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 if (std::all_of(V->user_begin(), V->user_end(),
409 [&](const User *U) { return EphValues.count(U); })) {
414 if (const User *U = dyn_cast<User>(V))
415 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
417 if (isSafeToSpeculativelyExecute(*J))
418 WorkSet.push_back(*J);
426 // Is this an intrinsic that cannot be speculated but also cannot trap?
427 static bool isAssumeLikeIntrinsic(const Instruction *I) {
428 if (const CallInst *CI = dyn_cast<CallInst>(I))
429 if (Function *F = CI->getCalledFunction())
430 switch (F->getIntrinsicID()) {
432 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
433 case Intrinsic::assume:
434 case Intrinsic::dbg_declare:
435 case Intrinsic::dbg_value:
436 case Intrinsic::invariant_start:
437 case Intrinsic::invariant_end:
438 case Intrinsic::lifetime_start:
439 case Intrinsic::lifetime_end:
440 case Intrinsic::objectsize:
441 case Intrinsic::ptr_annotation:
442 case Intrinsic::var_annotation:
449 static bool isValidAssumeForContext(Value *V, const Query &Q) {
450 Instruction *Inv = cast<Instruction>(V);
452 // There are two restrictions on the use of an assume:
453 // 1. The assume must dominate the context (or the control flow must
454 // reach the assume whenever it reaches the context).
455 // 2. The context must not be in the assume's set of ephemeral values
456 // (otherwise we will use the assume to prove that the condition
457 // feeding the assume is trivially true, thus causing the removal of
461 if (Q.DT->dominates(Inv, Q.CxtI)) {
463 } else if (Inv->getParent() == Q.CxtI->getParent()) {
464 // The context comes first, but they're both in the same block. Make sure
465 // there is nothing in between that might interrupt the control flow.
466 for (BasicBlock::const_iterator I =
467 std::next(BasicBlock::const_iterator(Q.CxtI)),
468 IE(Inv); I != IE; ++I)
469 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
472 return !isEphemeralValueOf(Inv, Q.CxtI);
478 // When we don't have a DT, we do a limited search...
479 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
481 } else if (Inv->getParent() == Q.CxtI->getParent()) {
482 // Search forward from the assume until we reach the context (or the end
483 // of the block); the common case is that the assume will come first.
484 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
485 IE = Inv->getParent()->end(); I != IE; ++I)
489 // The context must come first...
490 for (BasicBlock::const_iterator I =
491 std::next(BasicBlock::const_iterator(Q.CxtI)),
492 IE(Inv); I != IE; ++I)
493 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
496 return !isEphemeralValueOf(Inv, Q.CxtI);
502 bool llvm::isValidAssumeForContext(const Instruction *I,
503 const Instruction *CxtI,
504 const DominatorTree *DT) {
505 return ::isValidAssumeForContext(const_cast<Instruction *>(I),
506 Query(nullptr, CxtI, DT));
509 template<typename LHS, typename RHS>
510 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
511 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
512 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
513 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
516 template<typename LHS, typename RHS>
517 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
518 BinaryOp_match<RHS, LHS, Instruction::And>>
519 m_c_And(const LHS &L, const RHS &R) {
520 return m_CombineOr(m_And(L, R), m_And(R, L));
523 template<typename LHS, typename RHS>
524 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
525 BinaryOp_match<RHS, LHS, Instruction::Or>>
526 m_c_Or(const LHS &L, const RHS &R) {
527 return m_CombineOr(m_Or(L, R), m_Or(R, L));
530 template<typename LHS, typename RHS>
531 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
532 BinaryOp_match<RHS, LHS, Instruction::Xor>>
533 m_c_Xor(const LHS &L, const RHS &R) {
534 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
537 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
538 /// true (at the context instruction.) This is mostly a utility function for
539 /// the prototype dominating conditions reasoning below.
540 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
543 const DataLayout &DL,
544 unsigned Depth, const Query &Q) {
545 Value *LHS = Cmp->getOperand(0);
546 Value *RHS = Cmp->getOperand(1);
547 // TODO: We could potentially be more aggressive here. This would be worth
548 // evaluating. If we can, explore commoning this code with the assume
550 if (LHS != V && RHS != V)
553 const unsigned BitWidth = KnownZero.getBitWidth();
555 switch (Cmp->getPredicate()) {
557 // We know nothing from this condition
559 // TODO: implement unsigned bound from below (known one bits)
560 // TODO: common condition check implementations with assumes
561 // TODO: implement other patterns from assume (e.g. V & B == A)
562 case ICmpInst::ICMP_SGT:
564 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
565 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
566 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
567 // We know that the sign bit is zero.
568 KnownZero |= APInt::getSignBit(BitWidth);
572 case ICmpInst::ICMP_EQ:
574 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
576 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
578 computeKnownBits(LHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
580 llvm_unreachable("missing use?");
581 KnownZero |= KnownZeroTemp;
582 KnownOne |= KnownOneTemp;
585 case ICmpInst::ICMP_ULE:
587 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
588 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
589 // The known zero bits carry over
590 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
591 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
594 case ICmpInst::ICMP_ULT:
596 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
597 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
598 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
599 // power of 2, then one more).
600 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
601 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
603 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
609 /// Compute known bits in 'V' from conditions which are known to be true along
610 /// all paths leading to the context instruction. In particular, look for
611 /// cases where one branch of an interesting condition dominates the context
612 /// instruction. This does not do general dataflow.
613 /// NOTE: This code is EXPERIMENTAL and currently off by default.
614 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
616 const DataLayout &DL,
619 // Need both the dominator tree and the query location to do anything useful
620 if (!Q.DT || !Q.CxtI)
622 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
623 // The context instruction might be in a statically unreachable block. If
624 // so, asking dominator queries may yield suprising results. (e.g. the block
625 // may not have a dom tree node)
626 if (!Q.DT->isReachableFromEntry(Cxt->getParent()))
629 // Avoid useless work
630 if (auto VI = dyn_cast<Instruction>(V))
631 if (VI->getParent() == Cxt->getParent())
634 // Note: We currently implement two options. It's not clear which of these
635 // will survive long term, we need data for that.
636 // Option 1 - Try walking the dominator tree looking for conditions which
637 // might apply. This works well for local conditions (loop guards, etc..),
638 // but not as well for things far from the context instruction (presuming a
639 // low max blocks explored). If we can set an high enough limit, this would
641 // Option 2 - We restrict out search to those conditions which are uses of
642 // the value we're interested in. This is independent of dom structure,
643 // but is slightly less powerful without looking through lots of use chains.
644 // It does handle conditions far from the context instruction (e.g. early
645 // function exits on entry) really well though.
647 // Option 1 - Search the dom tree
648 unsigned NumBlocksExplored = 0;
649 BasicBlock *Current = Cxt->getParent();
651 // Stop searching if we've gone too far up the chain
652 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
656 if (!Q.DT->getNode(Current)->getIDom())
658 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
660 // found function entry
663 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
664 if (!BI || BI->isUnconditional())
666 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
670 // We're looking for conditions that are guaranteed to hold at the context
671 // instruction. Finding a condition where one path dominates the context
672 // isn't enough because both the true and false cases could merge before
673 // the context instruction we're actually interested in. Instead, we need
674 // to ensure that the taken *edge* dominates the context instruction. We
675 // know that the edge must be reachable since we started from a reachable
677 BasicBlock *BB0 = BI->getSuccessor(0);
678 BasicBlockEdge Edge(BI->getParent(), BB0);
679 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
682 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
686 // Option 2 - Search the other uses of V
687 unsigned NumUsesExplored = 0;
688 for (auto U : V->users()) {
689 // Avoid massive lists
690 if (NumUsesExplored >= DomConditionsMaxUses)
693 // Consider only compare instructions uniquely controlling a branch
694 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
698 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
701 for (auto *CmpU : Cmp->users()) {
702 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
703 if (!BI || BI->isUnconditional())
705 // We're looking for conditions that are guaranteed to hold at the
706 // context instruction. Finding a condition where one path dominates
707 // the context isn't enough because both the true and false cases could
708 // merge before the context instruction we're actually interested in.
709 // Instead, we need to ensure that the taken *edge* dominates the context
711 BasicBlock *BB0 = BI->getSuccessor(0);
712 BasicBlockEdge Edge(BI->getParent(), BB0);
713 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
716 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
722 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
723 APInt &KnownOne, const DataLayout &DL,
724 unsigned Depth, const Query &Q) {
725 // Use of assumptions is context-sensitive. If we don't have a context, we
727 if (!Q.AC || !Q.CxtI)
730 unsigned BitWidth = KnownZero.getBitWidth();
732 for (auto &AssumeVH : Q.AC->assumptions()) {
735 CallInst *I = cast<CallInst>(AssumeVH);
736 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
737 "Got assumption for the wrong function!");
738 if (Q.ExclInvs.count(I))
741 // Warning: This loop can end up being somewhat performance sensetive.
742 // We're running this loop for once for each value queried resulting in a
743 // runtime of ~O(#assumes * #values).
745 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
746 "must be an assume intrinsic");
748 Value *Arg = I->getArgOperand(0);
750 if (Arg == V && isValidAssumeForContext(I, Q)) {
751 assert(BitWidth == 1 && "assume operand is not i1?");
752 KnownZero.clearAllBits();
753 KnownOne.setAllBits();
757 // The remaining tests are all recursive, so bail out if we hit the limit.
758 if (Depth == MaxDepth)
762 auto m_V = m_CombineOr(m_Specific(V),
763 m_CombineOr(m_PtrToInt(m_Specific(V)),
764 m_BitCast(m_Specific(V))));
766 CmpInst::Predicate Pred;
769 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
770 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
771 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
772 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
773 KnownZero |= RHSKnownZero;
774 KnownOne |= RHSKnownOne;
776 } else if (match(Arg,
777 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
778 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
779 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
780 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
781 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
782 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
784 // For those bits in the mask that are known to be one, we can propagate
785 // known bits from the RHS to V.
786 KnownZero |= RHSKnownZero & MaskKnownOne;
787 KnownOne |= RHSKnownOne & MaskKnownOne;
788 // assume(~(v & b) = a)
789 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
791 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
792 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
793 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
794 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
795 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
797 // For those bits in the mask that are known to be one, we can propagate
798 // inverted known bits from the RHS to V.
799 KnownZero |= RHSKnownOne & MaskKnownOne;
800 KnownOne |= RHSKnownZero & MaskKnownOne;
802 } else if (match(Arg,
803 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
804 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
805 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
806 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
807 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
808 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
810 // For those bits in B that are known to be zero, we can propagate known
811 // bits from the RHS to V.
812 KnownZero |= RHSKnownZero & BKnownZero;
813 KnownOne |= RHSKnownOne & BKnownZero;
814 // assume(~(v | b) = a)
815 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
817 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
818 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
819 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
820 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
821 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
823 // For those bits in B that are known to be zero, we can propagate
824 // inverted known bits from the RHS to V.
825 KnownZero |= RHSKnownOne & BKnownZero;
826 KnownOne |= RHSKnownZero & BKnownZero;
828 } else if (match(Arg,
829 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
830 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
831 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
832 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
833 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
834 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
836 // For those bits in B that are known to be zero, we can propagate known
837 // bits from the RHS to V. For those bits in B that are known to be one,
838 // we can propagate inverted known bits from the RHS to V.
839 KnownZero |= RHSKnownZero & BKnownZero;
840 KnownOne |= RHSKnownOne & BKnownZero;
841 KnownZero |= RHSKnownOne & BKnownOne;
842 KnownOne |= RHSKnownZero & BKnownOne;
843 // assume(~(v ^ b) = a)
844 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
846 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
847 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
848 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
849 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
850 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
852 // For those bits in B that are known to be zero, we can propagate
853 // inverted known bits from the RHS to V. For those bits in B that are
854 // known to be one, we can propagate known bits from the RHS to V.
855 KnownZero |= RHSKnownOne & BKnownZero;
856 KnownOne |= RHSKnownZero & BKnownZero;
857 KnownZero |= RHSKnownZero & BKnownOne;
858 KnownOne |= RHSKnownOne & BKnownOne;
859 // assume(v << c = a)
860 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
862 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
863 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
864 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
865 // For those bits in RHS that are known, we can propagate them to known
866 // bits in V shifted to the right by C.
867 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
868 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
869 // assume(~(v << c) = a)
870 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
872 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
873 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
874 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
875 // For those bits in RHS that are known, we can propagate them inverted
876 // to known bits in V shifted to the right by C.
877 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
878 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
879 // assume(v >> c = a)
880 } else if (match(Arg,
881 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
882 m_AShr(m_V, m_ConstantInt(C))),
884 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
885 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
886 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
887 // For those bits in RHS that are known, we can propagate them to known
888 // bits in V shifted to the right by C.
889 KnownZero |= RHSKnownZero << C->getZExtValue();
890 KnownOne |= RHSKnownOne << C->getZExtValue();
891 // assume(~(v >> c) = a)
892 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
893 m_LShr(m_V, m_ConstantInt(C)),
894 m_AShr(m_V, m_ConstantInt(C)))),
896 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
897 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
898 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
899 // For those bits in RHS that are known, we can propagate them inverted
900 // to known bits in V shifted to the right by C.
901 KnownZero |= RHSKnownOne << C->getZExtValue();
902 KnownOne |= RHSKnownZero << C->getZExtValue();
903 // assume(v >=_s c) where c is non-negative
904 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
905 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
906 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
907 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
909 if (RHSKnownZero.isNegative()) {
910 // We know that the sign bit is zero.
911 KnownZero |= APInt::getSignBit(BitWidth);
913 // assume(v >_s c) where c is at least -1.
914 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
915 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
916 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
917 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
919 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
920 // We know that the sign bit is zero.
921 KnownZero |= APInt::getSignBit(BitWidth);
923 // assume(v <=_s c) where c is negative
924 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
925 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
926 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
927 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
929 if (RHSKnownOne.isNegative()) {
930 // We know that the sign bit is one.
931 KnownOne |= APInt::getSignBit(BitWidth);
933 // assume(v <_s c) where c is non-positive
934 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
935 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
936 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
937 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
939 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
940 // We know that the sign bit is one.
941 KnownOne |= APInt::getSignBit(BitWidth);
944 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
945 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
946 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
947 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
949 // Whatever high bits in c are zero are known to be zero.
951 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
953 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
954 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
955 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
956 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
958 // Whatever high bits in c are zero are known to be zero (if c is a power
959 // of 2, then one more).
960 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
962 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
965 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
970 // Compute known bits from a shift operator, including those with a
971 // non-constant shift amount. KnownZero and KnownOne are the outputs of this
972 // function. KnownZero2 and KnownOne2 are pre-allocated temporaries with the
973 // same bit width as KnownZero and KnownOne. KZF and KOF are operator-specific
974 // functors that, given the known-zero or known-one bits respectively, and a
975 // shift amount, compute the implied known-zero or known-one bits of the shift
976 // operator's result respectively for that shift amount. The results from calling
977 // KZF and KOF are conservatively combined for all permitted shift amounts.
978 template <typename KZFunctor, typename KOFunctor>
979 static void computeKnownBitsFromShiftOperator(Operator *I,
980 APInt &KnownZero, APInt &KnownOne,
981 APInt &KnownZero2, APInt &KnownOne2,
982 const DataLayout &DL, unsigned Depth, const Query &Q,
983 KZFunctor KZF, KOFunctor KOF) {
984 unsigned BitWidth = KnownZero.getBitWidth();
986 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
987 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
989 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
990 KnownZero = KZF(KnownZero, ShiftAmt);
991 KnownOne = KOF(KnownOne, ShiftAmt);
995 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
997 // Note: We cannot use KnownZero.getLimitedValue() here, because if
998 // BitWidth > 64 and any upper bits are known, we'll end up returning the
999 // limit value (which implies all bits are known).
1000 uint64_t ShiftAmtKZ = KnownZero.zextOrTrunc(64).getZExtValue();
1001 uint64_t ShiftAmtKO = KnownOne.zextOrTrunc(64).getZExtValue();
1003 // It would be more-clearly correct to use the two temporaries for this
1004 // calculation. Reusing the APInts here to prevent unnecessary allocations.
1005 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1007 // Early exit if we can't constrain any well-defined shift amount.
1008 if (!(ShiftAmtKZ & (BitWidth-1)) && !(ShiftAmtKO & (BitWidth-1)))
1011 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1013 KnownZero = KnownOne = APInt::getAllOnesValue(BitWidth);
1014 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
1015 // Combine the shifted known input bits only for those shift amounts
1016 // compatible with its known constraints.
1017 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
1019 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
1022 KnownZero &= KZF(KnownZero2, ShiftAmt);
1023 KnownOne &= KOF(KnownOne2, ShiftAmt);
1026 // If there are no compatible shift amounts, then we've proven that the shift
1027 // amount must be >= the BitWidth, and the result is undefined. We could
1028 // return anything we'd like, but we need to make sure the sets of known bits
1029 // stay disjoint (it should be better for some other code to actually
1030 // propagate the undef than to pick a value here using known bits).
1031 if ((KnownZero & KnownOne) != 0)
1032 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1035 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
1036 APInt &KnownOne, const DataLayout &DL,
1037 unsigned Depth, const Query &Q) {
1038 unsigned BitWidth = KnownZero.getBitWidth();
1040 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
1041 switch (I->getOpcode()) {
1043 case Instruction::Load:
1044 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
1045 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1047 case Instruction::And: {
1048 // If either the LHS or the RHS are Zero, the result is zero.
1049 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1050 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1052 // Output known-1 bits are only known if set in both the LHS & RHS.
1053 KnownOne &= KnownOne2;
1054 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1055 KnownZero |= KnownZero2;
1058 case Instruction::Or: {
1059 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1060 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1062 // Output known-0 bits are only known if clear in both the LHS & RHS.
1063 KnownZero &= KnownZero2;
1064 // Output known-1 are known to be set if set in either the LHS | RHS.
1065 KnownOne |= KnownOne2;
1068 case Instruction::Xor: {
1069 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1070 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1072 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1073 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1074 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1075 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1076 KnownZero = KnownZeroOut;
1079 case Instruction::Mul: {
1080 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1081 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1082 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1085 case Instruction::UDiv: {
1086 // For the purposes of computing leading zeros we can conservatively
1087 // treat a udiv as a logical right shift by the power of 2 known to
1088 // be less than the denominator.
1089 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1090 unsigned LeadZ = KnownZero2.countLeadingOnes();
1092 KnownOne2.clearAllBits();
1093 KnownZero2.clearAllBits();
1094 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1095 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1096 if (RHSUnknownLeadingOnes != BitWidth)
1097 LeadZ = std::min(BitWidth,
1098 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1100 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1103 case Instruction::Select:
1104 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1105 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1107 // Only known if known in both the LHS and RHS.
1108 KnownOne &= KnownOne2;
1109 KnownZero &= KnownZero2;
1111 case Instruction::FPTrunc:
1112 case Instruction::FPExt:
1113 case Instruction::FPToUI:
1114 case Instruction::FPToSI:
1115 case Instruction::SIToFP:
1116 case Instruction::UIToFP:
1117 break; // Can't work with floating point.
1118 case Instruction::PtrToInt:
1119 case Instruction::IntToPtr:
1120 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1121 // FALL THROUGH and handle them the same as zext/trunc.
1122 case Instruction::ZExt:
1123 case Instruction::Trunc: {
1124 Type *SrcTy = I->getOperand(0)->getType();
1126 unsigned SrcBitWidth;
1127 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1128 // which fall through here.
1129 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1131 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1132 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1133 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1134 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1135 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1136 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1137 // Any top bits are known to be zero.
1138 if (BitWidth > SrcBitWidth)
1139 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1142 case Instruction::BitCast: {
1143 Type *SrcTy = I->getOperand(0)->getType();
1144 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy() ||
1145 SrcTy->isFloatingPointTy()) &&
1146 // TODO: For now, not handling conversions like:
1147 // (bitcast i64 %x to <2 x i32>)
1148 !I->getType()->isVectorTy()) {
1149 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1154 case Instruction::SExt: {
1155 // Compute the bits in the result that are not present in the input.
1156 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1158 KnownZero = KnownZero.trunc(SrcBitWidth);
1159 KnownOne = KnownOne.trunc(SrcBitWidth);
1160 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1161 KnownZero = KnownZero.zext(BitWidth);
1162 KnownOne = KnownOne.zext(BitWidth);
1164 // If the sign bit of the input is known set or clear, then we know the
1165 // top bits of the result.
1166 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1167 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1168 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1169 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1172 case Instruction::Shl: {
1173 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1174 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1175 return (KnownZero << ShiftAmt) |
1176 APInt::getLowBitsSet(BitWidth, ShiftAmt); // Low bits known 0.
1179 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1180 return KnownOne << ShiftAmt;
1183 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1184 KnownZero2, KnownOne2, DL, Depth, Q,
1188 case Instruction::LShr: {
1189 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1190 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1191 return APIntOps::lshr(KnownZero, ShiftAmt) |
1192 // High bits known zero.
1193 APInt::getHighBitsSet(BitWidth, ShiftAmt);
1196 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1197 return APIntOps::lshr(KnownOne, ShiftAmt);
1200 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1201 KnownZero2, KnownOne2, DL, Depth, Q,
1205 case Instruction::AShr: {
1206 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1207 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1208 return APIntOps::ashr(KnownZero, ShiftAmt);
1211 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1212 return APIntOps::ashr(KnownOne, ShiftAmt);
1215 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1216 KnownZero2, KnownOne2, DL, Depth, Q,
1220 case Instruction::Sub: {
1221 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1222 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1223 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1227 case Instruction::Add: {
1228 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1229 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1230 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1234 case Instruction::SRem:
1235 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1236 APInt RA = Rem->getValue().abs();
1237 if (RA.isPowerOf2()) {
1238 APInt LowBits = RA - 1;
1239 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1242 // The low bits of the first operand are unchanged by the srem.
1243 KnownZero = KnownZero2 & LowBits;
1244 KnownOne = KnownOne2 & LowBits;
1246 // If the first operand is non-negative or has all low bits zero, then
1247 // the upper bits are all zero.
1248 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1249 KnownZero |= ~LowBits;
1251 // If the first operand is negative and not all low bits are zero, then
1252 // the upper bits are all one.
1253 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1254 KnownOne |= ~LowBits;
1256 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1260 // The sign bit is the LHS's sign bit, except when the result of the
1261 // remainder is zero.
1262 if (KnownZero.isNonNegative()) {
1263 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1264 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1266 // If it's known zero, our sign bit is also zero.
1267 if (LHSKnownZero.isNegative())
1268 KnownZero.setBit(BitWidth - 1);
1272 case Instruction::URem: {
1273 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1274 APInt RA = Rem->getValue();
1275 if (RA.isPowerOf2()) {
1276 APInt LowBits = (RA - 1);
1277 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1279 KnownZero |= ~LowBits;
1280 KnownOne &= LowBits;
1285 // Since the result is less than or equal to either operand, any leading
1286 // zero bits in either operand must also exist in the result.
1287 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1288 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1290 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1291 KnownZero2.countLeadingOnes());
1292 KnownOne.clearAllBits();
1293 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1297 case Instruction::Alloca: {
1298 AllocaInst *AI = cast<AllocaInst>(I);
1299 unsigned Align = AI->getAlignment();
1301 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1304 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1307 case Instruction::GetElementPtr: {
1308 // Analyze all of the subscripts of this getelementptr instruction
1309 // to determine if we can prove known low zero bits.
1310 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1311 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1313 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1315 gep_type_iterator GTI = gep_type_begin(I);
1316 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1317 Value *Index = I->getOperand(i);
1318 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1319 // Handle struct member offset arithmetic.
1321 // Handle case when index is vector zeroinitializer
1322 Constant *CIndex = cast<Constant>(Index);
1323 if (CIndex->isZeroValue())
1326 if (CIndex->getType()->isVectorTy())
1327 Index = CIndex->getSplatValue();
1329 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1330 const StructLayout *SL = DL.getStructLayout(STy);
1331 uint64_t Offset = SL->getElementOffset(Idx);
1332 TrailZ = std::min<unsigned>(TrailZ,
1333 countTrailingZeros(Offset));
1335 // Handle array index arithmetic.
1336 Type *IndexedTy = GTI.getIndexedType();
1337 if (!IndexedTy->isSized()) {
1341 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1342 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1343 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1344 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1346 TrailZ = std::min(TrailZ,
1347 unsigned(countTrailingZeros(TypeSize) +
1348 LocalKnownZero.countTrailingOnes()));
1352 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1355 case Instruction::PHI: {
1356 PHINode *P = cast<PHINode>(I);
1357 // Handle the case of a simple two-predecessor recurrence PHI.
1358 // There's a lot more that could theoretically be done here, but
1359 // this is sufficient to catch some interesting cases.
1360 if (P->getNumIncomingValues() == 2) {
1361 for (unsigned i = 0; i != 2; ++i) {
1362 Value *L = P->getIncomingValue(i);
1363 Value *R = P->getIncomingValue(!i);
1364 Operator *LU = dyn_cast<Operator>(L);
1367 unsigned Opcode = LU->getOpcode();
1368 // Check for operations that have the property that if
1369 // both their operands have low zero bits, the result
1370 // will have low zero bits.
1371 if (Opcode == Instruction::Add ||
1372 Opcode == Instruction::Sub ||
1373 Opcode == Instruction::And ||
1374 Opcode == Instruction::Or ||
1375 Opcode == Instruction::Mul) {
1376 Value *LL = LU->getOperand(0);
1377 Value *LR = LU->getOperand(1);
1378 // Find a recurrence.
1385 // Ok, we have a PHI of the form L op= R. Check for low
1387 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1389 // We need to take the minimum number of known bits
1390 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1391 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1393 KnownZero = APInt::getLowBitsSet(BitWidth,
1394 std::min(KnownZero2.countTrailingOnes(),
1395 KnownZero3.countTrailingOnes()));
1401 // Unreachable blocks may have zero-operand PHI nodes.
1402 if (P->getNumIncomingValues() == 0)
1405 // Otherwise take the unions of the known bit sets of the operands,
1406 // taking conservative care to avoid excessive recursion.
1407 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1408 // Skip if every incoming value references to ourself.
1409 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1412 KnownZero = APInt::getAllOnesValue(BitWidth);
1413 KnownOne = APInt::getAllOnesValue(BitWidth);
1414 for (Value *IncValue : P->incoming_values()) {
1415 // Skip direct self references.
1416 if (IncValue == P) continue;
1418 KnownZero2 = APInt(BitWidth, 0);
1419 KnownOne2 = APInt(BitWidth, 0);
1420 // Recurse, but cap the recursion to one level, because we don't
1421 // want to waste time spinning around in loops.
1422 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1424 KnownZero &= KnownZero2;
1425 KnownOne &= KnownOne2;
1426 // If all bits have been ruled out, there's no need to check
1428 if (!KnownZero && !KnownOne)
1434 case Instruction::Call:
1435 case Instruction::Invoke:
1436 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1437 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1438 // If a range metadata is attached to this IntrinsicInst, intersect the
1439 // explicit range specified by the metadata and the implicit range of
1441 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1442 switch (II->getIntrinsicID()) {
1444 case Intrinsic::bswap:
1445 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1447 KnownZero |= KnownZero2.byteSwap();
1448 KnownOne |= KnownOne2.byteSwap();
1450 case Intrinsic::ctlz:
1451 case Intrinsic::cttz: {
1452 unsigned LowBits = Log2_32(BitWidth)+1;
1453 // If this call is undefined for 0, the result will be less than 2^n.
1454 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1456 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1459 case Intrinsic::ctpop: {
1460 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1462 // We can bound the space the count needs. Also, bits known to be zero
1463 // can't contribute to the population.
1464 unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation();
1465 unsigned LeadingZeros =
1466 APInt(BitWidth, BitsPossiblySet).countLeadingZeros();
1467 assert(LeadingZeros <= BitWidth);
1468 KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros);
1469 KnownOne &= ~KnownZero;
1470 // TODO: we could bound KnownOne using the lower bound on the number
1471 // of bits which might be set provided by popcnt KnownOne2.
1474 case Intrinsic::fabs: {
1475 Type *Ty = II->getType();
1476 APInt SignBit = APInt::getSignBit(Ty->getScalarSizeInBits());
1477 KnownZero |= APInt::getSplat(Ty->getPrimitiveSizeInBits(), SignBit);
1480 case Intrinsic::x86_sse42_crc32_64_64:
1481 KnownZero |= APInt::getHighBitsSet(64, 32);
1486 case Instruction::ExtractValue:
1487 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1488 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1489 if (EVI->getNumIndices() != 1) break;
1490 if (EVI->getIndices()[0] == 0) {
1491 switch (II->getIntrinsicID()) {
1493 case Intrinsic::uadd_with_overflow:
1494 case Intrinsic::sadd_with_overflow:
1495 computeKnownBitsAddSub(true, II->getArgOperand(0),
1496 II->getArgOperand(1), false, KnownZero,
1497 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1499 case Intrinsic::usub_with_overflow:
1500 case Intrinsic::ssub_with_overflow:
1501 computeKnownBitsAddSub(false, II->getArgOperand(0),
1502 II->getArgOperand(1), false, KnownZero,
1503 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1505 case Intrinsic::umul_with_overflow:
1506 case Intrinsic::smul_with_overflow:
1507 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1508 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1517 static unsigned getAlignment(const Value *V, const DataLayout &DL) {
1519 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1520 Align = GO->getAlignment();
1522 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1523 Type *ObjectType = GVar->getType()->getElementType();
1524 if (ObjectType->isSized()) {
1525 // If the object is defined in the current Module, we'll be giving
1526 // it the preferred alignment. Otherwise, we have to assume that it
1527 // may only have the minimum ABI alignment.
1528 if (GVar->isStrongDefinitionForLinker())
1529 Align = DL.getPreferredAlignment(GVar);
1531 Align = DL.getABITypeAlignment(ObjectType);
1535 } else if (const Argument *A = dyn_cast<Argument>(V)) {
1536 Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1538 if (!Align && A->hasStructRetAttr()) {
1539 // An sret parameter has at least the ABI alignment of the return type.
1540 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1541 if (EltTy->isSized())
1542 Align = DL.getABITypeAlignment(EltTy);
1544 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
1545 Align = AI->getAlignment();
1546 else if (auto CS = ImmutableCallSite(V))
1547 Align = CS.getAttributes().getParamAlignment(AttributeSet::ReturnIndex);
1548 else if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1549 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_align)) {
1550 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
1551 Align = CI->getLimitedValue();
1557 /// Determine which bits of V are known to be either zero or one and return
1558 /// them in the KnownZero/KnownOne bit sets.
1560 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1561 /// we cannot optimize based on the assumption that it is zero without changing
1562 /// it to be an explicit zero. If we don't change it to zero, other code could
1563 /// optimized based on the contradictory assumption that it is non-zero.
1564 /// Because instcombine aggressively folds operations with undef args anyway,
1565 /// this won't lose us code quality.
1567 /// This function is defined on values with integer type, values with pointer
1568 /// type, and vectors of integers. In the case
1569 /// where V is a vector, known zero, and known one values are the
1570 /// same width as the vector element, and the bit is set only if it is true
1571 /// for all of the elements in the vector.
1572 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1573 const DataLayout &DL, unsigned Depth, const Query &Q) {
1574 assert(V && "No Value?");
1575 assert(Depth <= MaxDepth && "Limit Search Depth");
1576 unsigned BitWidth = KnownZero.getBitWidth();
1578 assert((V->getType()->isIntOrIntVectorTy() ||
1579 V->getType()->isFPOrFPVectorTy() ||
1580 V->getType()->getScalarType()->isPointerTy()) &&
1581 "Not integer, floating point, or pointer type!");
1582 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1583 (!V->getType()->isIntOrIntVectorTy() ||
1584 V->getType()->getScalarSizeInBits() == BitWidth) &&
1585 KnownZero.getBitWidth() == BitWidth &&
1586 KnownOne.getBitWidth() == BitWidth &&
1587 "V, KnownOne and KnownZero should have same BitWidth");
1589 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1590 // We know all of the bits for a constant!
1591 KnownOne = CI->getValue();
1592 KnownZero = ~KnownOne;
1595 // Null and aggregate-zero are all-zeros.
1596 if (isa<ConstantPointerNull>(V) ||
1597 isa<ConstantAggregateZero>(V)) {
1598 KnownOne.clearAllBits();
1599 KnownZero = APInt::getAllOnesValue(BitWidth);
1602 // Handle a constant vector by taking the intersection of the known bits of
1603 // each element. There is no real need to handle ConstantVector here, because
1604 // we don't handle undef in any particularly useful way.
1605 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1606 // We know that CDS must be a vector of integers. Take the intersection of
1608 KnownZero.setAllBits(); KnownOne.setAllBits();
1609 APInt Elt(KnownZero.getBitWidth(), 0);
1610 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1611 Elt = CDS->getElementAsInteger(i);
1618 // Start out not knowing anything.
1619 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1621 // Limit search depth.
1622 // All recursive calls that increase depth must come after this.
1623 if (Depth == MaxDepth)
1626 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1627 // the bits of its aliasee.
1628 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1629 if (!GA->mayBeOverridden())
1630 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1634 if (Operator *I = dyn_cast<Operator>(V))
1635 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1637 // Aligned pointers have trailing zeros - refine KnownZero set
1638 if (V->getType()->isPointerTy()) {
1639 unsigned Align = getAlignment(V, DL);
1641 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1644 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1645 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1646 // computeKnownBitsFromOperator.
1648 // Check whether a nearby assume intrinsic can determine some known bits.
1649 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1651 // Check whether there's a dominating condition which implies something about
1652 // this value at the given context.
1653 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1654 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1657 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1660 /// Determine whether the sign bit is known to be zero or one.
1661 /// Convenience wrapper around computeKnownBits.
1662 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1663 const DataLayout &DL, unsigned Depth, const Query &Q) {
1664 unsigned BitWidth = getBitWidth(V->getType(), DL);
1670 APInt ZeroBits(BitWidth, 0);
1671 APInt OneBits(BitWidth, 0);
1672 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1673 KnownOne = OneBits[BitWidth - 1];
1674 KnownZero = ZeroBits[BitWidth - 1];
1677 /// Return true if the given value is known to have exactly one
1678 /// bit set when defined. For vectors return true if every element is known to
1679 /// be a power of two when defined. Supports values with integer or pointer
1680 /// types and vectors of integers.
1681 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1682 const Query &Q, const DataLayout &DL) {
1683 if (Constant *C = dyn_cast<Constant>(V)) {
1684 if (C->isNullValue())
1686 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1687 return CI->getValue().isPowerOf2();
1688 // TODO: Handle vector constants.
1691 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1692 // it is shifted off the end then the result is undefined.
1693 if (match(V, m_Shl(m_One(), m_Value())))
1696 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1697 // bottom. If it is shifted off the bottom then the result is undefined.
1698 if (match(V, m_LShr(m_SignBit(), m_Value())))
1701 // The remaining tests are all recursive, so bail out if we hit the limit.
1702 if (Depth++ == MaxDepth)
1705 Value *X = nullptr, *Y = nullptr;
1706 // A shift of a power of two is a power of two or zero.
1707 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1708 match(V, m_Shr(m_Value(X), m_Value()))))
1709 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1711 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1712 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1714 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1715 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1716 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1718 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1719 // A power of two and'd with anything is a power of two or zero.
1720 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1721 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1723 // X & (-X) is always a power of two or zero.
1724 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1729 // Adding a power-of-two or zero to the same power-of-two or zero yields
1730 // either the original power-of-two, a larger power-of-two or zero.
1731 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1732 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1733 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1734 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1735 match(X, m_And(m_Value(), m_Specific(Y))))
1736 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1738 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1739 match(Y, m_And(m_Value(), m_Specific(X))))
1740 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1743 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1744 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1745 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1747 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1748 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1749 // If i8 V is a power of two or zero:
1750 // ZeroBits: 1 1 1 0 1 1 1 1
1751 // ~ZeroBits: 0 0 0 1 0 0 0 0
1752 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1753 // If OrZero isn't set, we cannot give back a zero result.
1754 // Make sure either the LHS or RHS has a bit set.
1755 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1760 // An exact divide or right shift can only shift off zero bits, so the result
1761 // is a power of two only if the first operand is a power of two and not
1762 // copying a sign bit (sdiv int_min, 2).
1763 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1764 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1765 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1772 /// \brief Test whether a GEP's result is known to be non-null.
1774 /// Uses properties inherent in a GEP to try to determine whether it is known
1777 /// Currently this routine does not support vector GEPs.
1778 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1779 unsigned Depth, const Query &Q) {
1780 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1783 // FIXME: Support vector-GEPs.
1784 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1786 // If the base pointer is non-null, we cannot walk to a null address with an
1787 // inbounds GEP in address space zero.
1788 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1791 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1792 // If so, then the GEP cannot produce a null pointer, as doing so would
1793 // inherently violate the inbounds contract within address space zero.
1794 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1795 GTI != GTE; ++GTI) {
1796 // Struct types are easy -- they must always be indexed by a constant.
1797 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1798 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1799 unsigned ElementIdx = OpC->getZExtValue();
1800 const StructLayout *SL = DL.getStructLayout(STy);
1801 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1802 if (ElementOffset > 0)
1807 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1808 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1811 // Fast path the constant operand case both for efficiency and so we don't
1812 // increment Depth when just zipping down an all-constant GEP.
1813 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1819 // We post-increment Depth here because while isKnownNonZero increments it
1820 // as well, when we pop back up that increment won't persist. We don't want
1821 // to recurse 10k times just because we have 10k GEP operands. We don't
1822 // bail completely out because we want to handle constant GEPs regardless
1824 if (Depth++ >= MaxDepth)
1827 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1834 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1835 /// ensure that the value it's attached to is never Value? 'RangeType' is
1836 /// is the type of the value described by the range.
1837 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1838 const APInt& Value) {
1839 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1840 assert(NumRanges >= 1);
1841 for (unsigned i = 0; i < NumRanges; ++i) {
1842 ConstantInt *Lower =
1843 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1844 ConstantInt *Upper =
1845 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1846 ConstantRange Range(Lower->getValue(), Upper->getValue());
1847 if (Range.contains(Value))
1853 /// Return true if the given value is known to be non-zero when defined.
1854 /// For vectors return true if every element is known to be non-zero when
1855 /// defined. Supports values with integer or pointer type and vectors of
1857 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1859 if (Constant *C = dyn_cast<Constant>(V)) {
1860 if (C->isNullValue())
1862 if (isa<ConstantInt>(C))
1863 // Must be non-zero due to null test above.
1865 // TODO: Handle vectors
1869 if (Instruction* I = dyn_cast<Instruction>(V)) {
1870 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1871 // If the possible ranges don't contain zero, then the value is
1872 // definitely non-zero.
1873 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1874 const APInt ZeroValue(Ty->getBitWidth(), 0);
1875 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1881 // The remaining tests are all recursive, so bail out if we hit the limit.
1882 if (Depth++ >= MaxDepth)
1885 // Check for pointer simplifications.
1886 if (V->getType()->isPointerTy()) {
1887 if (isKnownNonNull(V))
1889 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1890 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1894 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1896 // X | Y != 0 if X != 0 or Y != 0.
1897 Value *X = nullptr, *Y = nullptr;
1898 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1899 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1901 // ext X != 0 if X != 0.
1902 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1903 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1905 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1906 // if the lowest bit is shifted off the end.
1907 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1908 // shl nuw can't remove any non-zero bits.
1909 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1910 if (BO->hasNoUnsignedWrap())
1911 return isKnownNonZero(X, DL, Depth, Q);
1913 APInt KnownZero(BitWidth, 0);
1914 APInt KnownOne(BitWidth, 0);
1915 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1919 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1920 // defined if the sign bit is shifted off the end.
1921 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1922 // shr exact can only shift out zero bits.
1923 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1925 return isKnownNonZero(X, DL, Depth, Q);
1927 bool XKnownNonNegative, XKnownNegative;
1928 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1932 // If the shifter operand is a constant, and all of the bits shifted
1933 // out are known to be zero, and X is known non-zero then at least one
1934 // non-zero bit must remain.
1935 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1936 APInt KnownZero(BitWidth, 0);
1937 APInt KnownOne(BitWidth, 0);
1938 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1940 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1941 // Is there a known one in the portion not shifted out?
1942 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1944 // Are all the bits to be shifted out known zero?
1945 if (KnownZero.countTrailingOnes() >= ShiftVal)
1946 return isKnownNonZero(X, DL, Depth, Q);
1949 // div exact can only produce a zero if the dividend is zero.
1950 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1951 return isKnownNonZero(X, DL, Depth, Q);
1954 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1955 bool XKnownNonNegative, XKnownNegative;
1956 bool YKnownNonNegative, YKnownNegative;
1957 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1958 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1960 // If X and Y are both non-negative (as signed values) then their sum is not
1961 // zero unless both X and Y are zero.
1962 if (XKnownNonNegative && YKnownNonNegative)
1963 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1966 // If X and Y are both negative (as signed values) then their sum is not
1967 // zero unless both X and Y equal INT_MIN.
1968 if (BitWidth && XKnownNegative && YKnownNegative) {
1969 APInt KnownZero(BitWidth, 0);
1970 APInt KnownOne(BitWidth, 0);
1971 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1972 // The sign bit of X is set. If some other bit is set then X is not equal
1974 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1975 if ((KnownOne & Mask) != 0)
1977 // The sign bit of Y is set. If some other bit is set then Y is not equal
1979 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1980 if ((KnownOne & Mask) != 0)
1984 // The sum of a non-negative number and a power of two is not zero.
1985 if (XKnownNonNegative &&
1986 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1988 if (YKnownNonNegative &&
1989 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1993 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1994 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1995 // If X and Y are non-zero then so is X * Y as long as the multiplication
1996 // does not overflow.
1997 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1998 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
2001 // (C ? X : Y) != 0 if X != 0 and Y != 0.
2002 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2003 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
2004 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
2008 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
2009 // Try and detect a recurrence that monotonically increases from a
2010 // starting value, as these are common as induction variables.
2011 if (PN->getNumIncomingValues() == 2) {
2012 Value *Start = PN->getIncomingValue(0);
2013 Value *Induction = PN->getIncomingValue(1);
2014 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2015 std::swap(Start, Induction);
2016 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2017 if (!C->isZero() && !C->isNegative()) {
2019 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2020 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2028 if (!BitWidth) return false;
2029 APInt KnownZero(BitWidth, 0);
2030 APInt KnownOne(BitWidth, 0);
2031 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2032 return KnownOne != 0;
2035 /// Return true if V2 == V1 + X, where X is known non-zero.
2036 static bool isAddOfNonZero(Value *V1, Value *V2, const DataLayout &DL,
2038 BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2039 if (!BO || BO->getOpcode() != Instruction::Add)
2041 Value *Op = nullptr;
2042 if (V2 == BO->getOperand(0))
2043 Op = BO->getOperand(1);
2044 else if (V2 == BO->getOperand(1))
2045 Op = BO->getOperand(0);
2048 return isKnownNonZero(Op, DL, 0, Q);
2051 /// Return true if it is known that V1 != V2.
2052 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
2054 if (V1->getType()->isVectorTy() || V1 == V2)
2056 if (V1->getType() != V2->getType())
2057 // We can't look through casts yet.
2059 if (isAddOfNonZero(V1, V2, DL, Q) || isAddOfNonZero(V2, V1, DL, Q))
2062 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
2063 // Are any known bits in V1 contradictory to known bits in V2? If V1
2064 // has a known zero where V2 has a known one, they must not be equal.
2065 auto BitWidth = Ty->getBitWidth();
2066 APInt KnownZero1(BitWidth, 0);
2067 APInt KnownOne1(BitWidth, 0);
2068 computeKnownBits(V1, KnownZero1, KnownOne1, DL, 0, Q);
2069 APInt KnownZero2(BitWidth, 0);
2070 APInt KnownOne2(BitWidth, 0);
2071 computeKnownBits(V2, KnownZero2, KnownOne2, DL, 0, Q);
2073 auto OppositeBits = (KnownZero1 & KnownOne2) | (KnownZero2 & KnownOne1);
2074 if (OppositeBits.getBoolValue())
2080 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2081 /// simplify operations downstream. Mask is known to be zero for bits that V
2084 /// This function is defined on values with integer type, values with pointer
2085 /// type, and vectors of integers. In the case
2086 /// where V is a vector, the mask, known zero, and known one values are the
2087 /// same width as the vector element, and the bit is set only if it is true
2088 /// for all of the elements in the vector.
2089 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
2090 unsigned Depth, const Query &Q) {
2091 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
2092 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2093 return (KnownZero & Mask) == Mask;
2098 /// Return the number of times the sign bit of the register is replicated into
2099 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2100 /// (itself), but other cases can give us information. For example, immediately
2101 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2102 /// other, so we return 3.
2104 /// 'Op' must have a scalar integer type.
2106 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
2108 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
2110 unsigned FirstAnswer = 1;
2112 // Note that ConstantInt is handled by the general computeKnownBits case
2116 return 1; // Limit search depth.
2118 Operator *U = dyn_cast<Operator>(V);
2119 switch (Operator::getOpcode(V)) {
2121 case Instruction::SExt:
2122 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2123 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
2125 case Instruction::SDiv: {
2126 const APInt *Denominator;
2127 // sdiv X, C -> adds log(C) sign bits.
2128 if (match(U->getOperand(1), m_APInt(Denominator))) {
2130 // Ignore non-positive denominator.
2131 if (!Denominator->isStrictlyPositive())
2134 // Calculate the incoming numerator bits.
2135 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2137 // Add floor(log(C)) bits to the numerator bits.
2138 return std::min(TyBits, NumBits + Denominator->logBase2());
2143 case Instruction::SRem: {
2144 const APInt *Denominator;
2145 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2146 // positive constant. This let us put a lower bound on the number of sign
2148 if (match(U->getOperand(1), m_APInt(Denominator))) {
2150 // Ignore non-positive denominator.
2151 if (!Denominator->isStrictlyPositive())
2154 // Calculate the incoming numerator bits. SRem by a positive constant
2155 // can't lower the number of sign bits.
2157 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2159 // Calculate the leading sign bit constraints by examining the
2160 // denominator. Given that the denominator is positive, there are two
2163 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2164 // (1 << ceilLogBase2(C)).
2166 // 2. the numerator is negative. Then the result range is (-C,0] and
2167 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2169 // Thus a lower bound on the number of sign bits is `TyBits -
2170 // ceilLogBase2(C)`.
2172 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2173 return std::max(NumrBits, ResBits);
2178 case Instruction::AShr: {
2179 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2180 // ashr X, C -> adds C sign bits. Vectors too.
2182 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2183 Tmp += ShAmt->getZExtValue();
2184 if (Tmp > TyBits) Tmp = TyBits;
2188 case Instruction::Shl: {
2190 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2191 // shl destroys sign bits.
2192 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2193 Tmp2 = ShAmt->getZExtValue();
2194 if (Tmp2 >= TyBits || // Bad shift.
2195 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2200 case Instruction::And:
2201 case Instruction::Or:
2202 case Instruction::Xor: // NOT is handled here.
2203 // Logical binary ops preserve the number of sign bits at the worst.
2204 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2206 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2207 FirstAnswer = std::min(Tmp, Tmp2);
2208 // We computed what we know about the sign bits as our first
2209 // answer. Now proceed to the generic code that uses
2210 // computeKnownBits, and pick whichever answer is better.
2214 case Instruction::Select:
2215 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2216 if (Tmp == 1) return 1; // Early out.
2217 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2218 return std::min(Tmp, Tmp2);
2220 case Instruction::Add:
2221 // Add can have at most one carry bit. Thus we know that the output
2222 // is, at worst, one more bit than the inputs.
2223 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2224 if (Tmp == 1) return 1; // Early out.
2226 // Special case decrementing a value (ADD X, -1):
2227 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2228 if (CRHS->isAllOnesValue()) {
2229 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2230 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2233 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2235 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2238 // If we are subtracting one from a positive number, there is no carry
2239 // out of the result.
2240 if (KnownZero.isNegative())
2244 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2245 if (Tmp2 == 1) return 1;
2246 return std::min(Tmp, Tmp2)-1;
2248 case Instruction::Sub:
2249 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2250 if (Tmp2 == 1) return 1;
2253 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2254 if (CLHS->isNullValue()) {
2255 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2256 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2258 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2260 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2263 // If the input is known to be positive (the sign bit is known clear),
2264 // the output of the NEG has the same number of sign bits as the input.
2265 if (KnownZero.isNegative())
2268 // Otherwise, we treat this like a SUB.
2271 // Sub can have at most one carry bit. Thus we know that the output
2272 // is, at worst, one more bit than the inputs.
2273 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2274 if (Tmp == 1) return 1; // Early out.
2275 return std::min(Tmp, Tmp2)-1;
2277 case Instruction::PHI: {
2278 PHINode *PN = cast<PHINode>(U);
2279 unsigned NumIncomingValues = PN->getNumIncomingValues();
2280 // Don't analyze large in-degree PHIs.
2281 if (NumIncomingValues > 4) break;
2282 // Unreachable blocks may have zero-operand PHI nodes.
2283 if (NumIncomingValues == 0) break;
2285 // Take the minimum of all incoming values. This can't infinitely loop
2286 // because of our depth threshold.
2287 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2288 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2289 if (Tmp == 1) return Tmp;
2291 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2296 case Instruction::Trunc:
2297 // FIXME: it's tricky to do anything useful for this, but it is an important
2298 // case for targets like X86.
2302 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2303 // use this information.
2304 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2306 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2308 if (KnownZero.isNegative()) { // sign bit is 0
2310 } else if (KnownOne.isNegative()) { // sign bit is 1;
2317 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2318 // the number of identical bits in the top of the input value.
2320 Mask <<= Mask.getBitWidth()-TyBits;
2321 // Return # leading zeros. We use 'min' here in case Val was zero before
2322 // shifting. We don't want to return '64' as for an i32 "0".
2323 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2326 /// This function computes the integer multiple of Base that equals V.
2327 /// If successful, it returns true and returns the multiple in
2328 /// Multiple. If unsuccessful, it returns false. It looks
2329 /// through SExt instructions only if LookThroughSExt is true.
2330 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2331 bool LookThroughSExt, unsigned Depth) {
2332 const unsigned MaxDepth = 6;
2334 assert(V && "No Value?");
2335 assert(Depth <= MaxDepth && "Limit Search Depth");
2336 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2338 Type *T = V->getType();
2340 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2350 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2351 Constant *BaseVal = ConstantInt::get(T, Base);
2352 if (CO && CO == BaseVal) {
2354 Multiple = ConstantInt::get(T, 1);
2358 if (CI && CI->getZExtValue() % Base == 0) {
2359 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2363 if (Depth == MaxDepth) return false; // Limit search depth.
2365 Operator *I = dyn_cast<Operator>(V);
2366 if (!I) return false;
2368 switch (I->getOpcode()) {
2370 case Instruction::SExt:
2371 if (!LookThroughSExt) return false;
2372 // otherwise fall through to ZExt
2373 case Instruction::ZExt:
2374 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2375 LookThroughSExt, Depth+1);
2376 case Instruction::Shl:
2377 case Instruction::Mul: {
2378 Value *Op0 = I->getOperand(0);
2379 Value *Op1 = I->getOperand(1);
2381 if (I->getOpcode() == Instruction::Shl) {
2382 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2383 if (!Op1CI) return false;
2384 // Turn Op0 << Op1 into Op0 * 2^Op1
2385 APInt Op1Int = Op1CI->getValue();
2386 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2387 APInt API(Op1Int.getBitWidth(), 0);
2388 API.setBit(BitToSet);
2389 Op1 = ConstantInt::get(V->getContext(), API);
2392 Value *Mul0 = nullptr;
2393 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2394 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2395 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2396 if (Op1C->getType()->getPrimitiveSizeInBits() <
2397 MulC->getType()->getPrimitiveSizeInBits())
2398 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2399 if (Op1C->getType()->getPrimitiveSizeInBits() >
2400 MulC->getType()->getPrimitiveSizeInBits())
2401 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2403 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2404 Multiple = ConstantExpr::getMul(MulC, Op1C);
2408 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2409 if (Mul0CI->getValue() == 1) {
2410 // V == Base * Op1, so return Op1
2416 Value *Mul1 = nullptr;
2417 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2418 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2419 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2420 if (Op0C->getType()->getPrimitiveSizeInBits() <
2421 MulC->getType()->getPrimitiveSizeInBits())
2422 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2423 if (Op0C->getType()->getPrimitiveSizeInBits() >
2424 MulC->getType()->getPrimitiveSizeInBits())
2425 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2427 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2428 Multiple = ConstantExpr::getMul(MulC, Op0C);
2432 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2433 if (Mul1CI->getValue() == 1) {
2434 // V == Base * Op0, so return Op0
2442 // We could not determine if V is a multiple of Base.
2446 /// Return true if we can prove that the specified FP value is never equal to
2449 /// NOTE: this function will need to be revisited when we support non-default
2452 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2453 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2454 return !CFP->getValueAPF().isNegZero();
2456 // FIXME: Magic number! At the least, this should be given a name because it's
2457 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2458 // expose it as a parameter, so it can be used for testing / experimenting.
2460 return false; // Limit search depth.
2462 const Operator *I = dyn_cast<Operator>(V);
2463 if (!I) return false;
2465 // Check if the nsz fast-math flag is set
2466 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2467 if (FPO->hasNoSignedZeros())
2470 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2471 if (I->getOpcode() == Instruction::FAdd)
2472 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2473 if (CFP->isNullValue())
2476 // sitofp and uitofp turn into +0.0 for zero.
2477 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2480 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2481 // sqrt(-0.0) = -0.0, no other negative results are possible.
2482 if (II->getIntrinsicID() == Intrinsic::sqrt)
2483 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2485 if (const CallInst *CI = dyn_cast<CallInst>(I))
2486 if (const Function *F = CI->getCalledFunction()) {
2487 if (F->isDeclaration()) {
2489 if (F->getName() == "abs") return true;
2490 // fabs[lf](x) != -0.0
2491 if (F->getName() == "fabs") return true;
2492 if (F->getName() == "fabsf") return true;
2493 if (F->getName() == "fabsl") return true;
2494 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2495 F->getName() == "sqrtl")
2496 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2503 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2504 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2505 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2507 // FIXME: Magic number! At the least, this should be given a name because it's
2508 // used similarly in CannotBeNegativeZero(). A better fix may be to
2509 // expose it as a parameter, so it can be used for testing / experimenting.
2511 return false; // Limit search depth.
2513 const Operator *I = dyn_cast<Operator>(V);
2514 if (!I) return false;
2516 switch (I->getOpcode()) {
2518 case Instruction::FMul:
2519 // x*x is always non-negative or a NaN.
2520 if (I->getOperand(0) == I->getOperand(1))
2523 case Instruction::FAdd:
2524 case Instruction::FDiv:
2525 case Instruction::FRem:
2526 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2527 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2528 case Instruction::FPExt:
2529 case Instruction::FPTrunc:
2530 // Widening/narrowing never change sign.
2531 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2532 case Instruction::Call:
2533 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2534 switch (II->getIntrinsicID()) {
2536 case Intrinsic::exp:
2537 case Intrinsic::exp2:
2538 case Intrinsic::fabs:
2539 case Intrinsic::sqrt:
2541 case Intrinsic::powi:
2542 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2543 // powi(x,n) is non-negative if n is even.
2544 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2547 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2548 case Intrinsic::fma:
2549 case Intrinsic::fmuladd:
2550 // x*x+y is non-negative if y is non-negative.
2551 return I->getOperand(0) == I->getOperand(1) &&
2552 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2559 /// If the specified value can be set by repeating the same byte in memory,
2560 /// return the i8 value that it is represented with. This is
2561 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2562 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2563 /// byte store (e.g. i16 0x1234), return null.
2564 Value *llvm::isBytewiseValue(Value *V) {
2565 // All byte-wide stores are splatable, even of arbitrary variables.
2566 if (V->getType()->isIntegerTy(8)) return V;
2568 // Handle 'null' ConstantArrayZero etc.
2569 if (Constant *C = dyn_cast<Constant>(V))
2570 if (C->isNullValue())
2571 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2573 // Constant float and double values can be handled as integer values if the
2574 // corresponding integer value is "byteable". An important case is 0.0.
2575 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2576 if (CFP->getType()->isFloatTy())
2577 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2578 if (CFP->getType()->isDoubleTy())
2579 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2580 // Don't handle long double formats, which have strange constraints.
2583 // We can handle constant integers that are multiple of 8 bits.
2584 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2585 if (CI->getBitWidth() % 8 == 0) {
2586 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2588 if (!CI->getValue().isSplat(8))
2590 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2594 // A ConstantDataArray/Vector is splatable if all its members are equal and
2596 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2597 Value *Elt = CA->getElementAsConstant(0);
2598 Value *Val = isBytewiseValue(Elt);
2602 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2603 if (CA->getElementAsConstant(I) != Elt)
2609 // Conceptually, we could handle things like:
2610 // %a = zext i8 %X to i16
2611 // %b = shl i16 %a, 8
2612 // %c = or i16 %a, %b
2613 // but until there is an example that actually needs this, it doesn't seem
2614 // worth worrying about.
2619 // This is the recursive version of BuildSubAggregate. It takes a few different
2620 // arguments. Idxs is the index within the nested struct From that we are
2621 // looking at now (which is of type IndexedType). IdxSkip is the number of
2622 // indices from Idxs that should be left out when inserting into the resulting
2623 // struct. To is the result struct built so far, new insertvalue instructions
2625 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2626 SmallVectorImpl<unsigned> &Idxs,
2628 Instruction *InsertBefore) {
2629 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2631 // Save the original To argument so we can modify it
2633 // General case, the type indexed by Idxs is a struct
2634 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2635 // Process each struct element recursively
2638 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2642 // Couldn't find any inserted value for this index? Cleanup
2643 while (PrevTo != OrigTo) {
2644 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2645 PrevTo = Del->getAggregateOperand();
2646 Del->eraseFromParent();
2648 // Stop processing elements
2652 // If we successfully found a value for each of our subaggregates
2656 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2657 // the struct's elements had a value that was inserted directly. In the latter
2658 // case, perhaps we can't determine each of the subelements individually, but
2659 // we might be able to find the complete struct somewhere.
2661 // Find the value that is at that particular spot
2662 Value *V = FindInsertedValue(From, Idxs);
2667 // Insert the value in the new (sub) aggregrate
2668 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2669 "tmp", InsertBefore);
2672 // This helper takes a nested struct and extracts a part of it (which is again a
2673 // struct) into a new value. For example, given the struct:
2674 // { a, { b, { c, d }, e } }
2675 // and the indices "1, 1" this returns
2678 // It does this by inserting an insertvalue for each element in the resulting
2679 // struct, as opposed to just inserting a single struct. This will only work if
2680 // each of the elements of the substruct are known (ie, inserted into From by an
2681 // insertvalue instruction somewhere).
2683 // All inserted insertvalue instructions are inserted before InsertBefore
2684 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2685 Instruction *InsertBefore) {
2686 assert(InsertBefore && "Must have someplace to insert!");
2687 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2689 Value *To = UndefValue::get(IndexedType);
2690 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2691 unsigned IdxSkip = Idxs.size();
2693 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2696 /// Given an aggregrate and an sequence of indices, see if
2697 /// the scalar value indexed is already around as a register, for example if it
2698 /// were inserted directly into the aggregrate.
2700 /// If InsertBefore is not null, this function will duplicate (modified)
2701 /// insertvalues when a part of a nested struct is extracted.
2702 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2703 Instruction *InsertBefore) {
2704 // Nothing to index? Just return V then (this is useful at the end of our
2706 if (idx_range.empty())
2708 // We have indices, so V should have an indexable type.
2709 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2710 "Not looking at a struct or array?");
2711 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2712 "Invalid indices for type?");
2714 if (Constant *C = dyn_cast<Constant>(V)) {
2715 C = C->getAggregateElement(idx_range[0]);
2716 if (!C) return nullptr;
2717 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2720 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2721 // Loop the indices for the insertvalue instruction in parallel with the
2722 // requested indices
2723 const unsigned *req_idx = idx_range.begin();
2724 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2725 i != e; ++i, ++req_idx) {
2726 if (req_idx == idx_range.end()) {
2727 // We can't handle this without inserting insertvalues
2731 // The requested index identifies a part of a nested aggregate. Handle
2732 // this specially. For example,
2733 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2734 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2735 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2736 // This can be changed into
2737 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2738 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2739 // which allows the unused 0,0 element from the nested struct to be
2741 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2745 // This insert value inserts something else than what we are looking for.
2746 // See if the (aggregate) value inserted into has the value we are
2747 // looking for, then.
2749 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2752 // If we end up here, the indices of the insertvalue match with those
2753 // requested (though possibly only partially). Now we recursively look at
2754 // the inserted value, passing any remaining indices.
2755 return FindInsertedValue(I->getInsertedValueOperand(),
2756 makeArrayRef(req_idx, idx_range.end()),
2760 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2761 // If we're extracting a value from an aggregate that was extracted from
2762 // something else, we can extract from that something else directly instead.
2763 // However, we will need to chain I's indices with the requested indices.
2765 // Calculate the number of indices required
2766 unsigned size = I->getNumIndices() + idx_range.size();
2767 // Allocate some space to put the new indices in
2768 SmallVector<unsigned, 5> Idxs;
2770 // Add indices from the extract value instruction
2771 Idxs.append(I->idx_begin(), I->idx_end());
2773 // Add requested indices
2774 Idxs.append(idx_range.begin(), idx_range.end());
2776 assert(Idxs.size() == size
2777 && "Number of indices added not correct?");
2779 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2781 // Otherwise, we don't know (such as, extracting from a function return value
2782 // or load instruction)
2786 /// Analyze the specified pointer to see if it can be expressed as a base
2787 /// pointer plus a constant offset. Return the base and offset to the caller.
2788 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2789 const DataLayout &DL) {
2790 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2791 APInt ByteOffset(BitWidth, 0);
2793 if (Ptr->getType()->isVectorTy())
2796 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2797 APInt GEPOffset(BitWidth, 0);
2798 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2801 ByteOffset += GEPOffset;
2803 Ptr = GEP->getPointerOperand();
2804 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2805 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2806 Ptr = cast<Operator>(Ptr)->getOperand(0);
2807 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2808 if (GA->mayBeOverridden())
2810 Ptr = GA->getAliasee();
2815 Offset = ByteOffset.getSExtValue();
2820 /// This function computes the length of a null-terminated C string pointed to
2821 /// by V. If successful, it returns true and returns the string in Str.
2822 /// If unsuccessful, it returns false.
2823 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2824 uint64_t Offset, bool TrimAtNul) {
2827 // Look through bitcast instructions and geps.
2828 V = V->stripPointerCasts();
2830 // If the value is a GEP instruction or constant expression, treat it as an
2832 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2833 // Make sure the GEP has exactly three arguments.
2834 if (GEP->getNumOperands() != 3)
2837 // Make sure the index-ee is a pointer to array of i8.
2838 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2839 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2840 if (!AT || !AT->getElementType()->isIntegerTy(8))
2843 // Check to make sure that the first operand of the GEP is an integer and
2844 // has value 0 so that we are sure we're indexing into the initializer.
2845 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2846 if (!FirstIdx || !FirstIdx->isZero())
2849 // If the second index isn't a ConstantInt, then this is a variable index
2850 // into the array. If this occurs, we can't say anything meaningful about
2852 uint64_t StartIdx = 0;
2853 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2854 StartIdx = CI->getZExtValue();
2857 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2861 // The GEP instruction, constant or instruction, must reference a global
2862 // variable that is a constant and is initialized. The referenced constant
2863 // initializer is the array that we'll use for optimization.
2864 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2865 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2868 // Handle the all-zeros case
2869 if (GV->getInitializer()->isNullValue()) {
2870 // This is a degenerate case. The initializer is constant zero so the
2871 // length of the string must be zero.
2876 // Must be a Constant Array
2877 const ConstantDataArray *Array =
2878 dyn_cast<ConstantDataArray>(GV->getInitializer());
2879 if (!Array || !Array->isString())
2882 // Get the number of elements in the array
2883 uint64_t NumElts = Array->getType()->getArrayNumElements();
2885 // Start out with the entire array in the StringRef.
2886 Str = Array->getAsString();
2888 if (Offset > NumElts)
2891 // Skip over 'offset' bytes.
2892 Str = Str.substr(Offset);
2895 // Trim off the \0 and anything after it. If the array is not nul
2896 // terminated, we just return the whole end of string. The client may know
2897 // some other way that the string is length-bound.
2898 Str = Str.substr(0, Str.find('\0'));
2903 // These next two are very similar to the above, but also look through PHI
2905 // TODO: See if we can integrate these two together.
2907 /// If we can compute the length of the string pointed to by
2908 /// the specified pointer, return 'len+1'. If we can't, return 0.
2909 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2910 // Look through noop bitcast instructions.
2911 V = V->stripPointerCasts();
2913 // If this is a PHI node, there are two cases: either we have already seen it
2915 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2916 if (!PHIs.insert(PN).second)
2917 return ~0ULL; // already in the set.
2919 // If it was new, see if all the input strings are the same length.
2920 uint64_t LenSoFar = ~0ULL;
2921 for (Value *IncValue : PN->incoming_values()) {
2922 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2923 if (Len == 0) return 0; // Unknown length -> unknown.
2925 if (Len == ~0ULL) continue;
2927 if (Len != LenSoFar && LenSoFar != ~0ULL)
2928 return 0; // Disagree -> unknown.
2932 // Success, all agree.
2936 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2937 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2938 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2939 if (Len1 == 0) return 0;
2940 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2941 if (Len2 == 0) return 0;
2942 if (Len1 == ~0ULL) return Len2;
2943 if (Len2 == ~0ULL) return Len1;
2944 if (Len1 != Len2) return 0;
2948 // Otherwise, see if we can read the string.
2950 if (!getConstantStringInfo(V, StrData))
2953 return StrData.size()+1;
2956 /// If we can compute the length of the string pointed to by
2957 /// the specified pointer, return 'len+1'. If we can't, return 0.
2958 uint64_t llvm::GetStringLength(Value *V) {
2959 if (!V->getType()->isPointerTy()) return 0;
2961 SmallPtrSet<PHINode*, 32> PHIs;
2962 uint64_t Len = GetStringLengthH(V, PHIs);
2963 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2964 // an empty string as a length.
2965 return Len == ~0ULL ? 1 : Len;
2968 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2969 /// previous iteration of the loop was referring to the same object as \p PN.
2970 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2971 // Find the loop-defined value.
2972 Loop *L = LI->getLoopFor(PN->getParent());
2973 if (PN->getNumIncomingValues() != 2)
2976 // Find the value from previous iteration.
2977 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2978 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2979 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2980 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2983 // If a new pointer is loaded in the loop, the pointer references a different
2984 // object in every iteration. E.g.:
2988 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2989 if (!L->isLoopInvariant(Load->getPointerOperand()))
2994 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2995 unsigned MaxLookup) {
2996 if (!V->getType()->isPointerTy())
2998 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2999 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3000 V = GEP->getPointerOperand();
3001 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3002 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3003 V = cast<Operator>(V)->getOperand(0);
3004 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3005 if (GA->mayBeOverridden())
3007 V = GA->getAliasee();
3009 // See if InstructionSimplify knows any relevant tricks.
3010 if (Instruction *I = dyn_cast<Instruction>(V))
3011 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3012 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
3019 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3024 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3025 const DataLayout &DL, LoopInfo *LI,
3026 unsigned MaxLookup) {
3027 SmallPtrSet<Value *, 4> Visited;
3028 SmallVector<Value *, 4> Worklist;
3029 Worklist.push_back(V);
3031 Value *P = Worklist.pop_back_val();
3032 P = GetUnderlyingObject(P, DL, MaxLookup);
3034 if (!Visited.insert(P).second)
3037 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3038 Worklist.push_back(SI->getTrueValue());
3039 Worklist.push_back(SI->getFalseValue());
3043 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3044 // If this PHI changes the underlying object in every iteration of the
3045 // loop, don't look through it. Consider:
3048 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3052 // Prev is tracking Curr one iteration behind so they refer to different
3053 // underlying objects.
3054 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3055 isSameUnderlyingObjectInLoop(PN, LI))
3056 for (Value *IncValue : PN->incoming_values())
3057 Worklist.push_back(IncValue);
3061 Objects.push_back(P);
3062 } while (!Worklist.empty());
3065 /// Return true if the only users of this pointer are lifetime markers.
3066 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3067 for (const User *U : V->users()) {
3068 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3069 if (!II) return false;
3071 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3072 II->getIntrinsicID() != Intrinsic::lifetime_end)
3078 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
3079 Type *Ty, const DataLayout &DL,
3080 const Instruction *CtxI,
3081 const DominatorTree *DT,
3082 const TargetLibraryInfo *TLI) {
3083 assert(Offset.isNonNegative() && "offset can't be negative");
3084 assert(Ty->isSized() && "must be sized");
3086 APInt DerefBytes(Offset.getBitWidth(), 0);
3087 bool CheckForNonNull = false;
3088 if (const Argument *A = dyn_cast<Argument>(BV)) {
3089 DerefBytes = A->getDereferenceableBytes();
3090 if (!DerefBytes.getBoolValue()) {
3091 DerefBytes = A->getDereferenceableOrNullBytes();
3092 CheckForNonNull = true;
3094 } else if (auto CS = ImmutableCallSite(BV)) {
3095 DerefBytes = CS.getDereferenceableBytes(0);
3096 if (!DerefBytes.getBoolValue()) {
3097 DerefBytes = CS.getDereferenceableOrNullBytes(0);
3098 CheckForNonNull = true;
3100 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
3101 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
3102 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3103 DerefBytes = CI->getLimitedValue();
3105 if (!DerefBytes.getBoolValue()) {
3107 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
3108 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3109 DerefBytes = CI->getLimitedValue();
3111 CheckForNonNull = true;
3115 if (DerefBytes.getBoolValue())
3116 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
3117 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
3123 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
3124 const Instruction *CtxI,
3125 const DominatorTree *DT,
3126 const TargetLibraryInfo *TLI) {
3127 Type *VTy = V->getType();
3128 Type *Ty = VTy->getPointerElementType();
3132 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3133 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
3136 static bool isAligned(const Value *Base, APInt Offset, unsigned Align,
3137 const DataLayout &DL) {
3138 APInt BaseAlign(Offset.getBitWidth(), getAlignment(Base, DL));
3141 Type *Ty = Base->getType()->getPointerElementType();
3142 BaseAlign = DL.getABITypeAlignment(Ty);
3145 APInt Alignment(Offset.getBitWidth(), Align);
3147 assert(Alignment.isPowerOf2() && "must be a power of 2!");
3148 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1));
3151 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) {
3152 APInt Offset(DL.getTypeStoreSizeInBits(Base->getType()), 0);
3153 return isAligned(Base, Offset, Align, DL);
3156 /// Test if V is always a pointer to allocated and suitably aligned memory for
3157 /// a simple load or store.
3158 static bool isDereferenceableAndAlignedPointer(
3159 const Value *V, unsigned Align, const DataLayout &DL,
3160 const Instruction *CtxI, const DominatorTree *DT,
3161 const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited) {
3162 // Note that it is not safe to speculate into a malloc'd region because
3163 // malloc may return null.
3165 // These are obviously ok if aligned.
3166 if (isa<AllocaInst>(V))
3167 return isAligned(V, Align, DL);
3169 // It's not always safe to follow a bitcast, for example:
3170 // bitcast i8* (alloca i8) to i32*
3171 // would result in a 4-byte load from a 1-byte alloca. However,
3172 // if we're casting from a pointer from a type of larger size
3173 // to a type of smaller size (or the same size), and the alignment
3174 // is at least as large as for the resulting pointer type, then
3175 // we can look through the bitcast.
3176 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
3177 Type *STy = BC->getSrcTy()->getPointerElementType(),
3178 *DTy = BC->getDestTy()->getPointerElementType();
3179 if (STy->isSized() && DTy->isSized() &&
3180 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
3181 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
3182 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, DL,
3183 CtxI, DT, TLI, Visited);
3186 // Global variables which can't collapse to null are ok.
3187 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
3188 if (!GV->hasExternalWeakLinkage())
3189 return isAligned(V, Align, DL);
3191 // byval arguments are okay.
3192 if (const Argument *A = dyn_cast<Argument>(V))
3193 if (A->hasByValAttr())
3194 return isAligned(V, Align, DL);
3196 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
3197 return isAligned(V, Align, DL);
3199 // For GEPs, determine if the indexing lands within the allocated object.
3200 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3201 Type *VTy = GEP->getType();
3202 Type *Ty = VTy->getPointerElementType();
3203 const Value *Base = GEP->getPointerOperand();
3205 // Conservatively require that the base pointer be fully dereferenceable
3207 if (!Visited.insert(Base).second)
3209 if (!isDereferenceableAndAlignedPointer(Base, Align, DL, CtxI, DT, TLI,
3213 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
3214 if (!GEP->accumulateConstantOffset(DL, Offset))
3217 // Check if the load is within the bounds of the underlying object
3218 // and offset is aligned.
3219 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3220 Type *BaseType = Base->getType()->getPointerElementType();
3221 assert(isPowerOf2_32(Align) && "must be a power of 2!");
3222 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType)) &&
3223 !(Offset & APInt(Offset.getBitWidth(), Align-1));
3226 // For gc.relocate, look through relocations
3227 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3228 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3229 GCRelocateOperands RelocateInst(I);
3230 return isDereferenceableAndAlignedPointer(
3231 RelocateInst.getDerivedPtr(), Align, DL, CtxI, DT, TLI, Visited);
3234 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3235 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, DL,
3236 CtxI, DT, TLI, Visited);
3238 // If we don't know, assume the worst.
3242 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align,
3243 const DataLayout &DL,
3244 const Instruction *CtxI,
3245 const DominatorTree *DT,
3246 const TargetLibraryInfo *TLI) {
3247 // When dereferenceability information is provided by a dereferenceable
3248 // attribute, we know exactly how many bytes are dereferenceable. If we can
3249 // determine the exact offset to the attributed variable, we can use that
3250 // information here.
3251 Type *VTy = V->getType();
3252 Type *Ty = VTy->getPointerElementType();
3254 // Require ABI alignment for loads without alignment specification
3256 Align = DL.getABITypeAlignment(Ty);
3258 if (Ty->isSized()) {
3259 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3260 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3262 if (Offset.isNonNegative())
3263 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL, CtxI, DT, TLI) &&
3264 isAligned(BV, Offset, Align, DL))
3268 SmallPtrSet<const Value *, 32> Visited;
3269 return ::isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT, TLI,
3273 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3274 const Instruction *CtxI,
3275 const DominatorTree *DT,
3276 const TargetLibraryInfo *TLI) {
3277 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT, TLI);
3280 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3281 const Instruction *CtxI,
3282 const DominatorTree *DT,
3283 const TargetLibraryInfo *TLI) {
3284 const Operator *Inst = dyn_cast<Operator>(V);
3288 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3289 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3293 switch (Inst->getOpcode()) {
3296 case Instruction::UDiv:
3297 case Instruction::URem: {
3298 // x / y is undefined if y == 0.
3300 if (match(Inst->getOperand(1), m_APInt(V)))
3304 case Instruction::SDiv:
3305 case Instruction::SRem: {
3306 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3307 const APInt *Numerator, *Denominator;
3308 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3310 // We cannot hoist this division if the denominator is 0.
3311 if (*Denominator == 0)
3313 // It's safe to hoist if the denominator is not 0 or -1.
3314 if (*Denominator != -1)
3316 // At this point we know that the denominator is -1. It is safe to hoist as
3317 // long we know that the numerator is not INT_MIN.
3318 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3319 return !Numerator->isMinSignedValue();
3320 // The numerator *might* be MinSignedValue.
3323 case Instruction::Load: {
3324 const LoadInst *LI = cast<LoadInst>(Inst);
3325 if (!LI->isUnordered() ||
3326 // Speculative load may create a race that did not exist in the source.
3327 LI->getParent()->getParent()->hasFnAttribute(
3328 Attribute::SanitizeThread) ||
3329 // Speculative load may load data from dirty regions.
3330 LI->getParent()->getParent()->hasFnAttribute(
3331 Attribute::SanitizeAddress))
3333 const DataLayout &DL = LI->getModule()->getDataLayout();
3334 return isDereferenceableAndAlignedPointer(
3335 LI->getPointerOperand(), LI->getAlignment(), DL, CtxI, DT, TLI);
3337 case Instruction::Call: {
3338 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3339 switch (II->getIntrinsicID()) {
3340 // These synthetic intrinsics have no side-effects and just mark
3341 // information about their operands.
3342 // FIXME: There are other no-op synthetic instructions that potentially
3343 // should be considered at least *safe* to speculate...
3344 case Intrinsic::dbg_declare:
3345 case Intrinsic::dbg_value:
3348 case Intrinsic::bswap:
3349 case Intrinsic::ctlz:
3350 case Intrinsic::ctpop:
3351 case Intrinsic::cttz:
3352 case Intrinsic::objectsize:
3353 case Intrinsic::sadd_with_overflow:
3354 case Intrinsic::smul_with_overflow:
3355 case Intrinsic::ssub_with_overflow:
3356 case Intrinsic::uadd_with_overflow:
3357 case Intrinsic::umul_with_overflow:
3358 case Intrinsic::usub_with_overflow:
3360 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3361 // errno like libm sqrt would.
3362 case Intrinsic::sqrt:
3363 case Intrinsic::fma:
3364 case Intrinsic::fmuladd:
3365 case Intrinsic::fabs:
3366 case Intrinsic::minnum:
3367 case Intrinsic::maxnum:
3369 // TODO: some fp intrinsics are marked as having the same error handling
3370 // as libm. They're safe to speculate when they won't error.
3371 // TODO: are convert_{from,to}_fp16 safe?
3372 // TODO: can we list target-specific intrinsics here?
3376 return false; // The called function could have undefined behavior or
3377 // side-effects, even if marked readnone nounwind.
3379 case Instruction::VAArg:
3380 case Instruction::Alloca:
3381 case Instruction::Invoke:
3382 case Instruction::PHI:
3383 case Instruction::Store:
3384 case Instruction::Ret:
3385 case Instruction::Br:
3386 case Instruction::IndirectBr:
3387 case Instruction::Switch:
3388 case Instruction::Unreachable:
3389 case Instruction::Fence:
3390 case Instruction::AtomicRMW:
3391 case Instruction::AtomicCmpXchg:
3392 case Instruction::LandingPad:
3393 case Instruction::Resume:
3394 case Instruction::CatchPad:
3395 case Instruction::CatchEndPad:
3396 case Instruction::CatchRet:
3397 case Instruction::CleanupPad:
3398 case Instruction::CleanupEndPad:
3399 case Instruction::CleanupRet:
3400 case Instruction::TerminatePad:
3401 return false; // Misc instructions which have effects
3405 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3406 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3409 /// Return true if we know that the specified value is never null.
3410 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3411 assert(V->getType()->isPointerTy() && "V must be pointer type");
3413 // Alloca never returns null, malloc might.
3414 if (isa<AllocaInst>(V)) return true;
3416 // A byval, inalloca, or nonnull argument is never null.
3417 if (const Argument *A = dyn_cast<Argument>(V))
3418 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3420 // A global variable in address space 0 is non null unless extern weak.
3421 // Other address spaces may have null as a valid address for a global,
3422 // so we can't assume anything.
3423 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3424 return !GV->hasExternalWeakLinkage() &&
3425 GV->getType()->getAddressSpace() == 0;
3427 // A Load tagged w/nonnull metadata is never null.
3428 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3429 return LI->getMetadata(LLVMContext::MD_nonnull);
3431 if (auto CS = ImmutableCallSite(V))
3432 if (CS.isReturnNonNull())
3435 // operator new never returns null.
3436 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3442 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3443 const Instruction *CtxI,
3444 const DominatorTree *DT) {
3445 assert(V->getType()->isPointerTy() && "V must be pointer type");
3447 unsigned NumUsesExplored = 0;
3448 for (auto U : V->users()) {
3449 // Avoid massive lists
3450 if (NumUsesExplored >= DomConditionsMaxUses)
3453 // Consider only compare instructions uniquely controlling a branch
3454 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3458 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3461 for (auto *CmpU : Cmp->users()) {
3462 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3466 assert(BI->isConditional() && "uses a comparison!");
3468 BasicBlock *NonNullSuccessor = nullptr;
3469 CmpInst::Predicate Pred;
3471 if (match(const_cast<ICmpInst*>(Cmp),
3472 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3473 if (Pred == ICmpInst::ICMP_EQ)
3474 NonNullSuccessor = BI->getSuccessor(1);
3475 else if (Pred == ICmpInst::ICMP_NE)
3476 NonNullSuccessor = BI->getSuccessor(0);
3479 if (NonNullSuccessor) {
3480 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3481 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3490 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3491 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3492 if (isKnownNonNull(V, TLI))
3495 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3498 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3499 const DataLayout &DL,
3500 AssumptionCache *AC,
3501 const Instruction *CxtI,
3502 const DominatorTree *DT) {
3503 // Multiplying n * m significant bits yields a result of n + m significant
3504 // bits. If the total number of significant bits does not exceed the
3505 // result bit width (minus 1), there is no overflow.
3506 // This means if we have enough leading zero bits in the operands
3507 // we can guarantee that the result does not overflow.
3508 // Ref: "Hacker's Delight" by Henry Warren
3509 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3510 APInt LHSKnownZero(BitWidth, 0);
3511 APInt LHSKnownOne(BitWidth, 0);
3512 APInt RHSKnownZero(BitWidth, 0);
3513 APInt RHSKnownOne(BitWidth, 0);
3514 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3516 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3518 // Note that underestimating the number of zero bits gives a more
3519 // conservative answer.
3520 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3521 RHSKnownZero.countLeadingOnes();
3522 // First handle the easy case: if we have enough zero bits there's
3523 // definitely no overflow.
3524 if (ZeroBits >= BitWidth)
3525 return OverflowResult::NeverOverflows;
3527 // Get the largest possible values for each operand.
3528 APInt LHSMax = ~LHSKnownZero;
3529 APInt RHSMax = ~RHSKnownZero;
3531 // We know the multiply operation doesn't overflow if the maximum values for
3532 // each operand will not overflow after we multiply them together.
3534 LHSMax.umul_ov(RHSMax, MaxOverflow);
3536 return OverflowResult::NeverOverflows;
3538 // We know it always overflows if multiplying the smallest possible values for
3539 // the operands also results in overflow.
3541 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3543 return OverflowResult::AlwaysOverflows;
3545 return OverflowResult::MayOverflow;
3548 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3549 const DataLayout &DL,
3550 AssumptionCache *AC,
3551 const Instruction *CxtI,
3552 const DominatorTree *DT) {
3553 bool LHSKnownNonNegative, LHSKnownNegative;
3554 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3556 if (LHSKnownNonNegative || LHSKnownNegative) {
3557 bool RHSKnownNonNegative, RHSKnownNegative;
3558 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3561 if (LHSKnownNegative && RHSKnownNegative) {
3562 // The sign bit is set in both cases: this MUST overflow.
3563 // Create a simple add instruction, and insert it into the struct.
3564 return OverflowResult::AlwaysOverflows;
3567 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3568 // The sign bit is clear in both cases: this CANNOT overflow.
3569 // Create a simple add instruction, and insert it into the struct.
3570 return OverflowResult::NeverOverflows;
3574 return OverflowResult::MayOverflow;
3577 static OverflowResult computeOverflowForSignedAdd(
3578 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3579 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3580 if (Add && Add->hasNoSignedWrap()) {
3581 return OverflowResult::NeverOverflows;
3584 bool LHSKnownNonNegative, LHSKnownNegative;
3585 bool RHSKnownNonNegative, RHSKnownNegative;
3586 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3588 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3591 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3592 (LHSKnownNegative && RHSKnownNonNegative)) {
3593 // The sign bits are opposite: this CANNOT overflow.
3594 return OverflowResult::NeverOverflows;
3597 // The remaining code needs Add to be available. Early returns if not so.
3599 return OverflowResult::MayOverflow;
3601 // If the sign of Add is the same as at least one of the operands, this add
3602 // CANNOT overflow. This is particularly useful when the sum is
3603 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3605 bool LHSOrRHSKnownNonNegative =
3606 (LHSKnownNonNegative || RHSKnownNonNegative);
3607 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3608 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3609 bool AddKnownNonNegative, AddKnownNegative;
3610 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3611 /*Depth=*/0, AC, CxtI, DT);
3612 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3613 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3614 return OverflowResult::NeverOverflows;
3618 return OverflowResult::MayOverflow;
3621 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3622 const DataLayout &DL,
3623 AssumptionCache *AC,
3624 const Instruction *CxtI,
3625 const DominatorTree *DT) {
3626 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3627 Add, DL, AC, CxtI, DT);
3630 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3631 const DataLayout &DL,
3632 AssumptionCache *AC,
3633 const Instruction *CxtI,
3634 const DominatorTree *DT) {
3635 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3638 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3639 // FIXME: This conservative implementation can be relaxed. E.g. most
3640 // atomic operations are guaranteed to terminate on most platforms
3641 // and most functions terminate.
3643 return !I->isAtomic() && // atomics may never succeed on some platforms
3644 !isa<CallInst>(I) && // could throw and might not terminate
3645 !isa<InvokeInst>(I) && // might not terminate and could throw to
3646 // non-successor (see bug 24185 for details).
3647 !isa<ResumeInst>(I) && // has no successors
3648 !isa<ReturnInst>(I); // has no successors
3651 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3653 // The loop header is guaranteed to be executed for every iteration.
3655 // FIXME: Relax this constraint to cover all basic blocks that are
3656 // guaranteed to be executed at every iteration.
3657 if (I->getParent() != L->getHeader()) return false;
3659 for (const Instruction &LI : *L->getHeader()) {
3660 if (&LI == I) return true;
3661 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3663 llvm_unreachable("Instruction not contained in its own parent basic block.");
3666 bool llvm::propagatesFullPoison(const Instruction *I) {
3667 switch (I->getOpcode()) {
3668 case Instruction::Add:
3669 case Instruction::Sub:
3670 case Instruction::Xor:
3671 case Instruction::Trunc:
3672 case Instruction::BitCast:
3673 case Instruction::AddrSpaceCast:
3674 // These operations all propagate poison unconditionally. Note that poison
3675 // is not any particular value, so xor or subtraction of poison with
3676 // itself still yields poison, not zero.
3679 case Instruction::AShr:
3680 case Instruction::SExt:
3681 // For these operations, one bit of the input is replicated across
3682 // multiple output bits. A replicated poison bit is still poison.
3685 case Instruction::Shl: {
3686 // Left shift *by* a poison value is poison. The number of
3687 // positions to shift is unsigned, so no negative values are
3688 // possible there. Left shift by zero places preserves poison. So
3689 // it only remains to consider left shift of poison by a positive
3690 // number of places.
3692 // A left shift by a positive number of places leaves the lowest order bit
3693 // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3694 // make the poison operand violate that flag, yielding a fresh full-poison
3696 auto *OBO = cast<OverflowingBinaryOperator>(I);
3697 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3700 case Instruction::Mul: {
3701 // A multiplication by zero yields a non-poison zero result, so we need to
3702 // rule out zero as an operand. Conservatively, multiplication by a
3703 // non-zero constant is not multiplication by zero.
3705 // Multiplication by a non-zero constant can leave some bits
3706 // non-poisoned. For example, a multiplication by 2 leaves the lowest
3707 // order bit unpoisoned. So we need to consider that.
3709 // Multiplication by 1 preserves poison. If the multiplication has a
3710 // no-wrap flag, then we can make the poison operand violate that flag
3711 // when multiplied by any integer other than 0 and 1.
3712 auto *OBO = cast<OverflowingBinaryOperator>(I);
3713 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3714 for (Value *V : OBO->operands()) {
3715 if (auto *CI = dyn_cast<ConstantInt>(V)) {
3716 // A ConstantInt cannot yield poison, so we can assume that it is
3717 // the other operand that is poison.
3718 return !CI->isZero();
3725 case Instruction::GetElementPtr:
3726 // A GEP implicitly represents a sequence of additions, subtractions,
3727 // truncations, sign extensions and multiplications. The multiplications
3728 // are by the non-zero sizes of some set of types, so we do not have to be
3729 // concerned with multiplication by zero. If the GEP is in-bounds, then
3730 // these operations are implicitly no-signed-wrap so poison is propagated
3731 // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3732 return cast<GEPOperator>(I)->isInBounds();
3739 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3740 switch (I->getOpcode()) {
3741 case Instruction::Store:
3742 return cast<StoreInst>(I)->getPointerOperand();
3744 case Instruction::Load:
3745 return cast<LoadInst>(I)->getPointerOperand();
3747 case Instruction::AtomicCmpXchg:
3748 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3750 case Instruction::AtomicRMW:
3751 return cast<AtomicRMWInst>(I)->getPointerOperand();
3753 case Instruction::UDiv:
3754 case Instruction::SDiv:
3755 case Instruction::URem:
3756 case Instruction::SRem:
3757 return I->getOperand(1);
3764 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3765 // We currently only look for uses of poison values within the same basic
3766 // block, as that makes it easier to guarantee that the uses will be
3767 // executed given that PoisonI is executed.
3769 // FIXME: Expand this to consider uses beyond the same basic block. To do
3770 // this, look out for the distinction between post-dominance and strong
3772 const BasicBlock *BB = PoisonI->getParent();
3774 // Set of instructions that we have proved will yield poison if PoisonI
3776 SmallSet<const Value *, 16> YieldsPoison;
3777 YieldsPoison.insert(PoisonI);
3779 for (BasicBlock::const_iterator I = PoisonI->getIterator(), E = BB->end();
3781 if (&*I != PoisonI) {
3782 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&*I);
3783 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) return true;
3784 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
3788 // Mark poison that propagates from I through uses of I.
3789 if (YieldsPoison.count(&*I)) {
3790 for (const User *User : I->users()) {
3791 const Instruction *UserI = cast<Instruction>(User);
3792 if (UserI->getParent() == BB && propagatesFullPoison(UserI))
3793 YieldsPoison.insert(User);
3800 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3804 if (auto *C = dyn_cast<ConstantFP>(V))
3809 static bool isKnownNonZero(Value *V) {
3810 if (auto *C = dyn_cast<ConstantFP>(V))
3811 return !C->isZero();
3815 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3817 Value *CmpLHS, Value *CmpRHS,
3818 Value *TrueVal, Value *FalseVal,
3819 Value *&LHS, Value *&RHS) {
3823 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3824 // return inconsistent results between implementations.
3825 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3826 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3827 // Therefore we behave conservatively and only proceed if at least one of the
3828 // operands is known to not be zero, or if we don't care about signed zeroes.
3831 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3832 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3833 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3834 !isKnownNonZero(CmpRHS))
3835 return {SPF_UNKNOWN, SPNB_NA, false};
3838 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3839 bool Ordered = false;
3841 // When given one NaN and one non-NaN input:
3842 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3843 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3844 // ordered comparison fails), which could be NaN or non-NaN.
3845 // so here we discover exactly what NaN behavior is required/accepted.
3846 if (CmpInst::isFPPredicate(Pred)) {
3847 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3848 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3850 if (LHSSafe && RHSSafe) {
3851 // Both operands are known non-NaN.
3852 NaNBehavior = SPNB_RETURNS_ANY;
3853 } else if (CmpInst::isOrdered(Pred)) {
3854 // An ordered comparison will return false when given a NaN, so it
3858 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3859 NaNBehavior = SPNB_RETURNS_NAN;
3861 NaNBehavior = SPNB_RETURNS_OTHER;
3863 // Completely unsafe.
3864 return {SPF_UNKNOWN, SPNB_NA, false};
3867 // An unordered comparison will return true when given a NaN, so it
3870 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3871 NaNBehavior = SPNB_RETURNS_OTHER;
3873 NaNBehavior = SPNB_RETURNS_NAN;
3875 // Completely unsafe.
3876 return {SPF_UNKNOWN, SPNB_NA, false};
3880 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3881 std::swap(CmpLHS, CmpRHS);
3882 Pred = CmpInst::getSwappedPredicate(Pred);
3883 if (NaNBehavior == SPNB_RETURNS_NAN)
3884 NaNBehavior = SPNB_RETURNS_OTHER;
3885 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3886 NaNBehavior = SPNB_RETURNS_NAN;
3890 // ([if]cmp X, Y) ? X : Y
3891 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3893 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3894 case ICmpInst::ICMP_UGT:
3895 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3896 case ICmpInst::ICMP_SGT:
3897 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3898 case ICmpInst::ICMP_ULT:
3899 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3900 case ICmpInst::ICMP_SLT:
3901 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3902 case FCmpInst::FCMP_UGT:
3903 case FCmpInst::FCMP_UGE:
3904 case FCmpInst::FCMP_OGT:
3905 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3906 case FCmpInst::FCMP_ULT:
3907 case FCmpInst::FCMP_ULE:
3908 case FCmpInst::FCMP_OLT:
3909 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3913 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3914 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3915 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3917 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3918 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3919 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3920 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
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_SLT && (C1->isZero() || C1->isOne())) {
3926 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3930 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3931 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3932 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3933 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3934 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3937 return {SPF_SMIN, SPNB_NA, false};
3942 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3944 return {SPF_UNKNOWN, SPNB_NA, false};
3947 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
3948 Instruction::CastOps *CastOp) {
3949 CastInst *CI = dyn_cast<CastInst>(V1);
3950 Constant *C = dyn_cast<Constant>(V2);
3951 CastInst *CI2 = dyn_cast<CastInst>(V2);
3954 *CastOp = CI->getOpcode();
3957 // If V1 and V2 are both the same cast from the same type, we can look
3959 if (CI2->getOpcode() == CI->getOpcode() &&
3960 CI2->getSrcTy() == CI->getSrcTy())
3961 return CI2->getOperand(0);
3967 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3968 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3969 // This is only valid if the truncated value can be sign-extended
3970 // back to the original value.
3971 if (ConstantExpr::getSExt(T, C->getType()) == C)
3975 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3976 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3978 if (isa<TruncInst>(CI))
3979 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3981 if (isa<FPToUIInst>(CI))
3982 return ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
3984 if (isa<FPToSIInst>(CI))
3985 return ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
3987 if (isa<UIToFPInst>(CI))
3988 return ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
3990 if (isa<SIToFPInst>(CI))
3991 return ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
3993 if (isa<FPTruncInst>(CI))
3994 return ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
3996 if (isa<FPExtInst>(CI))
3997 return ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
4002 SelectPatternResult llvm::matchSelectPattern(Value *V,
4003 Value *&LHS, Value *&RHS,
4004 Instruction::CastOps *CastOp) {
4005 SelectInst *SI = dyn_cast<SelectInst>(V);
4006 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4008 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4009 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4011 CmpInst::Predicate Pred = CmpI->getPredicate();
4012 Value *CmpLHS = CmpI->getOperand(0);
4013 Value *CmpRHS = CmpI->getOperand(1);
4014 Value *TrueVal = SI->getTrueValue();
4015 Value *FalseVal = SI->getFalseValue();
4017 if (isa<FPMathOperator>(CmpI))
4018 FMF = CmpI->getFastMathFlags();
4021 if (CmpI->isEquality())
4022 return {SPF_UNKNOWN, SPNB_NA, false};
4024 // Deal with type mismatches.
4025 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4026 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
4027 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4028 cast<CastInst>(TrueVal)->getOperand(0), C,
4030 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
4031 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4032 C, cast<CastInst>(FalseVal)->getOperand(0),
4035 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4039 ConstantRange llvm::getConstantRangeFromMetadata(MDNode &Ranges) {
4040 const unsigned NumRanges = Ranges.getNumOperands() / 2;
4041 assert(NumRanges >= 1 && "Must have at least one range!");
4042 assert(Ranges.getNumOperands() % 2 == 0 && "Must be a sequence of pairs");
4044 auto *FirstLow = mdconst::extract<ConstantInt>(Ranges.getOperand(0));
4045 auto *FirstHigh = mdconst::extract<ConstantInt>(Ranges.getOperand(1));
4047 ConstantRange CR(FirstLow->getValue(), FirstHigh->getValue());
4049 for (unsigned i = 1; i < NumRanges; ++i) {
4050 auto *Low = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
4051 auto *High = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
4053 // Note: unionWith will potentially create a range that contains values not
4054 // contained in any of the original N ranges.
4055 CR = CR.unionWith(ConstantRange(Low->getValue(), High->getValue()));