)
107 : ExclInvs(Q.ExclInvs), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT) {
108 ExclInvs.insert(NewExcl);
111 } // end anonymous namespace
113 // Given the provided Value and, potentially, a context instruction, return
114 // the preferred context instruction (if any).
115 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
116 // If we've been provided with a context instruction, then use that (provided
117 // it has been inserted).
118 if (CxtI && CxtI->getParent())
121 // If the value is really an already-inserted instruction, then use that.
122 CxtI = dyn_cast<Instruction>(V);
123 if (CxtI && CxtI->getParent())
129 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
130 const DataLayout &DL, unsigned Depth,
133 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
134 const DataLayout &DL, unsigned Depth,
135 AssumptionCache *AC, const Instruction *CxtI,
136 const DominatorTree *DT) {
137 ::computeKnownBits(V, KnownZero, KnownOne, DL, Depth,
138 Query(AC, safeCxtI(V, CxtI), DT));
141 bool llvm::haveNoCommonBitsSet(Value *LHS, Value *RHS, const DataLayout &DL,
142 AssumptionCache *AC, const Instruction *CxtI,
143 const DominatorTree *DT) {
144 assert(LHS->getType() == RHS->getType() &&
145 "LHS and RHS should have the same type");
146 assert(LHS->getType()->isIntOrIntVectorTy() &&
147 "LHS and RHS should be integers");
148 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
149 APInt LHSKnownZero(IT->getBitWidth(), 0), LHSKnownOne(IT->getBitWidth(), 0);
150 APInt RHSKnownZero(IT->getBitWidth(), 0), RHSKnownOne(IT->getBitWidth(), 0);
151 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, 0, AC, CxtI, DT);
152 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, 0, AC, CxtI, DT);
153 return (LHSKnownZero | RHSKnownZero).isAllOnesValue();
156 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
157 const DataLayout &DL, unsigned Depth,
160 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
161 const DataLayout &DL, unsigned Depth,
162 AssumptionCache *AC, const Instruction *CxtI,
163 const DominatorTree *DT) {
164 ::ComputeSignBit(V, KnownZero, KnownOne, DL, Depth,
165 Query(AC, safeCxtI(V, CxtI), DT));
168 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
169 const Query &Q, const DataLayout &DL);
171 bool llvm::isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, bool OrZero,
172 unsigned Depth, AssumptionCache *AC,
173 const Instruction *CxtI,
174 const DominatorTree *DT) {
175 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
176 Query(AC, safeCxtI(V, CxtI), DT), DL);
179 static bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
182 bool llvm::isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
183 AssumptionCache *AC, const Instruction *CxtI,
184 const DominatorTree *DT) {
185 return ::isKnownNonZero(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
188 bool llvm::isKnownNonNegative(Value *V, const DataLayout &DL, unsigned Depth,
189 AssumptionCache *AC, const Instruction *CxtI,
190 const DominatorTree *DT) {
191 bool NonNegative, Negative;
192 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
196 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
199 bool llvm::isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
200 AssumptionCache *AC, const Instruction *CxtI,
201 const DominatorTree *DT) {
202 return ::isKnownNonEqual(V1, V2, DL, Query(AC,
203 safeCxtI(V1, safeCxtI(V2, CxtI)),
207 static bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
208 unsigned Depth, const Query &Q);
210 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
211 unsigned Depth, AssumptionCache *AC,
212 const Instruction *CxtI, const DominatorTree *DT) {
213 return ::MaskedValueIsZero(V, Mask, DL, Depth,
214 Query(AC, safeCxtI(V, CxtI), DT));
217 static unsigned ComputeNumSignBits(Value *V, const DataLayout &DL,
218 unsigned Depth, const Query &Q);
220 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL,
221 unsigned Depth, AssumptionCache *AC,
222 const Instruction *CxtI,
223 const DominatorTree *DT) {
224 return ::ComputeNumSignBits(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
227 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
228 APInt &KnownZero, APInt &KnownOne,
229 APInt &KnownZero2, APInt &KnownOne2,
230 const DataLayout &DL, unsigned Depth,
233 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
234 // We know that the top bits of C-X are clear if X contains less bits
235 // than C (i.e. no wrap-around can happen). For example, 20-X is
236 // positive if we can prove that X is >= 0 and < 16.
237 if (!CLHS->getValue().isNegative()) {
238 unsigned BitWidth = KnownZero.getBitWidth();
239 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
240 // NLZ can't be BitWidth with no sign bit
241 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
242 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
244 // If all of the MaskV bits are known to be zero, then we know the
245 // output top bits are zero, because we now know that the output is
247 if ((KnownZero2 & MaskV) == MaskV) {
248 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
249 // Top bits known zero.
250 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
256 unsigned BitWidth = KnownZero.getBitWidth();
258 // If an initial sequence of bits in the result is not needed, the
259 // corresponding bits in the operands are not needed.
260 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
261 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, DL, Depth + 1, Q);
262 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
264 // Carry in a 1 for a subtract, rather than a 0.
265 APInt CarryIn(BitWidth, 0);
267 // Sum = LHS + ~RHS + 1
268 std::swap(KnownZero2, KnownOne2);
272 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
273 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
275 // Compute known bits of the carry.
276 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
277 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
279 // Compute set of known bits (where all three relevant bits are known).
280 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
281 APInt RHSKnown = KnownZero2 | KnownOne2;
282 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
283 APInt Known = LHSKnown & RHSKnown & CarryKnown;
285 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
286 "known bits of sum differ");
288 // Compute known bits of the result.
289 KnownZero = ~PossibleSumOne & Known;
290 KnownOne = PossibleSumOne & Known;
292 // Are we still trying to solve for the sign bit?
293 if (!Known.isNegative()) {
295 // Adding two non-negative numbers, or subtracting a negative number from
296 // a non-negative one, can't wrap into negative.
297 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
298 KnownZero |= APInt::getSignBit(BitWidth);
299 // Adding two negative numbers, or subtracting a non-negative number from
300 // a negative one, can't wrap into non-negative.
301 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
302 KnownOne |= APInt::getSignBit(BitWidth);
307 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
308 APInt &KnownZero, APInt &KnownOne,
309 APInt &KnownZero2, APInt &KnownOne2,
310 const DataLayout &DL, unsigned Depth,
312 unsigned BitWidth = KnownZero.getBitWidth();
313 computeKnownBits(Op1, KnownZero, KnownOne, DL, Depth + 1, Q);
314 computeKnownBits(Op0, KnownZero2, KnownOne2, DL, Depth + 1, Q);
316 bool isKnownNegative = false;
317 bool isKnownNonNegative = false;
318 // If the multiplication is known not to overflow, compute the sign bit.
321 // The product of a number with itself is non-negative.
322 isKnownNonNegative = true;
324 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
325 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
326 bool isKnownNegativeOp1 = KnownOne.isNegative();
327 bool isKnownNegativeOp0 = KnownOne2.isNegative();
328 // The product of two numbers with the same sign is non-negative.
329 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
330 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
331 // The product of a negative number and a non-negative number is either
333 if (!isKnownNonNegative)
334 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
335 isKnownNonZero(Op0, DL, Depth, Q)) ||
336 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
337 isKnownNonZero(Op1, DL, Depth, Q));
341 // If low bits are zero in either operand, output low known-0 bits.
342 // Also compute a conservative estimate for high known-0 bits.
343 // More trickiness is possible, but this is sufficient for the
344 // interesting case of alignment computation.
345 KnownOne.clearAllBits();
346 unsigned TrailZ = KnownZero.countTrailingOnes() +
347 KnownZero2.countTrailingOnes();
348 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
349 KnownZero2.countLeadingOnes(),
350 BitWidth) - BitWidth;
352 TrailZ = std::min(TrailZ, BitWidth);
353 LeadZ = std::min(LeadZ, BitWidth);
354 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
355 APInt::getHighBitsSet(BitWidth, LeadZ);
357 // Only make use of no-wrap flags if we failed to compute the sign bit
358 // directly. This matters if the multiplication always overflows, in
359 // which case we prefer to follow the result of the direct computation,
360 // though as the program is invoking undefined behaviour we can choose
361 // whatever we like here.
362 if (isKnownNonNegative && !KnownOne.isNegative())
363 KnownZero.setBit(BitWidth - 1);
364 else if (isKnownNegative && !KnownZero.isNegative())
365 KnownOne.setBit(BitWidth - 1);
368 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
370 unsigned BitWidth = KnownZero.getBitWidth();
371 unsigned NumRanges = Ranges.getNumOperands() / 2;
372 assert(NumRanges >= 1);
374 // Use the high end of the ranges to find leading zeros.
375 unsigned MinLeadingZeros = BitWidth;
376 for (unsigned i = 0; i < NumRanges; ++i) {
378 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
380 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
381 ConstantRange Range(Lower->getValue(), Upper->getValue());
382 if (Range.isWrappedSet())
383 MinLeadingZeros = 0; // -1 has no zeros
384 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
385 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
388 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
391 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
392 SmallVector<const Value *, 16> WorkSet(1, I);
393 SmallPtrSet<const Value *, 32> Visited;
394 SmallPtrSet<const Value *, 16> EphValues;
396 // The instruction defining an assumption's condition itself is always
397 // considered ephemeral to that assumption (even if it has other
398 // non-ephemeral users). See r246696's test case for an example.
399 if (std::find(I->op_begin(), I->op_end(), E) != I->op_end())
402 while (!WorkSet.empty()) {
403 const Value *V = WorkSet.pop_back_val();
404 if (!Visited.insert(V).second)
407 // If all uses of this value are ephemeral, then so is this value.
408 bool FoundNEUse = false;
409 for (const User *I : V->users())
410 if (!EphValues.count(I)) {
420 if (const User *U = dyn_cast<User>(V))
421 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
423 if (isSafeToSpeculativelyExecute(*J))
424 WorkSet.push_back(*J);
432 // Is this an intrinsic that cannot be speculated but also cannot trap?
433 static bool isAssumeLikeIntrinsic(const Instruction *I) {
434 if (const CallInst *CI = dyn_cast<CallInst>(I))
435 if (Function *F = CI->getCalledFunction())
436 switch (F->getIntrinsicID()) {
438 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
439 case Intrinsic::assume:
440 case Intrinsic::dbg_declare:
441 case Intrinsic::dbg_value:
442 case Intrinsic::invariant_start:
443 case Intrinsic::invariant_end:
444 case Intrinsic::lifetime_start:
445 case Intrinsic::lifetime_end:
446 case Intrinsic::objectsize:
447 case Intrinsic::ptr_annotation:
448 case Intrinsic::var_annotation:
455 static bool isValidAssumeForContext(Value *V, const Query &Q) {
456 Instruction *Inv = cast<Instruction>(V);
458 // There are two restrictions on the use of an assume:
459 // 1. The assume must dominate the context (or the control flow must
460 // reach the assume whenever it reaches the context).
461 // 2. The context must not be in the assume's set of ephemeral values
462 // (otherwise we will use the assume to prove that the condition
463 // feeding the assume is trivially true, thus causing the removal of
467 if (Q.DT->dominates(Inv, Q.CxtI)) {
469 } else if (Inv->getParent() == Q.CxtI->getParent()) {
470 // The context comes first, but they're both in the same block. Make sure
471 // there is nothing in between that might interrupt the control flow.
472 for (BasicBlock::const_iterator I =
473 std::next(BasicBlock::const_iterator(Q.CxtI)),
474 IE(Inv); I != IE; ++I)
475 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
478 return !isEphemeralValueOf(Inv, Q.CxtI);
484 // When we don't have a DT, we do a limited search...
485 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
487 } else if (Inv->getParent() == Q.CxtI->getParent()) {
488 // Search forward from the assume until we reach the context (or the end
489 // of the block); the common case is that the assume will come first.
490 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
491 IE = Inv->getParent()->end(); I != IE; ++I)
495 // The context must come first...
496 for (BasicBlock::const_iterator I =
497 std::next(BasicBlock::const_iterator(Q.CxtI)),
498 IE(Inv); I != IE; ++I)
499 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
502 return !isEphemeralValueOf(Inv, Q.CxtI);
508 bool llvm::isValidAssumeForContext(const Instruction *I,
509 const Instruction *CxtI,
510 const DominatorTree *DT) {
511 return ::isValidAssumeForContext(const_cast<Instruction *>(I),
512 Query(nullptr, CxtI, DT));
515 template<typename LHS, typename RHS>
516 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
517 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
518 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
519 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
522 template<typename LHS, typename RHS>
523 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
524 BinaryOp_match<RHS, LHS, Instruction::And>>
525 m_c_And(const LHS &L, const RHS &R) {
526 return m_CombineOr(m_And(L, R), m_And(R, L));
529 template<typename LHS, typename RHS>
530 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
531 BinaryOp_match<RHS, LHS, Instruction::Or>>
532 m_c_Or(const LHS &L, const RHS &R) {
533 return m_CombineOr(m_Or(L, R), m_Or(R, L));
536 template<typename LHS, typename RHS>
537 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
538 BinaryOp_match<RHS, LHS, Instruction::Xor>>
539 m_c_Xor(const LHS &L, const RHS &R) {
540 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
543 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
544 /// true (at the context instruction.) This is mostly a utility function for
545 /// the prototype dominating conditions reasoning below.
546 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
549 const DataLayout &DL,
550 unsigned Depth, const Query &Q) {
551 Value *LHS = Cmp->getOperand(0);
552 Value *RHS = Cmp->getOperand(1);
553 // TODO: We could potentially be more aggressive here. This would be worth
554 // evaluating. If we can, explore commoning this code with the assume
556 if (LHS != V && RHS != V)
559 const unsigned BitWidth = KnownZero.getBitWidth();
561 switch (Cmp->getPredicate()) {
563 // We know nothing from this condition
565 // TODO: implement unsigned bound from below (known one bits)
566 // TODO: common condition check implementations with assumes
567 // TODO: implement other patterns from assume (e.g. V & B == A)
568 case ICmpInst::ICMP_SGT:
570 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
571 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
572 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
573 // We know that the sign bit is zero.
574 KnownZero |= APInt::getSignBit(BitWidth);
578 case ICmpInst::ICMP_EQ:
580 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
582 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
584 computeKnownBits(LHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
586 llvm_unreachable("missing use?");
587 KnownZero |= KnownZeroTemp;
588 KnownOne |= KnownOneTemp;
591 case ICmpInst::ICMP_ULE:
593 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
594 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
595 // The known zero bits carry over
596 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
597 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
600 case ICmpInst::ICMP_ULT:
602 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
603 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
604 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
605 // power of 2, then one more).
606 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
607 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
609 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
615 /// Compute known bits in 'V' from conditions which are known to be true along
616 /// all paths leading to the context instruction. In particular, look for
617 /// cases where one branch of an interesting condition dominates the context
618 /// instruction. This does not do general dataflow.
619 /// NOTE: This code is EXPERIMENTAL and currently off by default.
620 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
622 const DataLayout &DL,
625 // Need both the dominator tree and the query location to do anything useful
626 if (!Q.DT || !Q.CxtI)
628 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
629 // The context instruction might be in a statically unreachable block. If
630 // so, asking dominator queries may yield suprising results. (e.g. the block
631 // may not have a dom tree node)
632 if (!Q.DT->isReachableFromEntry(Cxt->getParent()))
635 // Avoid useless work
636 if (auto VI = dyn_cast<Instruction>(V))
637 if (VI->getParent() == Cxt->getParent())
640 // Note: We currently implement two options. It's not clear which of these
641 // will survive long term, we need data for that.
642 // Option 1 - Try walking the dominator tree looking for conditions which
643 // might apply. This works well for local conditions (loop guards, etc..),
644 // but not as well for things far from the context instruction (presuming a
645 // low max blocks explored). If we can set an high enough limit, this would
647 // Option 2 - We restrict out search to those conditions which are uses of
648 // the value we're interested in. This is independent of dom structure,
649 // but is slightly less powerful without looking through lots of use chains.
650 // It does handle conditions far from the context instruction (e.g. early
651 // function exits on entry) really well though.
653 // Option 1 - Search the dom tree
654 unsigned NumBlocksExplored = 0;
655 BasicBlock *Current = Cxt->getParent();
657 // Stop searching if we've gone too far up the chain
658 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
662 if (!Q.DT->getNode(Current)->getIDom())
664 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
666 // found function entry
669 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
670 if (!BI || BI->isUnconditional())
672 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
676 // We're looking for conditions that are guaranteed to hold at the context
677 // instruction. Finding a condition where one path dominates the context
678 // isn't enough because both the true and false cases could merge before
679 // the context instruction we're actually interested in. Instead, we need
680 // to ensure that the taken *edge* dominates the context instruction. We
681 // know that the edge must be reachable since we started from a reachable
683 BasicBlock *BB0 = BI->getSuccessor(0);
684 BasicBlockEdge Edge(BI->getParent(), BB0);
685 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
688 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
692 // Option 2 - Search the other uses of V
693 unsigned NumUsesExplored = 0;
694 for (auto U : V->users()) {
695 // Avoid massive lists
696 if (NumUsesExplored >= DomConditionsMaxUses)
699 // Consider only compare instructions uniquely controlling a branch
700 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
704 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
707 for (auto *CmpU : Cmp->users()) {
708 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
709 if (!BI || BI->isUnconditional())
711 // We're looking for conditions that are guaranteed to hold at the
712 // context instruction. Finding a condition where one path dominates
713 // the context isn't enough because both the true and false cases could
714 // merge before the context instruction we're actually interested in.
715 // Instead, we need to ensure that the taken *edge* dominates the context
717 BasicBlock *BB0 = BI->getSuccessor(0);
718 BasicBlockEdge Edge(BI->getParent(), BB0);
719 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
722 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
728 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
729 APInt &KnownOne, const DataLayout &DL,
730 unsigned Depth, const Query &Q) {
731 // Use of assumptions is context-sensitive. If we don't have a context, we
733 if (!Q.AC || !Q.CxtI)
736 unsigned BitWidth = KnownZero.getBitWidth();
738 for (auto &AssumeVH : Q.AC->assumptions()) {
741 CallInst *I = cast<CallInst>(AssumeVH);
742 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
743 "Got assumption for the wrong function!");
744 if (Q.ExclInvs.count(I))
747 // Warning: This loop can end up being somewhat performance sensetive.
748 // We're running this loop for once for each value queried resulting in a
749 // runtime of ~O(#assumes * #values).
751 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
752 "must be an assume intrinsic");
754 Value *Arg = I->getArgOperand(0);
756 if (Arg == V && isValidAssumeForContext(I, Q)) {
757 assert(BitWidth == 1 && "assume operand is not i1?");
758 KnownZero.clearAllBits();
759 KnownOne.setAllBits();
763 // The remaining tests are all recursive, so bail out if we hit the limit.
764 if (Depth == MaxDepth)
768 auto m_V = m_CombineOr(m_Specific(V),
769 m_CombineOr(m_PtrToInt(m_Specific(V)),
770 m_BitCast(m_Specific(V))));
772 CmpInst::Predicate Pred;
775 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
776 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
777 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
778 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
779 KnownZero |= RHSKnownZero;
780 KnownOne |= RHSKnownOne;
782 } else if (match(Arg,
783 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
784 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
785 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
786 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
787 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
788 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
790 // For those bits in the mask that are known to be one, we can propagate
791 // known bits from the RHS to V.
792 KnownZero |= RHSKnownZero & MaskKnownOne;
793 KnownOne |= RHSKnownOne & MaskKnownOne;
794 // assume(~(v & b) = a)
795 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
797 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
798 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
799 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
800 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
801 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
803 // For those bits in the mask that are known to be one, we can propagate
804 // inverted known bits from the RHS to V.
805 KnownZero |= RHSKnownOne & MaskKnownOne;
806 KnownOne |= RHSKnownZero & MaskKnownOne;
808 } else if (match(Arg,
809 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
810 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
811 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
812 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
813 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
814 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
816 // For those bits in B that are known to be zero, we can propagate known
817 // bits from the RHS to V.
818 KnownZero |= RHSKnownZero & BKnownZero;
819 KnownOne |= RHSKnownOne & BKnownZero;
820 // assume(~(v | b) = a)
821 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
823 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
824 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
825 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
826 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
827 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
829 // For those bits in B that are known to be zero, we can propagate
830 // inverted known bits from the RHS to V.
831 KnownZero |= RHSKnownOne & BKnownZero;
832 KnownOne |= RHSKnownZero & BKnownZero;
834 } else if (match(Arg,
835 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
836 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
837 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
838 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
839 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
840 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
842 // For those bits in B that are known to be zero, we can propagate known
843 // bits from the RHS to V. For those bits in B that are known to be one,
844 // we can propagate inverted known bits from the RHS to V.
845 KnownZero |= RHSKnownZero & BKnownZero;
846 KnownOne |= RHSKnownOne & BKnownZero;
847 KnownZero |= RHSKnownOne & BKnownOne;
848 KnownOne |= RHSKnownZero & BKnownOne;
849 // assume(~(v ^ b) = a)
850 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
852 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
853 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
854 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
855 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
856 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
858 // For those bits in B that are known to be zero, we can propagate
859 // inverted known bits from the RHS to V. For those bits in B that are
860 // known to be one, we can propagate known bits from the RHS to V.
861 KnownZero |= RHSKnownOne & BKnownZero;
862 KnownOne |= RHSKnownZero & BKnownZero;
863 KnownZero |= RHSKnownZero & BKnownOne;
864 KnownOne |= RHSKnownOne & BKnownOne;
865 // assume(v << c = a)
866 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
868 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
869 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
870 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
871 // For those bits in RHS that are known, we can propagate them to known
872 // bits in V shifted to the right by C.
873 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
874 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
875 // assume(~(v << c) = a)
876 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
878 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
879 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
880 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
881 // For those bits in RHS that are known, we can propagate them inverted
882 // to known bits in V shifted to the right by C.
883 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
884 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
885 // assume(v >> c = a)
886 } else if (match(Arg,
887 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
888 m_AShr(m_V, m_ConstantInt(C))),
890 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
891 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
892 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
893 // For those bits in RHS that are known, we can propagate them to known
894 // bits in V shifted to the right by C.
895 KnownZero |= RHSKnownZero << C->getZExtValue();
896 KnownOne |= RHSKnownOne << C->getZExtValue();
897 // assume(~(v >> c) = a)
898 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
899 m_LShr(m_V, m_ConstantInt(C)),
900 m_AShr(m_V, m_ConstantInt(C)))),
902 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
903 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
904 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
905 // For those bits in RHS that are known, we can propagate them inverted
906 // to known bits in V shifted to the right by C.
907 KnownZero |= RHSKnownOne << C->getZExtValue();
908 KnownOne |= RHSKnownZero << C->getZExtValue();
909 // assume(v >=_s c) where c is non-negative
910 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
911 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
912 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
913 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
915 if (RHSKnownZero.isNegative()) {
916 // We know that the sign bit is zero.
917 KnownZero |= APInt::getSignBit(BitWidth);
919 // assume(v >_s c) where c is at least -1.
920 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
921 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
922 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
923 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
925 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
926 // We know that the sign bit is zero.
927 KnownZero |= APInt::getSignBit(BitWidth);
929 // assume(v <=_s c) where c is negative
930 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
931 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
932 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
933 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
935 if (RHSKnownOne.isNegative()) {
936 // We know that the sign bit is one.
937 KnownOne |= APInt::getSignBit(BitWidth);
939 // assume(v <_s c) where c is non-positive
940 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
941 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
942 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
943 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
945 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
946 // We know that the sign bit is one.
947 KnownOne |= APInt::getSignBit(BitWidth);
950 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
951 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
952 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
953 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
955 // Whatever high bits in c are zero are known to be zero.
957 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
959 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
960 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
961 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
962 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
964 // Whatever high bits in c are zero are known to be zero (if c is a power
965 // of 2, then one more).
966 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
968 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
971 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
976 // Compute known bits from a shift operator, including those with a
977 // non-constant shift amount. KnownZero and KnownOne are the outputs of this
978 // function. KnownZero2 and KnownOne2 are pre-allocated temporaries with the
979 // same bit width as KnownZero and KnownOne. KZF and KOF are operator-specific
980 // functors that, given the known-zero or known-one bits respectively, and a
981 // shift amount, compute the implied known-zero or known-one bits of the shift
982 // operator's result respectively for that shift amount. The results from calling
983 // KZF and KOF are conservatively combined for all permitted shift amounts.
984 template <typename KZFunctor, typename KOFunctor>
985 static void computeKnownBitsFromShiftOperator(Operator *I,
986 APInt &KnownZero, APInt &KnownOne,
987 APInt &KnownZero2, APInt &KnownOne2,
988 const DataLayout &DL, unsigned Depth, const Query &Q,
989 KZFunctor KZF, KOFunctor KOF) {
990 unsigned BitWidth = KnownZero.getBitWidth();
992 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
993 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
995 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
996 KnownZero = KZF(KnownZero, ShiftAmt);
997 KnownOne = KOF(KnownOne, ShiftAmt);
1001 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1003 // Note: We cannot use KnownZero.getLimitedValue() here, because if
1004 // BitWidth > 64 and any upper bits are known, we'll end up returning the
1005 // limit value (which implies all bits are known).
1006 uint64_t ShiftAmtKZ = KnownZero.zextOrTrunc(64).getZExtValue();
1007 uint64_t ShiftAmtKO = KnownOne.zextOrTrunc(64).getZExtValue();
1009 // It would be more-clearly correct to use the two temporaries for this
1010 // calculation. Reusing the APInts here to prevent unnecessary allocations.
1011 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1013 // Early exit if we can't constrain any well-defined shift amount.
1014 if (!(ShiftAmtKZ & (BitWidth-1)) && !(ShiftAmtKO & (BitWidth-1)))
1017 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1019 KnownZero = KnownOne = APInt::getAllOnesValue(BitWidth);
1020 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
1021 // Combine the shifted known input bits only for those shift amounts
1022 // compatible with its known constraints.
1023 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
1025 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
1028 KnownZero &= KZF(KnownZero2, ShiftAmt);
1029 KnownOne &= KOF(KnownOne2, ShiftAmt);
1032 // If there are no compatible shift amounts, then we've proven that the shift
1033 // amount must be >= the BitWidth, and the result is undefined. We could
1034 // return anything we'd like, but we need to make sure the sets of known bits
1035 // stay disjoint (it should be better for some other code to actually
1036 // propagate the undef than to pick a value here using known bits).
1037 if ((KnownZero & KnownOne) != 0)
1038 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1041 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
1042 APInt &KnownOne, const DataLayout &DL,
1043 unsigned Depth, const Query &Q) {
1044 unsigned BitWidth = KnownZero.getBitWidth();
1046 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
1047 switch (I->getOpcode()) {
1049 case Instruction::Load:
1050 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
1051 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1053 case Instruction::And: {
1054 // If either the LHS or the RHS are Zero, the result is zero.
1055 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1056 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1058 // Output known-1 bits are only known if set in both the LHS & RHS.
1059 KnownOne &= KnownOne2;
1060 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1061 KnownZero |= KnownZero2;
1064 case Instruction::Or: {
1065 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1066 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1068 // Output known-0 bits are only known if clear in both the LHS & RHS.
1069 KnownZero &= KnownZero2;
1070 // Output known-1 are known to be set if set in either the LHS | RHS.
1071 KnownOne |= KnownOne2;
1074 case Instruction::Xor: {
1075 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1076 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1078 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1079 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1080 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1081 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1082 KnownZero = KnownZeroOut;
1085 case Instruction::Mul: {
1086 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1087 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1088 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1091 case Instruction::UDiv: {
1092 // For the purposes of computing leading zeros we can conservatively
1093 // treat a udiv as a logical right shift by the power of 2 known to
1094 // be less than the denominator.
1095 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1096 unsigned LeadZ = KnownZero2.countLeadingOnes();
1098 KnownOne2.clearAllBits();
1099 KnownZero2.clearAllBits();
1100 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1101 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1102 if (RHSUnknownLeadingOnes != BitWidth)
1103 LeadZ = std::min(BitWidth,
1104 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1106 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1109 case Instruction::Select:
1110 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1111 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1113 // Only known if known in both the LHS and RHS.
1114 KnownOne &= KnownOne2;
1115 KnownZero &= KnownZero2;
1117 case Instruction::FPTrunc:
1118 case Instruction::FPExt:
1119 case Instruction::FPToUI:
1120 case Instruction::FPToSI:
1121 case Instruction::SIToFP:
1122 case Instruction::UIToFP:
1123 break; // Can't work with floating point.
1124 case Instruction::PtrToInt:
1125 case Instruction::IntToPtr:
1126 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1127 // FALL THROUGH and handle them the same as zext/trunc.
1128 case Instruction::ZExt:
1129 case Instruction::Trunc: {
1130 Type *SrcTy = I->getOperand(0)->getType();
1132 unsigned SrcBitWidth;
1133 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1134 // which fall through here.
1135 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1137 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1138 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1139 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1140 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1141 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1142 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1143 // Any top bits are known to be zero.
1144 if (BitWidth > SrcBitWidth)
1145 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1148 case Instruction::BitCast: {
1149 Type *SrcTy = I->getOperand(0)->getType();
1150 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy() ||
1151 SrcTy->isFloatingPointTy()) &&
1152 // TODO: For now, not handling conversions like:
1153 // (bitcast i64 %x to <2 x i32>)
1154 !I->getType()->isVectorTy()) {
1155 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1160 case Instruction::SExt: {
1161 // Compute the bits in the result that are not present in the input.
1162 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1164 KnownZero = KnownZero.trunc(SrcBitWidth);
1165 KnownOne = KnownOne.trunc(SrcBitWidth);
1166 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1167 KnownZero = KnownZero.zext(BitWidth);
1168 KnownOne = KnownOne.zext(BitWidth);
1170 // If the sign bit of the input is known set or clear, then we know the
1171 // top bits of the result.
1172 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1173 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1174 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1175 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1178 case Instruction::Shl: {
1179 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1180 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1181 return (KnownZero << ShiftAmt) |
1182 APInt::getLowBitsSet(BitWidth, ShiftAmt); // Low bits known 0.
1185 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1186 return KnownOne << ShiftAmt;
1189 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1190 KnownZero2, KnownOne2, DL, Depth, Q,
1194 case Instruction::LShr: {
1195 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1196 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1197 return APIntOps::lshr(KnownZero, ShiftAmt) |
1198 // High bits known zero.
1199 APInt::getHighBitsSet(BitWidth, ShiftAmt);
1202 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1203 return APIntOps::lshr(KnownOne, ShiftAmt);
1206 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1207 KnownZero2, KnownOne2, DL, Depth, Q,
1211 case Instruction::AShr: {
1212 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1213 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1214 return APIntOps::ashr(KnownZero, ShiftAmt);
1217 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1218 return APIntOps::ashr(KnownOne, ShiftAmt);
1221 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1222 KnownZero2, KnownOne2, DL, Depth, Q,
1226 case Instruction::Sub: {
1227 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1228 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1229 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1233 case Instruction::Add: {
1234 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1235 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1236 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1240 case Instruction::SRem:
1241 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1242 APInt RA = Rem->getValue().abs();
1243 if (RA.isPowerOf2()) {
1244 APInt LowBits = RA - 1;
1245 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1248 // The low bits of the first operand are unchanged by the srem.
1249 KnownZero = KnownZero2 & LowBits;
1250 KnownOne = KnownOne2 & LowBits;
1252 // If the first operand is non-negative or has all low bits zero, then
1253 // the upper bits are all zero.
1254 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1255 KnownZero |= ~LowBits;
1257 // If the first operand is negative and not all low bits are zero, then
1258 // the upper bits are all one.
1259 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1260 KnownOne |= ~LowBits;
1262 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1266 // The sign bit is the LHS's sign bit, except when the result of the
1267 // remainder is zero.
1268 if (KnownZero.isNonNegative()) {
1269 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1270 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1272 // If it's known zero, our sign bit is also zero.
1273 if (LHSKnownZero.isNegative())
1274 KnownZero.setBit(BitWidth - 1);
1278 case Instruction::URem: {
1279 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1280 APInt RA = Rem->getValue();
1281 if (RA.isPowerOf2()) {
1282 APInt LowBits = (RA - 1);
1283 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1285 KnownZero |= ~LowBits;
1286 KnownOne &= LowBits;
1291 // Since the result is less than or equal to either operand, any leading
1292 // zero bits in either operand must also exist in the result.
1293 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1294 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1296 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1297 KnownZero2.countLeadingOnes());
1298 KnownOne.clearAllBits();
1299 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1303 case Instruction::Alloca: {
1304 AllocaInst *AI = cast<AllocaInst>(I);
1305 unsigned Align = AI->getAlignment();
1307 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1310 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1313 case Instruction::GetElementPtr: {
1314 // Analyze all of the subscripts of this getelementptr instruction
1315 // to determine if we can prove known low zero bits.
1316 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1317 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1319 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1321 gep_type_iterator GTI = gep_type_begin(I);
1322 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1323 Value *Index = I->getOperand(i);
1324 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1325 // Handle struct member offset arithmetic.
1327 // Handle case when index is vector zeroinitializer
1328 Constant *CIndex = cast<Constant>(Index);
1329 if (CIndex->isZeroValue())
1332 if (CIndex->getType()->isVectorTy())
1333 Index = CIndex->getSplatValue();
1335 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1336 const StructLayout *SL = DL.getStructLayout(STy);
1337 uint64_t Offset = SL->getElementOffset(Idx);
1338 TrailZ = std::min<unsigned>(TrailZ,
1339 countTrailingZeros(Offset));
1341 // Handle array index arithmetic.
1342 Type *IndexedTy = GTI.getIndexedType();
1343 if (!IndexedTy->isSized()) {
1347 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1348 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1349 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1350 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1352 TrailZ = std::min(TrailZ,
1353 unsigned(countTrailingZeros(TypeSize) +
1354 LocalKnownZero.countTrailingOnes()));
1358 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1361 case Instruction::PHI: {
1362 PHINode *P = cast<PHINode>(I);
1363 // Handle the case of a simple two-predecessor recurrence PHI.
1364 // There's a lot more that could theoretically be done here, but
1365 // this is sufficient to catch some interesting cases.
1366 if (P->getNumIncomingValues() == 2) {
1367 for (unsigned i = 0; i != 2; ++i) {
1368 Value *L = P->getIncomingValue(i);
1369 Value *R = P->getIncomingValue(!i);
1370 Operator *LU = dyn_cast<Operator>(L);
1373 unsigned Opcode = LU->getOpcode();
1374 // Check for operations that have the property that if
1375 // both their operands have low zero bits, the result
1376 // will have low zero bits.
1377 if (Opcode == Instruction::Add ||
1378 Opcode == Instruction::Sub ||
1379 Opcode == Instruction::And ||
1380 Opcode == Instruction::Or ||
1381 Opcode == Instruction::Mul) {
1382 Value *LL = LU->getOperand(0);
1383 Value *LR = LU->getOperand(1);
1384 // Find a recurrence.
1391 // Ok, we have a PHI of the form L op= R. Check for low
1393 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1395 // We need to take the minimum number of known bits
1396 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1397 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1399 KnownZero = APInt::getLowBitsSet(BitWidth,
1400 std::min(KnownZero2.countTrailingOnes(),
1401 KnownZero3.countTrailingOnes()));
1407 // Unreachable blocks may have zero-operand PHI nodes.
1408 if (P->getNumIncomingValues() == 0)
1411 // Otherwise take the unions of the known bit sets of the operands,
1412 // taking conservative care to avoid excessive recursion.
1413 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1414 // Skip if every incoming value references to ourself.
1415 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1418 KnownZero = APInt::getAllOnesValue(BitWidth);
1419 KnownOne = APInt::getAllOnesValue(BitWidth);
1420 for (Value *IncValue : P->incoming_values()) {
1421 // Skip direct self references.
1422 if (IncValue == P) continue;
1424 KnownZero2 = APInt(BitWidth, 0);
1425 KnownOne2 = APInt(BitWidth, 0);
1426 // Recurse, but cap the recursion to one level, because we don't
1427 // want to waste time spinning around in loops.
1428 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1430 KnownZero &= KnownZero2;
1431 KnownOne &= KnownOne2;
1432 // If all bits have been ruled out, there's no need to check
1434 if (!KnownZero && !KnownOne)
1440 case Instruction::Call:
1441 case Instruction::Invoke:
1442 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1443 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1444 // If a range metadata is attached to this IntrinsicInst, intersect the
1445 // explicit range specified by the metadata and the implicit range of
1447 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1448 switch (II->getIntrinsicID()) {
1450 case Intrinsic::bswap:
1451 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1453 KnownZero |= KnownZero2.byteSwap();
1454 KnownOne |= KnownOne2.byteSwap();
1456 case Intrinsic::ctlz:
1457 case Intrinsic::cttz: {
1458 unsigned LowBits = Log2_32(BitWidth)+1;
1459 // If this call is undefined for 0, the result will be less than 2^n.
1460 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1462 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1465 case Intrinsic::ctpop: {
1466 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1468 // We can bound the space the count needs. Also, bits known to be zero
1469 // can't contribute to the population.
1470 unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation();
1471 unsigned LeadingZeros =
1472 APInt(BitWidth, BitsPossiblySet).countLeadingZeros();
1473 assert(LeadingZeros <= BitWidth);
1474 KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros);
1475 KnownOne &= ~KnownZero;
1476 // TODO: we could bound KnownOne using the lower bound on the number
1477 // of bits which might be set provided by popcnt KnownOne2.
1480 case Intrinsic::fabs: {
1481 Type *Ty = II->getType();
1482 APInt SignBit = APInt::getSignBit(Ty->getScalarSizeInBits());
1483 KnownZero |= APInt::getSplat(Ty->getPrimitiveSizeInBits(), SignBit);
1486 case Intrinsic::x86_sse42_crc32_64_64:
1487 KnownZero |= APInt::getHighBitsSet(64, 32);
1492 case Instruction::ExtractValue:
1493 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1494 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1495 if (EVI->getNumIndices() != 1) break;
1496 if (EVI->getIndices()[0] == 0) {
1497 switch (II->getIntrinsicID()) {
1499 case Intrinsic::uadd_with_overflow:
1500 case Intrinsic::sadd_with_overflow:
1501 computeKnownBitsAddSub(true, II->getArgOperand(0),
1502 II->getArgOperand(1), false, KnownZero,
1503 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1505 case Intrinsic::usub_with_overflow:
1506 case Intrinsic::ssub_with_overflow:
1507 computeKnownBitsAddSub(false, II->getArgOperand(0),
1508 II->getArgOperand(1), false, KnownZero,
1509 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1511 case Intrinsic::umul_with_overflow:
1512 case Intrinsic::smul_with_overflow:
1513 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1514 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1523 static unsigned getAlignment(const Value *V, const DataLayout &DL) {
1525 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1526 Align = GO->getAlignment();
1528 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1529 Type *ObjectType = GVar->getType()->getElementType();
1530 if (ObjectType->isSized()) {
1531 // If the object is defined in the current Module, we'll be giving
1532 // it the preferred alignment. Otherwise, we have to assume that it
1533 // may only have the minimum ABI alignment.
1534 if (GVar->isStrongDefinitionForLinker())
1535 Align = DL.getPreferredAlignment(GVar);
1537 Align = DL.getABITypeAlignment(ObjectType);
1541 } else if (const Argument *A = dyn_cast<Argument>(V)) {
1542 Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1544 if (!Align && A->hasStructRetAttr()) {
1545 // An sret parameter has at least the ABI alignment of the return type.
1546 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1547 if (EltTy->isSized())
1548 Align = DL.getABITypeAlignment(EltTy);
1550 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
1551 Align = AI->getAlignment();
1552 else if (auto CS = ImmutableCallSite(V))
1553 Align = CS.getAttributes().getParamAlignment(AttributeSet::ReturnIndex);
1554 else if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1555 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_align)) {
1556 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
1557 Align = CI->getLimitedValue();
1563 /// Determine which bits of V are known to be either zero or one and return
1564 /// them in the KnownZero/KnownOne bit sets.
1566 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1567 /// we cannot optimize based on the assumption that it is zero without changing
1568 /// it to be an explicit zero. If we don't change it to zero, other code could
1569 /// optimized based on the contradictory assumption that it is non-zero.
1570 /// Because instcombine aggressively folds operations with undef args anyway,
1571 /// this won't lose us code quality.
1573 /// This function is defined on values with integer type, values with pointer
1574 /// type, and vectors of integers. In the case
1575 /// where V is a vector, known zero, and known one values are the
1576 /// same width as the vector element, and the bit is set only if it is true
1577 /// for all of the elements in the vector.
1578 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1579 const DataLayout &DL, unsigned Depth, const Query &Q) {
1580 assert(V && "No Value?");
1581 assert(Depth <= MaxDepth && "Limit Search Depth");
1582 unsigned BitWidth = KnownZero.getBitWidth();
1584 assert((V->getType()->isIntOrIntVectorTy() ||
1585 V->getType()->isFPOrFPVectorTy() ||
1586 V->getType()->getScalarType()->isPointerTy()) &&
1587 "Not integer, floating point, or pointer type!");
1588 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1589 (!V->getType()->isIntOrIntVectorTy() ||
1590 V->getType()->getScalarSizeInBits() == BitWidth) &&
1591 KnownZero.getBitWidth() == BitWidth &&
1592 KnownOne.getBitWidth() == BitWidth &&
1593 "V, KnownOne and KnownZero should have same BitWidth");
1595 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1596 // We know all of the bits for a constant!
1597 KnownOne = CI->getValue();
1598 KnownZero = ~KnownOne;
1601 // Null and aggregate-zero are all-zeros.
1602 if (isa<ConstantPointerNull>(V) ||
1603 isa<ConstantAggregateZero>(V)) {
1604 KnownOne.clearAllBits();
1605 KnownZero = APInt::getAllOnesValue(BitWidth);
1608 // Handle a constant vector by taking the intersection of the known bits of
1609 // each element. There is no real need to handle ConstantVector here, because
1610 // we don't handle undef in any particularly useful way.
1611 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1612 // We know that CDS must be a vector of integers. Take the intersection of
1614 KnownZero.setAllBits(); KnownOne.setAllBits();
1615 APInt Elt(KnownZero.getBitWidth(), 0);
1616 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1617 Elt = CDS->getElementAsInteger(i);
1624 // Start out not knowing anything.
1625 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1627 // Limit search depth.
1628 // All recursive calls that increase depth must come after this.
1629 if (Depth == MaxDepth)
1632 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1633 // the bits of its aliasee.
1634 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1635 if (!GA->mayBeOverridden())
1636 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1640 if (Operator *I = dyn_cast<Operator>(V))
1641 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1643 // Aligned pointers have trailing zeros - refine KnownZero set
1644 if (V->getType()->isPointerTy()) {
1645 unsigned Align = getAlignment(V, DL);
1647 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1650 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1651 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1652 // computeKnownBitsFromOperator.
1654 // Check whether a nearby assume intrinsic can determine some known bits.
1655 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1657 // Check whether there's a dominating condition which implies something about
1658 // this value at the given context.
1659 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1660 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1663 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1666 /// Determine whether the sign bit is known to be zero or one.
1667 /// Convenience wrapper around computeKnownBits.
1668 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1669 const DataLayout &DL, unsigned Depth, const Query &Q) {
1670 unsigned BitWidth = getBitWidth(V->getType(), DL);
1676 APInt ZeroBits(BitWidth, 0);
1677 APInt OneBits(BitWidth, 0);
1678 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1679 KnownOne = OneBits[BitWidth - 1];
1680 KnownZero = ZeroBits[BitWidth - 1];
1683 /// Return true if the given value is known to have exactly one
1684 /// bit set when defined. For vectors return true if every element is known to
1685 /// be a power of two when defined. Supports values with integer or pointer
1686 /// types and vectors of integers.
1687 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1688 const Query &Q, const DataLayout &DL) {
1689 if (Constant *C = dyn_cast<Constant>(V)) {
1690 if (C->isNullValue())
1692 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1693 return CI->getValue().isPowerOf2();
1694 // TODO: Handle vector constants.
1697 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1698 // it is shifted off the end then the result is undefined.
1699 if (match(V, m_Shl(m_One(), m_Value())))
1702 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1703 // bottom. If it is shifted off the bottom then the result is undefined.
1704 if (match(V, m_LShr(m_SignBit(), m_Value())))
1707 // The remaining tests are all recursive, so bail out if we hit the limit.
1708 if (Depth++ == MaxDepth)
1711 Value *X = nullptr, *Y = nullptr;
1712 // A shift of a power of two is a power of two or zero.
1713 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1714 match(V, m_Shr(m_Value(X), m_Value()))))
1715 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1717 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1718 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1720 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1721 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1722 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1724 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1725 // A power of two and'd with anything is a power of two or zero.
1726 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1727 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1729 // X & (-X) is always a power of two or zero.
1730 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1735 // Adding a power-of-two or zero to the same power-of-two or zero yields
1736 // either the original power-of-two, a larger power-of-two or zero.
1737 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1738 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1739 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1740 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1741 match(X, m_And(m_Value(), m_Specific(Y))))
1742 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1744 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1745 match(Y, m_And(m_Value(), m_Specific(X))))
1746 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1749 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1750 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1751 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1753 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1754 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1755 // If i8 V is a power of two or zero:
1756 // ZeroBits: 1 1 1 0 1 1 1 1
1757 // ~ZeroBits: 0 0 0 1 0 0 0 0
1758 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1759 // If OrZero isn't set, we cannot give back a zero result.
1760 // Make sure either the LHS or RHS has a bit set.
1761 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1766 // An exact divide or right shift can only shift off zero bits, so the result
1767 // is a power of two only if the first operand is a power of two and not
1768 // copying a sign bit (sdiv int_min, 2).
1769 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1770 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1771 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1778 /// \brief Test whether a GEP's result is known to be non-null.
1780 /// Uses properties inherent in a GEP to try to determine whether it is known
1783 /// Currently this routine does not support vector GEPs.
1784 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1785 unsigned Depth, const Query &Q) {
1786 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1789 // FIXME: Support vector-GEPs.
1790 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1792 // If the base pointer is non-null, we cannot walk to a null address with an
1793 // inbounds GEP in address space zero.
1794 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1797 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1798 // If so, then the GEP cannot produce a null pointer, as doing so would
1799 // inherently violate the inbounds contract within address space zero.
1800 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1801 GTI != GTE; ++GTI) {
1802 // Struct types are easy -- they must always be indexed by a constant.
1803 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1804 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1805 unsigned ElementIdx = OpC->getZExtValue();
1806 const StructLayout *SL = DL.getStructLayout(STy);
1807 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1808 if (ElementOffset > 0)
1813 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1814 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1817 // Fast path the constant operand case both for efficiency and so we don't
1818 // increment Depth when just zipping down an all-constant GEP.
1819 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1825 // We post-increment Depth here because while isKnownNonZero increments it
1826 // as well, when we pop back up that increment won't persist. We don't want
1827 // to recurse 10k times just because we have 10k GEP operands. We don't
1828 // bail completely out because we want to handle constant GEPs regardless
1830 if (Depth++ >= MaxDepth)
1833 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1840 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1841 /// ensure that the value it's attached to is never Value? 'RangeType' is
1842 /// is the type of the value described by the range.
1843 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1844 const APInt& Value) {
1845 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1846 assert(NumRanges >= 1);
1847 for (unsigned i = 0; i < NumRanges; ++i) {
1848 ConstantInt *Lower =
1849 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1850 ConstantInt *Upper =
1851 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1852 ConstantRange Range(Lower->getValue(), Upper->getValue());
1853 if (Range.contains(Value))
1859 /// Return true if the given value is known to be non-zero when defined.
1860 /// For vectors return true if every element is known to be non-zero when
1861 /// defined. Supports values with integer or pointer type and vectors of
1863 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1865 if (Constant *C = dyn_cast<Constant>(V)) {
1866 if (C->isNullValue())
1868 if (isa<ConstantInt>(C))
1869 // Must be non-zero due to null test above.
1871 // TODO: Handle vectors
1875 if (Instruction* I = dyn_cast<Instruction>(V)) {
1876 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1877 // If the possible ranges don't contain zero, then the value is
1878 // definitely non-zero.
1879 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1880 const APInt ZeroValue(Ty->getBitWidth(), 0);
1881 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1887 // The remaining tests are all recursive, so bail out if we hit the limit.
1888 if (Depth++ >= MaxDepth)
1891 // Check for pointer simplifications.
1892 if (V->getType()->isPointerTy()) {
1893 if (isKnownNonNull(V))
1895 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1896 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1900 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1902 // X | Y != 0 if X != 0 or Y != 0.
1903 Value *X = nullptr, *Y = nullptr;
1904 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1905 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1907 // ext X != 0 if X != 0.
1908 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1909 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1911 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1912 // if the lowest bit is shifted off the end.
1913 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1914 // shl nuw can't remove any non-zero bits.
1915 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1916 if (BO->hasNoUnsignedWrap())
1917 return isKnownNonZero(X, DL, Depth, Q);
1919 APInt KnownZero(BitWidth, 0);
1920 APInt KnownOne(BitWidth, 0);
1921 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1925 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1926 // defined if the sign bit is shifted off the end.
1927 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1928 // shr exact can only shift out zero bits.
1929 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1931 return isKnownNonZero(X, DL, Depth, Q);
1933 bool XKnownNonNegative, XKnownNegative;
1934 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1938 // If the shifter operand is a constant, and all of the bits shifted
1939 // out are known to be zero, and X is known non-zero then at least one
1940 // non-zero bit must remain.
1941 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1942 APInt KnownZero(BitWidth, 0);
1943 APInt KnownOne(BitWidth, 0);
1944 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1946 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1947 // Is there a known one in the portion not shifted out?
1948 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1950 // Are all the bits to be shifted out known zero?
1951 if (KnownZero.countTrailingOnes() >= ShiftVal)
1952 return isKnownNonZero(X, DL, Depth, Q);
1955 // div exact can only produce a zero if the dividend is zero.
1956 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1957 return isKnownNonZero(X, DL, Depth, Q);
1960 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1961 bool XKnownNonNegative, XKnownNegative;
1962 bool YKnownNonNegative, YKnownNegative;
1963 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1964 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1966 // If X and Y are both non-negative (as signed values) then their sum is not
1967 // zero unless both X and Y are zero.
1968 if (XKnownNonNegative && YKnownNonNegative)
1969 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1972 // If X and Y are both negative (as signed values) then their sum is not
1973 // zero unless both X and Y equal INT_MIN.
1974 if (BitWidth && XKnownNegative && YKnownNegative) {
1975 APInt KnownZero(BitWidth, 0);
1976 APInt KnownOne(BitWidth, 0);
1977 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1978 // The sign bit of X is set. If some other bit is set then X is not equal
1980 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1981 if ((KnownOne & Mask) != 0)
1983 // The sign bit of Y is set. If some other bit is set then Y is not equal
1985 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1986 if ((KnownOne & Mask) != 0)
1990 // The sum of a non-negative number and a power of two is not zero.
1991 if (XKnownNonNegative &&
1992 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1994 if (YKnownNonNegative &&
1995 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1999 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2000 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2001 // If X and Y are non-zero then so is X * Y as long as the multiplication
2002 // does not overflow.
2003 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
2004 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
2007 // (C ? X : Y) != 0 if X != 0 and Y != 0.
2008 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2009 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
2010 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
2014 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
2015 // Try and detect a recurrence that monotonically increases from a
2016 // starting value, as these are common as induction variables.
2017 if (PN->getNumIncomingValues() == 2) {
2018 Value *Start = PN->getIncomingValue(0);
2019 Value *Induction = PN->getIncomingValue(1);
2020 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2021 std::swap(Start, Induction);
2022 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2023 if (!C->isZero() && !C->isNegative()) {
2025 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2026 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2034 if (!BitWidth) return false;
2035 APInt KnownZero(BitWidth, 0);
2036 APInt KnownOne(BitWidth, 0);
2037 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2038 return KnownOne != 0;
2041 /// Return true if V2 == V1 + X, where X is known non-zero.
2042 static bool isAddOfNonZero(Value *V1, Value *V2, const DataLayout &DL,
2044 BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2045 if (!BO || BO->getOpcode() != Instruction::Add)
2047 Value *Op = nullptr;
2048 if (V2 == BO->getOperand(0))
2049 Op = BO->getOperand(1);
2050 else if (V2 == BO->getOperand(1))
2051 Op = BO->getOperand(0);
2054 return isKnownNonZero(Op, DL, 0, Q);
2057 /// Return true if it is known that V1 != V2.
2058 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
2060 if (V1->getType()->isVectorTy() || V1 == V2)
2062 if (V1->getType() != V2->getType())
2063 // We can't look through casts yet.
2065 if (isAddOfNonZero(V1, V2, DL, Q) || isAddOfNonZero(V2, V1, DL, Q))
2068 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
2069 // Are any known bits in V1 contradictory to known bits in V2? If V1
2070 // has a known zero where V2 has a known one, they must not be equal.
2071 auto BitWidth = Ty->getBitWidth();
2072 APInt KnownZero1(BitWidth, 0);
2073 APInt KnownOne1(BitWidth, 0);
2074 computeKnownBits(V1, KnownZero1, KnownOne1, DL, 0, Q);
2075 APInt KnownZero2(BitWidth, 0);
2076 APInt KnownOne2(BitWidth, 0);
2077 computeKnownBits(V2, KnownZero2, KnownOne2, DL, 0, Q);
2079 auto OppositeBits = (KnownZero1 & KnownOne2) | (KnownZero2 & KnownOne1);
2080 if (OppositeBits.getBoolValue())
2086 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2087 /// simplify operations downstream. Mask is known to be zero for bits that V
2090 /// This function is defined on values with integer type, values with pointer
2091 /// type, and vectors of integers. In the case
2092 /// where V is a vector, the mask, known zero, and known one values are the
2093 /// same width as the vector element, and the bit is set only if it is true
2094 /// for all of the elements in the vector.
2095 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
2096 unsigned Depth, const Query &Q) {
2097 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
2098 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2099 return (KnownZero & Mask) == Mask;
2104 /// Return the number of times the sign bit of the register is replicated into
2105 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2106 /// (itself), but other cases can give us information. For example, immediately
2107 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2108 /// other, so we return 3.
2110 /// 'Op' must have a scalar integer type.
2112 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
2114 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
2116 unsigned FirstAnswer = 1;
2118 // Note that ConstantInt is handled by the general computeKnownBits case
2122 return 1; // Limit search depth.
2124 Operator *U = dyn_cast<Operator>(V);
2125 switch (Operator::getOpcode(V)) {
2127 case Instruction::SExt:
2128 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2129 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
2131 case Instruction::SDiv: {
2132 const APInt *Denominator;
2133 // sdiv X, C -> adds log(C) sign bits.
2134 if (match(U->getOperand(1), m_APInt(Denominator))) {
2136 // Ignore non-positive denominator.
2137 if (!Denominator->isStrictlyPositive())
2140 // Calculate the incoming numerator bits.
2141 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2143 // Add floor(log(C)) bits to the numerator bits.
2144 return std::min(TyBits, NumBits + Denominator->logBase2());
2149 case Instruction::SRem: {
2150 const APInt *Denominator;
2151 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2152 // positive constant. This let us put a lower bound on the number of sign
2154 if (match(U->getOperand(1), m_APInt(Denominator))) {
2156 // Ignore non-positive denominator.
2157 if (!Denominator->isStrictlyPositive())
2160 // Calculate the incoming numerator bits. SRem by a positive constant
2161 // can't lower the number of sign bits.
2163 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2165 // Calculate the leading sign bit constraints by examining the
2166 // denominator. Given that the denominator is positive, there are two
2169 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2170 // (1 << ceilLogBase2(C)).
2172 // 2. the numerator is negative. Then the result range is (-C,0] and
2173 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2175 // Thus a lower bound on the number of sign bits is `TyBits -
2176 // ceilLogBase2(C)`.
2178 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2179 return std::max(NumrBits, ResBits);
2184 case Instruction::AShr: {
2185 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2186 // ashr X, C -> adds C sign bits. Vectors too.
2188 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2189 Tmp += ShAmt->getZExtValue();
2190 if (Tmp > TyBits) Tmp = TyBits;
2194 case Instruction::Shl: {
2196 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2197 // shl destroys sign bits.
2198 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2199 Tmp2 = ShAmt->getZExtValue();
2200 if (Tmp2 >= TyBits || // Bad shift.
2201 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2206 case Instruction::And:
2207 case Instruction::Or:
2208 case Instruction::Xor: // NOT is handled here.
2209 // Logical binary ops preserve the number of sign bits at the worst.
2210 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2212 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2213 FirstAnswer = std::min(Tmp, Tmp2);
2214 // We computed what we know about the sign bits as our first
2215 // answer. Now proceed to the generic code that uses
2216 // computeKnownBits, and pick whichever answer is better.
2220 case Instruction::Select:
2221 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2222 if (Tmp == 1) return 1; // Early out.
2223 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2224 return std::min(Tmp, Tmp2);
2226 case Instruction::Add:
2227 // Add can have at most one carry bit. Thus we know that the output
2228 // is, at worst, one more bit than the inputs.
2229 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2230 if (Tmp == 1) return 1; // Early out.
2232 // Special case decrementing a value (ADD X, -1):
2233 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2234 if (CRHS->isAllOnesValue()) {
2235 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2236 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2239 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2241 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2244 // If we are subtracting one from a positive number, there is no carry
2245 // out of the result.
2246 if (KnownZero.isNegative())
2250 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2251 if (Tmp2 == 1) return 1;
2252 return std::min(Tmp, Tmp2)-1;
2254 case Instruction::Sub:
2255 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2256 if (Tmp2 == 1) return 1;
2259 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2260 if (CLHS->isNullValue()) {
2261 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2262 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2264 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2266 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2269 // If the input is known to be positive (the sign bit is known clear),
2270 // the output of the NEG has the same number of sign bits as the input.
2271 if (KnownZero.isNegative())
2274 // Otherwise, we treat this like a SUB.
2277 // Sub can have at most one carry bit. Thus we know that the output
2278 // is, at worst, one more bit than the inputs.
2279 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2280 if (Tmp == 1) return 1; // Early out.
2281 return std::min(Tmp, Tmp2)-1;
2283 case Instruction::PHI: {
2284 PHINode *PN = cast<PHINode>(U);
2285 unsigned NumIncomingValues = PN->getNumIncomingValues();
2286 // Don't analyze large in-degree PHIs.
2287 if (NumIncomingValues > 4) break;
2288 // Unreachable blocks may have zero-operand PHI nodes.
2289 if (NumIncomingValues == 0) break;
2291 // Take the minimum of all incoming values. This can't infinitely loop
2292 // because of our depth threshold.
2293 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2294 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2295 if (Tmp == 1) return Tmp;
2297 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2302 case Instruction::Trunc:
2303 // FIXME: it's tricky to do anything useful for this, but it is an important
2304 // case for targets like X86.
2308 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2309 // use this information.
2310 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2312 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2314 if (KnownZero.isNegative()) { // sign bit is 0
2316 } else if (KnownOne.isNegative()) { // sign bit is 1;
2323 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2324 // the number of identical bits in the top of the input value.
2326 Mask <<= Mask.getBitWidth()-TyBits;
2327 // Return # leading zeros. We use 'min' here in case Val was zero before
2328 // shifting. We don't want to return '64' as for an i32 "0".
2329 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2332 /// This function computes the integer multiple of Base that equals V.
2333 /// If successful, it returns true and returns the multiple in
2334 /// Multiple. If unsuccessful, it returns false. It looks
2335 /// through SExt instructions only if LookThroughSExt is true.
2336 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2337 bool LookThroughSExt, unsigned Depth) {
2338 const unsigned MaxDepth = 6;
2340 assert(V && "No Value?");
2341 assert(Depth <= MaxDepth && "Limit Search Depth");
2342 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2344 Type *T = V->getType();
2346 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2356 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2357 Constant *BaseVal = ConstantInt::get(T, Base);
2358 if (CO && CO == BaseVal) {
2360 Multiple = ConstantInt::get(T, 1);
2364 if (CI && CI->getZExtValue() % Base == 0) {
2365 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2369 if (Depth == MaxDepth) return false; // Limit search depth.
2371 Operator *I = dyn_cast<Operator>(V);
2372 if (!I) return false;
2374 switch (I->getOpcode()) {
2376 case Instruction::SExt:
2377 if (!LookThroughSExt) return false;
2378 // otherwise fall through to ZExt
2379 case Instruction::ZExt:
2380 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2381 LookThroughSExt, Depth+1);
2382 case Instruction::Shl:
2383 case Instruction::Mul: {
2384 Value *Op0 = I->getOperand(0);
2385 Value *Op1 = I->getOperand(1);
2387 if (I->getOpcode() == Instruction::Shl) {
2388 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2389 if (!Op1CI) return false;
2390 // Turn Op0 << Op1 into Op0 * 2^Op1
2391 APInt Op1Int = Op1CI->getValue();
2392 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2393 APInt API(Op1Int.getBitWidth(), 0);
2394 API.setBit(BitToSet);
2395 Op1 = ConstantInt::get(V->getContext(), API);
2398 Value *Mul0 = nullptr;
2399 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2400 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2401 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2402 if (Op1C->getType()->getPrimitiveSizeInBits() <
2403 MulC->getType()->getPrimitiveSizeInBits())
2404 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2405 if (Op1C->getType()->getPrimitiveSizeInBits() >
2406 MulC->getType()->getPrimitiveSizeInBits())
2407 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2409 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2410 Multiple = ConstantExpr::getMul(MulC, Op1C);
2414 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2415 if (Mul0CI->getValue() == 1) {
2416 // V == Base * Op1, so return Op1
2422 Value *Mul1 = nullptr;
2423 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2424 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2425 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2426 if (Op0C->getType()->getPrimitiveSizeInBits() <
2427 MulC->getType()->getPrimitiveSizeInBits())
2428 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2429 if (Op0C->getType()->getPrimitiveSizeInBits() >
2430 MulC->getType()->getPrimitiveSizeInBits())
2431 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2433 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2434 Multiple = ConstantExpr::getMul(MulC, Op0C);
2438 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2439 if (Mul1CI->getValue() == 1) {
2440 // V == Base * Op0, so return Op0
2448 // We could not determine if V is a multiple of Base.
2452 /// Return true if we can prove that the specified FP value is never equal to
2455 /// NOTE: this function will need to be revisited when we support non-default
2458 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2459 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2460 return !CFP->getValueAPF().isNegZero();
2462 // FIXME: Magic number! At the least, this should be given a name because it's
2463 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2464 // expose it as a parameter, so it can be used for testing / experimenting.
2466 return false; // Limit search depth.
2468 const Operator *I = dyn_cast<Operator>(V);
2469 if (!I) return false;
2471 // Check if the nsz fast-math flag is set
2472 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2473 if (FPO->hasNoSignedZeros())
2476 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2477 if (I->getOpcode() == Instruction::FAdd)
2478 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2479 if (CFP->isNullValue())
2482 // sitofp and uitofp turn into +0.0 for zero.
2483 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2486 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2487 // sqrt(-0.0) = -0.0, no other negative results are possible.
2488 if (II->getIntrinsicID() == Intrinsic::sqrt)
2489 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2491 if (const CallInst *CI = dyn_cast<CallInst>(I))
2492 if (const Function *F = CI->getCalledFunction()) {
2493 if (F->isDeclaration()) {
2495 if (F->getName() == "abs") return true;
2496 // fabs[lf](x) != -0.0
2497 if (F->getName() == "fabs") return true;
2498 if (F->getName() == "fabsf") return true;
2499 if (F->getName() == "fabsl") return true;
2500 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2501 F->getName() == "sqrtl")
2502 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2509 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2510 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2511 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2513 // FIXME: Magic number! At the least, this should be given a name because it's
2514 // used similarly in CannotBeNegativeZero(). A better fix may be to
2515 // expose it as a parameter, so it can be used for testing / experimenting.
2517 return false; // Limit search depth.
2519 const Operator *I = dyn_cast<Operator>(V);
2520 if (!I) return false;
2522 switch (I->getOpcode()) {
2524 case Instruction::FMul:
2525 // x*x is always non-negative or a NaN.
2526 if (I->getOperand(0) == I->getOperand(1))
2529 case Instruction::FAdd:
2530 case Instruction::FDiv:
2531 case Instruction::FRem:
2532 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2533 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2534 case Instruction::FPExt:
2535 case Instruction::FPTrunc:
2536 // Widening/narrowing never change sign.
2537 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2538 case Instruction::Call:
2539 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2540 switch (II->getIntrinsicID()) {
2542 case Intrinsic::exp:
2543 case Intrinsic::exp2:
2544 case Intrinsic::fabs:
2545 case Intrinsic::sqrt:
2547 case Intrinsic::powi:
2548 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2549 // powi(x,n) is non-negative if n is even.
2550 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2553 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2554 case Intrinsic::fma:
2555 case Intrinsic::fmuladd:
2556 // x*x+y is non-negative if y is non-negative.
2557 return I->getOperand(0) == I->getOperand(1) &&
2558 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2565 /// If the specified value can be set by repeating the same byte in memory,
2566 /// return the i8 value that it is represented with. This is
2567 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2568 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2569 /// byte store (e.g. i16 0x1234), return null.
2570 Value *llvm::isBytewiseValue(Value *V) {
2571 // All byte-wide stores are splatable, even of arbitrary variables.
2572 if (V->getType()->isIntegerTy(8)) return V;
2574 // Handle 'null' ConstantArrayZero etc.
2575 if (Constant *C = dyn_cast<Constant>(V))
2576 if (C->isNullValue())
2577 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2579 // Constant float and double values can be handled as integer values if the
2580 // corresponding integer value is "byteable". An important case is 0.0.
2581 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2582 if (CFP->getType()->isFloatTy())
2583 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2584 if (CFP->getType()->isDoubleTy())
2585 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2586 // Don't handle long double formats, which have strange constraints.
2589 // We can handle constant integers that are multiple of 8 bits.
2590 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2591 if (CI->getBitWidth() % 8 == 0) {
2592 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2594 if (!CI->getValue().isSplat(8))
2596 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2600 // A ConstantDataArray/Vector is splatable if all its members are equal and
2602 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2603 Value *Elt = CA->getElementAsConstant(0);
2604 Value *Val = isBytewiseValue(Elt);
2608 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2609 if (CA->getElementAsConstant(I) != Elt)
2615 // Conceptually, we could handle things like:
2616 // %a = zext i8 %X to i16
2617 // %b = shl i16 %a, 8
2618 // %c = or i16 %a, %b
2619 // but until there is an example that actually needs this, it doesn't seem
2620 // worth worrying about.
2625 // This is the recursive version of BuildSubAggregate. It takes a few different
2626 // arguments. Idxs is the index within the nested struct From that we are
2627 // looking at now (which is of type IndexedType). IdxSkip is the number of
2628 // indices from Idxs that should be left out when inserting into the resulting
2629 // struct. To is the result struct built so far, new insertvalue instructions
2631 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2632 SmallVectorImpl<unsigned> &Idxs,
2634 Instruction *InsertBefore) {
2635 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2637 // Save the original To argument so we can modify it
2639 // General case, the type indexed by Idxs is a struct
2640 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2641 // Process each struct element recursively
2644 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2648 // Couldn't find any inserted value for this index? Cleanup
2649 while (PrevTo != OrigTo) {
2650 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2651 PrevTo = Del->getAggregateOperand();
2652 Del->eraseFromParent();
2654 // Stop processing elements
2658 // If we successfully found a value for each of our subaggregates
2662 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2663 // the struct's elements had a value that was inserted directly. In the latter
2664 // case, perhaps we can't determine each of the subelements individually, but
2665 // we might be able to find the complete struct somewhere.
2667 // Find the value that is at that particular spot
2668 Value *V = FindInsertedValue(From, Idxs);
2673 // Insert the value in the new (sub) aggregrate
2674 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2675 "tmp", InsertBefore);
2678 // This helper takes a nested struct and extracts a part of it (which is again a
2679 // struct) into a new value. For example, given the struct:
2680 // { a, { b, { c, d }, e } }
2681 // and the indices "1, 1" this returns
2684 // It does this by inserting an insertvalue for each element in the resulting
2685 // struct, as opposed to just inserting a single struct. This will only work if
2686 // each of the elements of the substruct are known (ie, inserted into From by an
2687 // insertvalue instruction somewhere).
2689 // All inserted insertvalue instructions are inserted before InsertBefore
2690 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2691 Instruction *InsertBefore) {
2692 assert(InsertBefore && "Must have someplace to insert!");
2693 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2695 Value *To = UndefValue::get(IndexedType);
2696 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2697 unsigned IdxSkip = Idxs.size();
2699 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2702 /// Given an aggregrate and an sequence of indices, see if
2703 /// the scalar value indexed is already around as a register, for example if it
2704 /// were inserted directly into the aggregrate.
2706 /// If InsertBefore is not null, this function will duplicate (modified)
2707 /// insertvalues when a part of a nested struct is extracted.
2708 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2709 Instruction *InsertBefore) {
2710 // Nothing to index? Just return V then (this is useful at the end of our
2712 if (idx_range.empty())
2714 // We have indices, so V should have an indexable type.
2715 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2716 "Not looking at a struct or array?");
2717 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2718 "Invalid indices for type?");
2720 if (Constant *C = dyn_cast<Constant>(V)) {
2721 C = C->getAggregateElement(idx_range[0]);
2722 if (!C) return nullptr;
2723 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2726 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2727 // Loop the indices for the insertvalue instruction in parallel with the
2728 // requested indices
2729 const unsigned *req_idx = idx_range.begin();
2730 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2731 i != e; ++i, ++req_idx) {
2732 if (req_idx == idx_range.end()) {
2733 // We can't handle this without inserting insertvalues
2737 // The requested index identifies a part of a nested aggregate. Handle
2738 // this specially. For example,
2739 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2740 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2741 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2742 // This can be changed into
2743 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2744 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2745 // which allows the unused 0,0 element from the nested struct to be
2747 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2751 // This insert value inserts something else than what we are looking for.
2752 // See if the (aggregate) value inserted into has the value we are
2753 // looking for, then.
2755 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2758 // If we end up here, the indices of the insertvalue match with those
2759 // requested (though possibly only partially). Now we recursively look at
2760 // the inserted value, passing any remaining indices.
2761 return FindInsertedValue(I->getInsertedValueOperand(),
2762 makeArrayRef(req_idx, idx_range.end()),
2766 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2767 // If we're extracting a value from an aggregate that was extracted from
2768 // something else, we can extract from that something else directly instead.
2769 // However, we will need to chain I's indices with the requested indices.
2771 // Calculate the number of indices required
2772 unsigned size = I->getNumIndices() + idx_range.size();
2773 // Allocate some space to put the new indices in
2774 SmallVector<unsigned, 5> Idxs;
2776 // Add indices from the extract value instruction
2777 Idxs.append(I->idx_begin(), I->idx_end());
2779 // Add requested indices
2780 Idxs.append(idx_range.begin(), idx_range.end());
2782 assert(Idxs.size() == size
2783 && "Number of indices added not correct?");
2785 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2787 // Otherwise, we don't know (such as, extracting from a function return value
2788 // or load instruction)
2792 /// Analyze the specified pointer to see if it can be expressed as a base
2793 /// pointer plus a constant offset. Return the base and offset to the caller.
2794 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2795 const DataLayout &DL) {
2796 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2797 APInt ByteOffset(BitWidth, 0);
2799 if (Ptr->getType()->isVectorTy())
2802 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2803 APInt GEPOffset(BitWidth, 0);
2804 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2807 ByteOffset += GEPOffset;
2809 Ptr = GEP->getPointerOperand();
2810 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2811 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2812 Ptr = cast<Operator>(Ptr)->getOperand(0);
2813 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2814 if (GA->mayBeOverridden())
2816 Ptr = GA->getAliasee();
2821 Offset = ByteOffset.getSExtValue();
2826 /// This function computes the length of a null-terminated C string pointed to
2827 /// by V. If successful, it returns true and returns the string in Str.
2828 /// If unsuccessful, it returns false.
2829 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2830 uint64_t Offset, bool TrimAtNul) {
2833 // Look through bitcast instructions and geps.
2834 V = V->stripPointerCasts();
2836 // If the value is a GEP instruction or constant expression, treat it as an
2838 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2839 // Make sure the GEP has exactly three arguments.
2840 if (GEP->getNumOperands() != 3)
2843 // Make sure the index-ee is a pointer to array of i8.
2844 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2845 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2846 if (!AT || !AT->getElementType()->isIntegerTy(8))
2849 // Check to make sure that the first operand of the GEP is an integer and
2850 // has value 0 so that we are sure we're indexing into the initializer.
2851 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2852 if (!FirstIdx || !FirstIdx->isZero())
2855 // If the second index isn't a ConstantInt, then this is a variable index
2856 // into the array. If this occurs, we can't say anything meaningful about
2858 uint64_t StartIdx = 0;
2859 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2860 StartIdx = CI->getZExtValue();
2863 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2867 // The GEP instruction, constant or instruction, must reference a global
2868 // variable that is a constant and is initialized. The referenced constant
2869 // initializer is the array that we'll use for optimization.
2870 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2871 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2874 // Handle the all-zeros case
2875 if (GV->getInitializer()->isNullValue()) {
2876 // This is a degenerate case. The initializer is constant zero so the
2877 // length of the string must be zero.
2882 // Must be a Constant Array
2883 const ConstantDataArray *Array =
2884 dyn_cast<ConstantDataArray>(GV->getInitializer());
2885 if (!Array || !Array->isString())
2888 // Get the number of elements in the array
2889 uint64_t NumElts = Array->getType()->getArrayNumElements();
2891 // Start out with the entire array in the StringRef.
2892 Str = Array->getAsString();
2894 if (Offset > NumElts)
2897 // Skip over 'offset' bytes.
2898 Str = Str.substr(Offset);
2901 // Trim off the \0 and anything after it. If the array is not nul
2902 // terminated, we just return the whole end of string. The client may know
2903 // some other way that the string is length-bound.
2904 Str = Str.substr(0, Str.find('\0'));
2909 // These next two are very similar to the above, but also look through PHI
2911 // TODO: See if we can integrate these two together.
2913 /// If we can compute the length of the string pointed to by
2914 /// the specified pointer, return 'len+1'. If we can't, return 0.
2915 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2916 // Look through noop bitcast instructions.
2917 V = V->stripPointerCasts();
2919 // If this is a PHI node, there are two cases: either we have already seen it
2921 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2922 if (!PHIs.insert(PN).second)
2923 return ~0ULL; // already in the set.
2925 // If it was new, see if all the input strings are the same length.
2926 uint64_t LenSoFar = ~0ULL;
2927 for (Value *IncValue : PN->incoming_values()) {
2928 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2929 if (Len == 0) return 0; // Unknown length -> unknown.
2931 if (Len == ~0ULL) continue;
2933 if (Len != LenSoFar && LenSoFar != ~0ULL)
2934 return 0; // Disagree -> unknown.
2938 // Success, all agree.
2942 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2943 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2944 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2945 if (Len1 == 0) return 0;
2946 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2947 if (Len2 == 0) return 0;
2948 if (Len1 == ~0ULL) return Len2;
2949 if (Len2 == ~0ULL) return Len1;
2950 if (Len1 != Len2) return 0;
2954 // Otherwise, see if we can read the string.
2956 if (!getConstantStringInfo(V, StrData))
2959 return StrData.size()+1;
2962 /// If we can compute the length of the string pointed to by
2963 /// the specified pointer, return 'len+1'. If we can't, return 0.
2964 uint64_t llvm::GetStringLength(Value *V) {
2965 if (!V->getType()->isPointerTy()) return 0;
2967 SmallPtrSet<PHINode*, 32> PHIs;
2968 uint64_t Len = GetStringLengthH(V, PHIs);
2969 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2970 // an empty string as a length.
2971 return Len == ~0ULL ? 1 : Len;
2974 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2975 /// previous iteration of the loop was referring to the same object as \p PN.
2976 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2977 // Find the loop-defined value.
2978 Loop *L = LI->getLoopFor(PN->getParent());
2979 if (PN->getNumIncomingValues() != 2)
2982 // Find the value from previous iteration.
2983 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2984 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2985 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2986 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2989 // If a new pointer is loaded in the loop, the pointer references a different
2990 // object in every iteration. E.g.:
2994 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2995 if (!L->isLoopInvariant(Load->getPointerOperand()))
3000 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3001 unsigned MaxLookup) {
3002 if (!V->getType()->isPointerTy())
3004 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3005 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3006 V = GEP->getPointerOperand();
3007 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3008 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3009 V = cast<Operator>(V)->getOperand(0);
3010 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3011 if (GA->mayBeOverridden())
3013 V = GA->getAliasee();
3015 // See if InstructionSimplify knows any relevant tricks.
3016 if (Instruction *I = dyn_cast<Instruction>(V))
3017 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3018 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
3025 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3030 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3031 const DataLayout &DL, LoopInfo *LI,
3032 unsigned MaxLookup) {
3033 SmallPtrSet<Value *, 4> Visited;
3034 SmallVector<Value *, 4> Worklist;
3035 Worklist.push_back(V);
3037 Value *P = Worklist.pop_back_val();
3038 P = GetUnderlyingObject(P, DL, MaxLookup);
3040 if (!Visited.insert(P).second)
3043 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3044 Worklist.push_back(SI->getTrueValue());
3045 Worklist.push_back(SI->getFalseValue());
3049 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3050 // If this PHI changes the underlying object in every iteration of the
3051 // loop, don't look through it. Consider:
3054 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3058 // Prev is tracking Curr one iteration behind so they refer to different
3059 // underlying objects.
3060 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3061 isSameUnderlyingObjectInLoop(PN, LI))
3062 for (Value *IncValue : PN->incoming_values())
3063 Worklist.push_back(IncValue);
3067 Objects.push_back(P);
3068 } while (!Worklist.empty());
3071 /// Return true if the only users of this pointer are lifetime markers.
3072 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3073 for (const User *U : V->users()) {
3074 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3075 if (!II) return false;
3077 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3078 II->getIntrinsicID() != Intrinsic::lifetime_end)
3084 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
3085 Type *Ty, const DataLayout &DL,
3086 const Instruction *CtxI,
3087 const DominatorTree *DT,
3088 const TargetLibraryInfo *TLI) {
3089 assert(Offset.isNonNegative() && "offset can't be negative");
3090 assert(Ty->isSized() && "must be sized");
3092 APInt DerefBytes(Offset.getBitWidth(), 0);
3093 bool CheckForNonNull = false;
3094 if (const Argument *A = dyn_cast<Argument>(BV)) {
3095 DerefBytes = A->getDereferenceableBytes();
3096 if (!DerefBytes.getBoolValue()) {
3097 DerefBytes = A->getDereferenceableOrNullBytes();
3098 CheckForNonNull = true;
3100 } else if (auto CS = ImmutableCallSite(BV)) {
3101 DerefBytes = CS.getDereferenceableBytes(0);
3102 if (!DerefBytes.getBoolValue()) {
3103 DerefBytes = CS.getDereferenceableOrNullBytes(0);
3104 CheckForNonNull = true;
3106 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
3107 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
3108 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3109 DerefBytes = CI->getLimitedValue();
3111 if (!DerefBytes.getBoolValue()) {
3113 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
3114 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3115 DerefBytes = CI->getLimitedValue();
3117 CheckForNonNull = true;
3121 if (DerefBytes.getBoolValue())
3122 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
3123 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
3129 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
3130 const Instruction *CtxI,
3131 const DominatorTree *DT,
3132 const TargetLibraryInfo *TLI) {
3133 Type *VTy = V->getType();
3134 Type *Ty = VTy->getPointerElementType();
3138 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3139 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
3142 static bool isAligned(const Value *Base, APInt Offset, unsigned Align,
3143 const DataLayout &DL) {
3144 APInt BaseAlign(Offset.getBitWidth(), getAlignment(Base, DL));
3147 Type *Ty = Base->getType()->getPointerElementType();
3148 BaseAlign = DL.getABITypeAlignment(Ty);
3151 APInt Alignment(Offset.getBitWidth(), Align);
3153 assert(Alignment.isPowerOf2() && "must be a power of 2!");
3154 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1));
3157 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) {
3158 APInt Offset(DL.getTypeStoreSizeInBits(Base->getType()), 0);
3159 return isAligned(Base, Offset, Align, DL);
3162 /// Test if V is always a pointer to allocated and suitably aligned memory for
3163 /// a simple load or store.
3164 static bool isDereferenceableAndAlignedPointer(
3165 const Value *V, unsigned Align, const DataLayout &DL,
3166 const Instruction *CtxI, const DominatorTree *DT,
3167 const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited) {
3168 // Note that it is not safe to speculate into a malloc'd region because
3169 // malloc may return null.
3171 // These are obviously ok if aligned.
3172 if (isa<AllocaInst>(V))
3173 return isAligned(V, Align, DL);
3175 // It's not always safe to follow a bitcast, for example:
3176 // bitcast i8* (alloca i8) to i32*
3177 // would result in a 4-byte load from a 1-byte alloca. However,
3178 // if we're casting from a pointer from a type of larger size
3179 // to a type of smaller size (or the same size), and the alignment
3180 // is at least as large as for the resulting pointer type, then
3181 // we can look through the bitcast.
3182 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
3183 Type *STy = BC->getSrcTy()->getPointerElementType(),
3184 *DTy = BC->getDestTy()->getPointerElementType();
3185 if (STy->isSized() && DTy->isSized() &&
3186 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
3187 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
3188 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, DL,
3189 CtxI, DT, TLI, Visited);
3192 // Global variables which can't collapse to null are ok.
3193 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
3194 if (!GV->hasExternalWeakLinkage())
3195 return isAligned(V, Align, DL);
3197 // byval arguments are okay.
3198 if (const Argument *A = dyn_cast<Argument>(V))
3199 if (A->hasByValAttr())
3200 return isAligned(V, Align, DL);
3202 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
3203 return isAligned(V, Align, DL);
3205 // For GEPs, determine if the indexing lands within the allocated object.
3206 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3207 Type *VTy = GEP->getType();
3208 Type *Ty = VTy->getPointerElementType();
3209 const Value *Base = GEP->getPointerOperand();
3211 // Conservatively require that the base pointer be fully dereferenceable
3213 if (!Visited.insert(Base).second)
3215 if (!isDereferenceableAndAlignedPointer(Base, Align, DL, CtxI, DT, TLI,
3219 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
3220 if (!GEP->accumulateConstantOffset(DL, Offset))
3223 // Check if the load is within the bounds of the underlying object
3224 // and offset is aligned.
3225 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3226 Type *BaseType = Base->getType()->getPointerElementType();
3227 assert(isPowerOf2_32(Align) && "must be a power of 2!");
3228 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType)) &&
3229 !(Offset & APInt(Offset.getBitWidth(), Align-1));
3232 // For gc.relocate, look through relocations
3233 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3234 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3235 GCRelocateOperands RelocateInst(I);
3236 return isDereferenceableAndAlignedPointer(
3237 RelocateInst.getDerivedPtr(), Align, DL, CtxI, DT, TLI, Visited);
3240 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3241 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, DL,
3242 CtxI, DT, TLI, Visited);
3244 // If we don't know, assume the worst.
3248 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align,
3249 const DataLayout &DL,
3250 const Instruction *CtxI,
3251 const DominatorTree *DT,
3252 const TargetLibraryInfo *TLI) {
3253 // When dereferenceability information is provided by a dereferenceable
3254 // attribute, we know exactly how many bytes are dereferenceable. If we can
3255 // determine the exact offset to the attributed variable, we can use that
3256 // information here.
3257 Type *VTy = V->getType();
3258 Type *Ty = VTy->getPointerElementType();
3260 // Require ABI alignment for loads without alignment specification
3262 Align = DL.getABITypeAlignment(Ty);
3264 if (Ty->isSized()) {
3265 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3266 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3268 if (Offset.isNonNegative())
3269 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL, CtxI, DT, TLI) &&
3270 isAligned(BV, Offset, Align, DL))
3274 SmallPtrSet<const Value *, 32> Visited;
3275 return ::isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT, TLI,
3279 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3280 const Instruction *CtxI,
3281 const DominatorTree *DT,
3282 const TargetLibraryInfo *TLI) {
3283 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT, TLI);
3286 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3287 const Instruction *CtxI,
3288 const DominatorTree *DT,
3289 const TargetLibraryInfo *TLI) {
3290 const Operator *Inst = dyn_cast<Operator>(V);
3294 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3295 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3299 switch (Inst->getOpcode()) {
3302 case Instruction::UDiv:
3303 case Instruction::URem: {
3304 // x / y is undefined if y == 0.
3306 if (match(Inst->getOperand(1), m_APInt(V)))
3310 case Instruction::SDiv:
3311 case Instruction::SRem: {
3312 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3313 const APInt *Numerator, *Denominator;
3314 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3316 // We cannot hoist this division if the denominator is 0.
3317 if (*Denominator == 0)
3319 // It's safe to hoist if the denominator is not 0 or -1.
3320 if (*Denominator != -1)
3322 // At this point we know that the denominator is -1. It is safe to hoist as
3323 // long we know that the numerator is not INT_MIN.
3324 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3325 return !Numerator->isMinSignedValue();
3326 // The numerator *might* be MinSignedValue.
3329 case Instruction::Load: {
3330 const LoadInst *LI = cast<LoadInst>(Inst);
3331 if (!LI->isUnordered() ||
3332 // Speculative load may create a race that did not exist in the source.
3333 LI->getParent()->getParent()->hasFnAttribute(
3334 Attribute::SanitizeThread) ||
3335 // Speculative load may load data from dirty regions.
3336 LI->getParent()->getParent()->hasFnAttribute(
3337 Attribute::SanitizeAddress))
3339 const DataLayout &DL = LI->getModule()->getDataLayout();
3340 return isDereferenceableAndAlignedPointer(
3341 LI->getPointerOperand(), LI->getAlignment(), DL, CtxI, DT, TLI);
3343 case Instruction::Call: {
3344 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3345 switch (II->getIntrinsicID()) {
3346 // These synthetic intrinsics have no side-effects and just mark
3347 // information about their operands.
3348 // FIXME: There are other no-op synthetic instructions that potentially
3349 // should be considered at least *safe* to speculate...
3350 case Intrinsic::dbg_declare:
3351 case Intrinsic::dbg_value:
3354 case Intrinsic::bswap:
3355 case Intrinsic::ctlz:
3356 case Intrinsic::ctpop:
3357 case Intrinsic::cttz:
3358 case Intrinsic::objectsize:
3359 case Intrinsic::sadd_with_overflow:
3360 case Intrinsic::smul_with_overflow:
3361 case Intrinsic::ssub_with_overflow:
3362 case Intrinsic::uadd_with_overflow:
3363 case Intrinsic::umul_with_overflow:
3364 case Intrinsic::usub_with_overflow:
3366 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3367 // errno like libm sqrt would.
3368 case Intrinsic::sqrt:
3369 case Intrinsic::fma:
3370 case Intrinsic::fmuladd:
3371 case Intrinsic::fabs:
3372 case Intrinsic::minnum:
3373 case Intrinsic::maxnum:
3375 // TODO: some fp intrinsics are marked as having the same error handling
3376 // as libm. They're safe to speculate when they won't error.
3377 // TODO: are convert_{from,to}_fp16 safe?
3378 // TODO: can we list target-specific intrinsics here?
3382 return false; // The called function could have undefined behavior or
3383 // side-effects, even if marked readnone nounwind.
3385 case Instruction::VAArg:
3386 case Instruction::Alloca:
3387 case Instruction::Invoke:
3388 case Instruction::PHI:
3389 case Instruction::Store:
3390 case Instruction::Ret:
3391 case Instruction::Br:
3392 case Instruction::IndirectBr:
3393 case Instruction::Switch:
3394 case Instruction::Unreachable:
3395 case Instruction::Fence:
3396 case Instruction::AtomicRMW:
3397 case Instruction::AtomicCmpXchg:
3398 case Instruction::LandingPad:
3399 case Instruction::Resume:
3400 case Instruction::CatchPad:
3401 case Instruction::CatchEndPad:
3402 case Instruction::CatchRet:
3403 case Instruction::CleanupPad:
3404 case Instruction::CleanupEndPad:
3405 case Instruction::CleanupRet:
3406 case Instruction::TerminatePad:
3407 return false; // Misc instructions which have effects
3411 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3412 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3415 /// Return true if we know that the specified value is never null.
3416 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3417 assert(V->getType()->isPointerTy() && "V must be pointer type");
3419 // Alloca never returns null, malloc might.
3420 if (isa<AllocaInst>(V)) return true;
3422 // A byval, inalloca, or nonnull argument is never null.
3423 if (const Argument *A = dyn_cast<Argument>(V))
3424 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3426 // A global variable in address space 0 is non null unless extern weak.
3427 // Other address spaces may have null as a valid address for a global,
3428 // so we can't assume anything.
3429 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3430 return !GV->hasExternalWeakLinkage() &&
3431 GV->getType()->getAddressSpace() == 0;
3433 // A Load tagged w/nonnull metadata is never null.
3434 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3435 return LI->getMetadata(LLVMContext::MD_nonnull);
3437 if (auto CS = ImmutableCallSite(V))
3438 if (CS.isReturnNonNull())
3441 // operator new never returns null.
3442 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3448 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3449 const Instruction *CtxI,
3450 const DominatorTree *DT) {
3451 assert(V->getType()->isPointerTy() && "V must be pointer type");
3453 unsigned NumUsesExplored = 0;
3454 for (auto U : V->users()) {
3455 // Avoid massive lists
3456 if (NumUsesExplored >= DomConditionsMaxUses)
3459 // Consider only compare instructions uniquely controlling a branch
3460 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3464 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3467 for (auto *CmpU : Cmp->users()) {
3468 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3472 assert(BI->isConditional() && "uses a comparison!");
3474 BasicBlock *NonNullSuccessor = nullptr;
3475 CmpInst::Predicate Pred;
3477 if (match(const_cast<ICmpInst*>(Cmp),
3478 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3479 if (Pred == ICmpInst::ICMP_EQ)
3480 NonNullSuccessor = BI->getSuccessor(1);
3481 else if (Pred == ICmpInst::ICMP_NE)
3482 NonNullSuccessor = BI->getSuccessor(0);
3485 if (NonNullSuccessor) {
3486 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3487 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3496 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3497 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3498 if (isKnownNonNull(V, TLI))
3501 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3504 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3505 const DataLayout &DL,
3506 AssumptionCache *AC,
3507 const Instruction *CxtI,
3508 const DominatorTree *DT) {
3509 // Multiplying n * m significant bits yields a result of n + m significant
3510 // bits. If the total number of significant bits does not exceed the
3511 // result bit width (minus 1), there is no overflow.
3512 // This means if we have enough leading zero bits in the operands
3513 // we can guarantee that the result does not overflow.
3514 // Ref: "Hacker's Delight" by Henry Warren
3515 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3516 APInt LHSKnownZero(BitWidth, 0);
3517 APInt LHSKnownOne(BitWidth, 0);
3518 APInt RHSKnownZero(BitWidth, 0);
3519 APInt RHSKnownOne(BitWidth, 0);
3520 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3522 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3524 // Note that underestimating the number of zero bits gives a more
3525 // conservative answer.
3526 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3527 RHSKnownZero.countLeadingOnes();
3528 // First handle the easy case: if we have enough zero bits there's
3529 // definitely no overflow.
3530 if (ZeroBits >= BitWidth)
3531 return OverflowResult::NeverOverflows;
3533 // Get the largest possible values for each operand.
3534 APInt LHSMax = ~LHSKnownZero;
3535 APInt RHSMax = ~RHSKnownZero;
3537 // We know the multiply operation doesn't overflow if the maximum values for
3538 // each operand will not overflow after we multiply them together.
3540 LHSMax.umul_ov(RHSMax, MaxOverflow);
3542 return OverflowResult::NeverOverflows;
3544 // We know it always overflows if multiplying the smallest possible values for
3545 // the operands also results in overflow.
3547 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3549 return OverflowResult::AlwaysOverflows;
3551 return OverflowResult::MayOverflow;
3554 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3555 const DataLayout &DL,
3556 AssumptionCache *AC,
3557 const Instruction *CxtI,
3558 const DominatorTree *DT) {
3559 bool LHSKnownNonNegative, LHSKnownNegative;
3560 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3562 if (LHSKnownNonNegative || LHSKnownNegative) {
3563 bool RHSKnownNonNegative, RHSKnownNegative;
3564 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3567 if (LHSKnownNegative && RHSKnownNegative) {
3568 // The sign bit is set in both cases: this MUST overflow.
3569 // Create a simple add instruction, and insert it into the struct.
3570 return OverflowResult::AlwaysOverflows;
3573 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3574 // The sign bit is clear in both cases: this CANNOT overflow.
3575 // Create a simple add instruction, and insert it into the struct.
3576 return OverflowResult::NeverOverflows;
3580 return OverflowResult::MayOverflow;
3583 static OverflowResult computeOverflowForSignedAdd(
3584 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3585 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3586 if (Add && Add->hasNoSignedWrap()) {
3587 return OverflowResult::NeverOverflows;
3590 bool LHSKnownNonNegative, LHSKnownNegative;
3591 bool RHSKnownNonNegative, RHSKnownNegative;
3592 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3594 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3597 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3598 (LHSKnownNegative && RHSKnownNonNegative)) {
3599 // The sign bits are opposite: this CANNOT overflow.
3600 return OverflowResult::NeverOverflows;
3603 // The remaining code needs Add to be available. Early returns if not so.
3605 return OverflowResult::MayOverflow;
3607 // If the sign of Add is the same as at least one of the operands, this add
3608 // CANNOT overflow. This is particularly useful when the sum is
3609 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3611 bool LHSOrRHSKnownNonNegative =
3612 (LHSKnownNonNegative || RHSKnownNonNegative);
3613 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3614 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3615 bool AddKnownNonNegative, AddKnownNegative;
3616 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3617 /*Depth=*/0, AC, CxtI, DT);
3618 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3619 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3620 return OverflowResult::NeverOverflows;
3624 return OverflowResult::MayOverflow;
3627 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3628 const DataLayout &DL,
3629 AssumptionCache *AC,
3630 const Instruction *CxtI,
3631 const DominatorTree *DT) {
3632 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3633 Add, DL, AC, CxtI, DT);
3636 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3637 const DataLayout &DL,
3638 AssumptionCache *AC,
3639 const Instruction *CxtI,
3640 const DominatorTree *DT) {
3641 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3644 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3645 // FIXME: This conservative implementation can be relaxed. E.g. most
3646 // atomic operations are guaranteed to terminate on most platforms
3647 // and most functions terminate.
3649 return !I->isAtomic() && // atomics may never succeed on some platforms
3650 !isa<CallInst>(I) && // could throw and might not terminate
3651 !isa<InvokeInst>(I) && // might not terminate and could throw to
3652 // non-successor (see bug 24185 for details).
3653 !isa<ResumeInst>(I) && // has no successors
3654 !isa<ReturnInst>(I); // has no successors
3657 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3659 // The loop header is guaranteed to be executed for every iteration.
3661 // FIXME: Relax this constraint to cover all basic blocks that are
3662 // guaranteed to be executed at every iteration.
3663 if (I->getParent() != L->getHeader()) return false;
3665 for (const Instruction &LI : *L->getHeader()) {
3666 if (&LI == I) return true;
3667 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3669 llvm_unreachable("Instruction not contained in its own parent basic block.");
3672 bool llvm::propagatesFullPoison(const Instruction *I) {
3673 switch (I->getOpcode()) {
3674 case Instruction::Add:
3675 case Instruction::Sub:
3676 case Instruction::Xor:
3677 case Instruction::Trunc:
3678 case Instruction::BitCast:
3679 case Instruction::AddrSpaceCast:
3680 // These operations all propagate poison unconditionally. Note that poison
3681 // is not any particular value, so xor or subtraction of poison with
3682 // itself still yields poison, not zero.
3685 case Instruction::AShr:
3686 case Instruction::SExt:
3687 // For these operations, one bit of the input is replicated across
3688 // multiple output bits. A replicated poison bit is still poison.
3691 case Instruction::Shl: {
3692 // Left shift *by* a poison value is poison. The number of
3693 // positions to shift is unsigned, so no negative values are
3694 // possible there. Left shift by zero places preserves poison. So
3695 // it only remains to consider left shift of poison by a positive
3696 // number of places.
3698 // A left shift by a positive number of places leaves the lowest order bit
3699 // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3700 // make the poison operand violate that flag, yielding a fresh full-poison
3702 auto *OBO = cast<OverflowingBinaryOperator>(I);
3703 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3706 case Instruction::Mul: {
3707 // A multiplication by zero yields a non-poison zero result, so we need to
3708 // rule out zero as an operand. Conservatively, multiplication by a
3709 // non-zero constant is not multiplication by zero.
3711 // Multiplication by a non-zero constant can leave some bits
3712 // non-poisoned. For example, a multiplication by 2 leaves the lowest
3713 // order bit unpoisoned. So we need to consider that.
3715 // Multiplication by 1 preserves poison. If the multiplication has a
3716 // no-wrap flag, then we can make the poison operand violate that flag
3717 // when multiplied by any integer other than 0 and 1.
3718 auto *OBO = cast<OverflowingBinaryOperator>(I);
3719 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3720 for (Value *V : OBO->operands()) {
3721 if (auto *CI = dyn_cast<ConstantInt>(V)) {
3722 // A ConstantInt cannot yield poison, so we can assume that it is
3723 // the other operand that is poison.
3724 return !CI->isZero();
3731 case Instruction::GetElementPtr:
3732 // A GEP implicitly represents a sequence of additions, subtractions,
3733 // truncations, sign extensions and multiplications. The multiplications
3734 // are by the non-zero sizes of some set of types, so we do not have to be
3735 // concerned with multiplication by zero. If the GEP is in-bounds, then
3736 // these operations are implicitly no-signed-wrap so poison is propagated
3737 // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3738 return cast<GEPOperator>(I)->isInBounds();
3745 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3746 switch (I->getOpcode()) {
3747 case Instruction::Store:
3748 return cast<StoreInst>(I)->getPointerOperand();
3750 case Instruction::Load:
3751 return cast<LoadInst>(I)->getPointerOperand();
3753 case Instruction::AtomicCmpXchg:
3754 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3756 case Instruction::AtomicRMW:
3757 return cast<AtomicRMWInst>(I)->getPointerOperand();
3759 case Instruction::UDiv:
3760 case Instruction::SDiv:
3761 case Instruction::URem:
3762 case Instruction::SRem:
3763 return I->getOperand(1);
3770 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3771 // We currently only look for uses of poison values within the same basic
3772 // block, as that makes it easier to guarantee that the uses will be
3773 // executed given that PoisonI is executed.
3775 // FIXME: Expand this to consider uses beyond the same basic block. To do
3776 // this, look out for the distinction between post-dominance and strong
3778 const BasicBlock *BB = PoisonI->getParent();
3780 // Set of instructions that we have proved will yield poison if PoisonI
3782 SmallSet<const Value *, 16> YieldsPoison;
3783 YieldsPoison.insert(PoisonI);
3785 for (BasicBlock::const_iterator I = PoisonI->getIterator(), E = BB->end();
3787 if (&*I != PoisonI) {
3788 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&*I);
3789 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) return true;
3790 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
3794 // Mark poison that propagates from I through uses of I.
3795 if (YieldsPoison.count(&*I)) {
3796 for (const User *User : I->users()) {
3797 const Instruction *UserI = cast<Instruction>(User);
3798 if (UserI->getParent() == BB && propagatesFullPoison(UserI))
3799 YieldsPoison.insert(User);
3806 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3810 if (auto *C = dyn_cast<ConstantFP>(V))
3815 static bool isKnownNonZero(Value *V) {
3816 if (auto *C = dyn_cast<ConstantFP>(V))
3817 return !C->isZero();
3821 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3823 Value *CmpLHS, Value *CmpRHS,
3824 Value *TrueVal, Value *FalseVal,
3825 Value *&LHS, Value *&RHS) {
3829 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3830 // return inconsistent results between implementations.
3831 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3832 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3833 // Therefore we behave conservatively and only proceed if at least one of the
3834 // operands is known to not be zero, or if we don't care about signed zeroes.
3837 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3838 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3839 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3840 !isKnownNonZero(CmpRHS))
3841 return {SPF_UNKNOWN, SPNB_NA, false};
3844 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3845 bool Ordered = false;
3847 // When given one NaN and one non-NaN input:
3848 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3849 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3850 // ordered comparison fails), which could be NaN or non-NaN.
3851 // so here we discover exactly what NaN behavior is required/accepted.
3852 if (CmpInst::isFPPredicate(Pred)) {
3853 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3854 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3856 if (LHSSafe && RHSSafe) {
3857 // Both operands are known non-NaN.
3858 NaNBehavior = SPNB_RETURNS_ANY;
3859 } else if (CmpInst::isOrdered(Pred)) {
3860 // An ordered comparison will return false when given a NaN, so it
3864 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3865 NaNBehavior = SPNB_RETURNS_NAN;
3867 NaNBehavior = SPNB_RETURNS_OTHER;
3869 // Completely unsafe.
3870 return {SPF_UNKNOWN, SPNB_NA, false};
3873 // An unordered comparison will return true when given a NaN, so it
3876 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3877 NaNBehavior = SPNB_RETURNS_OTHER;
3879 NaNBehavior = SPNB_RETURNS_NAN;
3881 // Completely unsafe.
3882 return {SPF_UNKNOWN, SPNB_NA, false};
3886 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3887 std::swap(CmpLHS, CmpRHS);
3888 Pred = CmpInst::getSwappedPredicate(Pred);
3889 if (NaNBehavior == SPNB_RETURNS_NAN)
3890 NaNBehavior = SPNB_RETURNS_OTHER;
3891 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3892 NaNBehavior = SPNB_RETURNS_NAN;
3896 // ([if]cmp X, Y) ? X : Y
3897 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3899 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3900 case ICmpInst::ICMP_UGT:
3901 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3902 case ICmpInst::ICMP_SGT:
3903 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3904 case ICmpInst::ICMP_ULT:
3905 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3906 case ICmpInst::ICMP_SLT:
3907 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3908 case FCmpInst::FCMP_UGT:
3909 case FCmpInst::FCMP_UGE:
3910 case FCmpInst::FCMP_OGT:
3911 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3912 case FCmpInst::FCMP_ULT:
3913 case FCmpInst::FCMP_ULE:
3914 case FCmpInst::FCMP_OLT:
3915 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3919 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3920 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3921 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3923 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3924 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3925 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3926 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3929 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3930 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3931 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3932 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3936 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3937 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3938 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3939 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3940 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3943 return {SPF_SMIN, SPNB_NA, false};
3948 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3950 return {SPF_UNKNOWN, SPNB_NA, false};
3953 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
3954 Instruction::CastOps *CastOp) {
3955 CastInst *CI = dyn_cast<CastInst>(V1);
3956 Constant *C = dyn_cast<Constant>(V2);
3957 CastInst *CI2 = dyn_cast<CastInst>(V2);
3960 *CastOp = CI->getOpcode();
3963 // If V1 and V2 are both the same cast from the same type, we can look
3965 if (CI2->getOpcode() == CI->getOpcode() &&
3966 CI2->getSrcTy() == CI->getSrcTy())
3967 return CI2->getOperand(0);
3973 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3974 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3975 // This is only valid if the truncated value can be sign-extended
3976 // back to the original value.
3977 if (ConstantExpr::getSExt(T, C->getType()) == C)
3981 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3982 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3984 if (isa<TruncInst>(CI))
3985 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3987 if (isa<FPToUIInst>(CI))
3988 return ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
3990 if (isa<FPToSIInst>(CI))
3991 return ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
3993 if (isa<UIToFPInst>(CI))
3994 return ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
3996 if (isa<SIToFPInst>(CI))
3997 return ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
3999 if (isa<FPTruncInst>(CI))
4000 return ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
4002 if (isa<FPExtInst>(CI))
4003 return ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
4008 SelectPatternResult llvm::matchSelectPattern(Value *V,
4009 Value *&LHS, Value *&RHS,
4010 Instruction::CastOps *CastOp) {
4011 SelectInst *SI = dyn_cast<SelectInst>(V);
4012 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4014 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4015 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4017 CmpInst::Predicate Pred = CmpI->getPredicate();
4018 Value *CmpLHS = CmpI->getOperand(0);
4019 Value *CmpRHS = CmpI->getOperand(1);
4020 Value *TrueVal = SI->getTrueValue();
4021 Value *FalseVal = SI->getFalseValue();
4023 if (isa<FPMathOperator>(CmpI))
4024 FMF = CmpI->getFastMathFlags();
4027 if (CmpI->isEquality())
4028 return {SPF_UNKNOWN, SPNB_NA, false};
4030 // Deal with type mismatches.
4031 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4032 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
4033 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4034 cast<CastInst>(TrueVal)->getOperand(0), C,
4036 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
4037 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4038 C, cast<CastInst>(FalseVal)->getOperand(0),
4041 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4045 ConstantRange llvm::getConstantRangeFromMetadata(MDNode &Ranges) {
4046 const unsigned NumRanges = Ranges.getNumOperands() / 2;
4047 assert(NumRanges >= 1 && "Must have at least one range!");
4048 assert(Ranges.getNumOperands() % 2 == 0 && "Must be a sequence of pairs");
4050 auto *FirstLow = mdconst::extract<ConstantInt>(Ranges.getOperand(0));
4051 auto *FirstHigh = mdconst::extract<ConstantInt>(Ranges.getOperand(1));
4053 ConstantRange CR(FirstLow->getValue(), FirstHigh->getValue());
4055 for (unsigned i = 1; i < NumRanges; ++i) {
4056 auto *Low = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
4057 auto *High = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
4059 // Note: unionWith will potentially create a range that contains values not
4060 // contained in any of the original N ranges.
4061 CR = CR.unionWith(ConstantRange(Low->getValue(), High->getValue()));