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 choosen 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(2000));
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 static bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
189 unsigned Depth, const Query &Q);
191 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
192 unsigned Depth, AssumptionCache *AC,
193 const Instruction *CxtI, const DominatorTree *DT) {
194 return ::MaskedValueIsZero(V, Mask, DL, Depth,
195 Query(AC, safeCxtI(V, CxtI), DT));
198 static unsigned ComputeNumSignBits(Value *V, const DataLayout &DL,
199 unsigned Depth, const Query &Q);
201 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL,
202 unsigned Depth, AssumptionCache *AC,
203 const Instruction *CxtI,
204 const DominatorTree *DT) {
205 return ::ComputeNumSignBits(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
208 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
209 APInt &KnownZero, APInt &KnownOne,
210 APInt &KnownZero2, APInt &KnownOne2,
211 const DataLayout &DL, unsigned Depth,
214 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
215 // We know that the top bits of C-X are clear if X contains less bits
216 // than C (i.e. no wrap-around can happen). For example, 20-X is
217 // positive if we can prove that X is >= 0 and < 16.
218 if (!CLHS->getValue().isNegative()) {
219 unsigned BitWidth = KnownZero.getBitWidth();
220 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
221 // NLZ can't be BitWidth with no sign bit
222 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
223 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
225 // If all of the MaskV bits are known to be zero, then we know the
226 // output top bits are zero, because we now know that the output is
228 if ((KnownZero2 & MaskV) == MaskV) {
229 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
230 // Top bits known zero.
231 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
237 unsigned BitWidth = KnownZero.getBitWidth();
239 // If an initial sequence of bits in the result is not needed, the
240 // corresponding bits in the operands are not needed.
241 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
242 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, DL, Depth + 1, Q);
243 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
245 // Carry in a 1 for a subtract, rather than a 0.
246 APInt CarryIn(BitWidth, 0);
248 // Sum = LHS + ~RHS + 1
249 std::swap(KnownZero2, KnownOne2);
253 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
254 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
256 // Compute known bits of the carry.
257 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
258 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
260 // Compute set of known bits (where all three relevant bits are known).
261 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
262 APInt RHSKnown = KnownZero2 | KnownOne2;
263 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
264 APInt Known = LHSKnown & RHSKnown & CarryKnown;
266 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
267 "known bits of sum differ");
269 // Compute known bits of the result.
270 KnownZero = ~PossibleSumOne & Known;
271 KnownOne = PossibleSumOne & Known;
273 // Are we still trying to solve for the sign bit?
274 if (!Known.isNegative()) {
276 // Adding two non-negative numbers, or subtracting a negative number from
277 // a non-negative one, can't wrap into negative.
278 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
279 KnownZero |= APInt::getSignBit(BitWidth);
280 // Adding two negative numbers, or subtracting a non-negative number from
281 // a negative one, can't wrap into non-negative.
282 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
283 KnownOne |= APInt::getSignBit(BitWidth);
288 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
289 APInt &KnownZero, APInt &KnownOne,
290 APInt &KnownZero2, APInt &KnownOne2,
291 const DataLayout &DL, unsigned Depth,
293 unsigned BitWidth = KnownZero.getBitWidth();
294 computeKnownBits(Op1, KnownZero, KnownOne, DL, Depth + 1, Q);
295 computeKnownBits(Op0, KnownZero2, KnownOne2, DL, Depth + 1, Q);
297 bool isKnownNegative = false;
298 bool isKnownNonNegative = false;
299 // If the multiplication is known not to overflow, compute the sign bit.
302 // The product of a number with itself is non-negative.
303 isKnownNonNegative = true;
305 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
306 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
307 bool isKnownNegativeOp1 = KnownOne.isNegative();
308 bool isKnownNegativeOp0 = KnownOne2.isNegative();
309 // The product of two numbers with the same sign is non-negative.
310 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
311 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
312 // The product of a negative number and a non-negative number is either
314 if (!isKnownNonNegative)
315 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
316 isKnownNonZero(Op0, DL, Depth, Q)) ||
317 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
318 isKnownNonZero(Op1, DL, Depth, Q));
322 // If low bits are zero in either operand, output low known-0 bits.
323 // Also compute a conserative estimate for high known-0 bits.
324 // More trickiness is possible, but this is sufficient for the
325 // interesting case of alignment computation.
326 KnownOne.clearAllBits();
327 unsigned TrailZ = KnownZero.countTrailingOnes() +
328 KnownZero2.countTrailingOnes();
329 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
330 KnownZero2.countLeadingOnes(),
331 BitWidth) - BitWidth;
333 TrailZ = std::min(TrailZ, BitWidth);
334 LeadZ = std::min(LeadZ, BitWidth);
335 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
336 APInt::getHighBitsSet(BitWidth, LeadZ);
338 // Only make use of no-wrap flags if we failed to compute the sign bit
339 // directly. This matters if the multiplication always overflows, in
340 // which case we prefer to follow the result of the direct computation,
341 // though as the program is invoking undefined behaviour we can choose
342 // whatever we like here.
343 if (isKnownNonNegative && !KnownOne.isNegative())
344 KnownZero.setBit(BitWidth - 1);
345 else if (isKnownNegative && !KnownZero.isNegative())
346 KnownOne.setBit(BitWidth - 1);
349 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
351 unsigned BitWidth = KnownZero.getBitWidth();
352 unsigned NumRanges = Ranges.getNumOperands() / 2;
353 assert(NumRanges >= 1);
355 // Use the high end of the ranges to find leading zeros.
356 unsigned MinLeadingZeros = BitWidth;
357 for (unsigned i = 0; i < NumRanges; ++i) {
359 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
361 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
362 ConstantRange Range(Lower->getValue(), Upper->getValue());
363 if (Range.isWrappedSet())
364 MinLeadingZeros = 0; // -1 has no zeros
365 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
366 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
369 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
372 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
373 SmallVector<const Value *, 16> WorkSet(1, I);
374 SmallPtrSet<const Value *, 32> Visited;
375 SmallPtrSet<const Value *, 16> EphValues;
377 while (!WorkSet.empty()) {
378 const Value *V = WorkSet.pop_back_val();
379 if (!Visited.insert(V).second)
382 // If all uses of this value are ephemeral, then so is this value.
383 bool FoundNEUse = false;
384 for (const User *I : V->users())
385 if (!EphValues.count(I)) {
395 if (const User *U = dyn_cast<User>(V))
396 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
398 if (isSafeToSpeculativelyExecute(*J))
399 WorkSet.push_back(*J);
407 // Is this an intrinsic that cannot be speculated but also cannot trap?
408 static bool isAssumeLikeIntrinsic(const Instruction *I) {
409 if (const CallInst *CI = dyn_cast<CallInst>(I))
410 if (Function *F = CI->getCalledFunction())
411 switch (F->getIntrinsicID()) {
413 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
414 case Intrinsic::assume:
415 case Intrinsic::dbg_declare:
416 case Intrinsic::dbg_value:
417 case Intrinsic::invariant_start:
418 case Intrinsic::invariant_end:
419 case Intrinsic::lifetime_start:
420 case Intrinsic::lifetime_end:
421 case Intrinsic::objectsize:
422 case Intrinsic::ptr_annotation:
423 case Intrinsic::var_annotation:
430 static bool isValidAssumeForContext(Value *V, const Query &Q) {
431 Instruction *Inv = cast<Instruction>(V);
433 // There are two restrictions on the use of an assume:
434 // 1. The assume must dominate the context (or the control flow must
435 // reach the assume whenever it reaches the context).
436 // 2. The context must not be in the assume's set of ephemeral values
437 // (otherwise we will use the assume to prove that the condition
438 // feeding the assume is trivially true, thus causing the removal of
442 if (Q.DT->dominates(Inv, Q.CxtI)) {
444 } else if (Inv->getParent() == Q.CxtI->getParent()) {
445 // The context comes first, but they're both in the same block. Make sure
446 // there is nothing in between that might interrupt the control flow.
447 for (BasicBlock::const_iterator I =
448 std::next(BasicBlock::const_iterator(Q.CxtI)),
449 IE(Inv); I != IE; ++I)
450 if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I))
453 return !isEphemeralValueOf(Inv, Q.CxtI);
459 // When we don't have a DT, we do a limited search...
460 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
462 } else if (Inv->getParent() == Q.CxtI->getParent()) {
463 // Search forward from the assume until we reach the context (or the end
464 // of the block); the common case is that the assume will come first.
465 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
466 IE = Inv->getParent()->end(); I != IE; ++I)
470 // The context must come first...
471 for (BasicBlock::const_iterator I =
472 std::next(BasicBlock::const_iterator(Q.CxtI)),
473 IE(Inv); I != IE; ++I)
474 if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I))
477 return !isEphemeralValueOf(Inv, Q.CxtI);
483 bool llvm::isValidAssumeForContext(const Instruction *I,
484 const Instruction *CxtI,
485 const DominatorTree *DT) {
486 return ::isValidAssumeForContext(const_cast<Instruction *>(I),
487 Query(nullptr, CxtI, DT));
490 template<typename LHS, typename RHS>
491 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
492 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
493 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
494 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
497 template<typename LHS, typename RHS>
498 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
499 BinaryOp_match<RHS, LHS, Instruction::And>>
500 m_c_And(const LHS &L, const RHS &R) {
501 return m_CombineOr(m_And(L, R), m_And(R, L));
504 template<typename LHS, typename RHS>
505 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
506 BinaryOp_match<RHS, LHS, Instruction::Or>>
507 m_c_Or(const LHS &L, const RHS &R) {
508 return m_CombineOr(m_Or(L, R), m_Or(R, L));
511 template<typename LHS, typename RHS>
512 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
513 BinaryOp_match<RHS, LHS, Instruction::Xor>>
514 m_c_Xor(const LHS &L, const RHS &R) {
515 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
518 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
519 /// true (at the context instruction.) This is mostly a utility function for
520 /// the prototype dominating conditions reasoning below.
521 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
524 const DataLayout &DL,
525 unsigned Depth, const Query &Q) {
526 Value *LHS = Cmp->getOperand(0);
527 Value *RHS = Cmp->getOperand(1);
528 // TODO: We could potentially be more aggressive here. This would be worth
529 // evaluating. If we can, explore commoning this code with the assume
531 if (LHS != V && RHS != V)
534 const unsigned BitWidth = KnownZero.getBitWidth();
536 switch (Cmp->getPredicate()) {
538 // We know nothing from this condition
540 // TODO: implement unsigned bound from below (known one bits)
541 // TODO: common condition check implementations with assumes
542 // TODO: implement other patterns from assume (e.g. V & B == A)
543 case ICmpInst::ICMP_SGT:
545 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
546 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
547 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
548 // We know that the sign bit is zero.
549 KnownZero |= APInt::getSignBit(BitWidth);
553 case ICmpInst::ICMP_EQ:
555 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
557 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
559 computeKnownBits(LHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
561 llvm_unreachable("missing use?");
562 KnownZero |= KnownZeroTemp;
563 KnownOne |= KnownOneTemp;
566 case ICmpInst::ICMP_ULE:
568 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
569 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
570 // The known zero bits carry over
571 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
572 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
575 case ICmpInst::ICMP_ULT:
577 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
578 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
579 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
580 // power of 2, then one more).
581 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
582 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
584 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
590 /// Compute known bits in 'V' from conditions which are known to be true along
591 /// all paths leading to the context instruction. In particular, look for
592 /// cases where one branch of an interesting condition dominates the context
593 /// instruction. This does not do general dataflow.
594 /// NOTE: This code is EXPERIMENTAL and currently off by default.
595 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
597 const DataLayout &DL,
600 // Need both the dominator tree and the query location to do anything useful
601 if (!Q.DT || !Q.CxtI)
603 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
605 // Avoid useless work
606 if (auto VI = dyn_cast<Instruction>(V))
607 if (VI->getParent() == Cxt->getParent())
610 // Note: We currently implement two options. It's not clear which of these
611 // will survive long term, we need data for that.
612 // Option 1 - Try walking the dominator tree looking for conditions which
613 // might apply. This works well for local conditions (loop guards, etc..),
614 // but not as well for things far from the context instruction (presuming a
615 // low max blocks explored). If we can set an high enough limit, this would
617 // Option 2 - We restrict out search to those conditions which are uses of
618 // the value we're interested in. This is independent of dom structure,
619 // but is slightly less powerful without looking through lots of use chains.
620 // It does handle conditions far from the context instruction (e.g. early
621 // function exits on entry) really well though.
623 // Option 1 - Search the dom tree
624 unsigned NumBlocksExplored = 0;
625 BasicBlock *Current = Cxt->getParent();
627 // Stop searching if we've gone too far up the chain
628 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
632 if (!Q.DT->getNode(Current)->getIDom())
634 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
636 // found function entry
639 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
640 if (!BI || BI->isUnconditional())
642 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
646 // We're looking for conditions that are guaranteed to hold at the context
647 // instruction. Finding a condition where one path dominates the context
648 // isn't enough because both the true and false cases could merge before
649 // the context instruction we're actually interested in. Instead, we need
650 // to ensure that the taken *edge* dominates the context instruction.
651 BasicBlock *BB0 = BI->getSuccessor(0);
652 BasicBlockEdge Edge(BI->getParent(), BB0);
653 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
656 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
660 // Option 2 - Search the other uses of V
661 unsigned NumUsesExplored = 0;
662 for (auto U : V->users()) {
663 // Avoid massive lists
664 if (NumUsesExplored >= DomConditionsMaxUses)
667 // Consider only compare instructions uniquely controlling a branch
668 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
672 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
675 for (auto *CmpU : Cmp->users()) {
676 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
677 if (!BI || BI->isUnconditional())
679 // We're looking for conditions that are guaranteed to hold at the
680 // context instruction. Finding a condition where one path dominates
681 // the context isn't enough because both the true and false cases could
682 // merge before the context instruction we're actually interested in.
683 // Instead, we need to ensure that the taken *edge* dominates the context
685 BasicBlock *BB0 = BI->getSuccessor(0);
686 BasicBlockEdge Edge(BI->getParent(), BB0);
687 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
690 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
696 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
697 APInt &KnownOne, const DataLayout &DL,
698 unsigned Depth, const Query &Q) {
699 // Use of assumptions is context-sensitive. If we don't have a context, we
701 if (!Q.AC || !Q.CxtI)
704 unsigned BitWidth = KnownZero.getBitWidth();
706 for (auto &AssumeVH : Q.AC->assumptions()) {
709 CallInst *I = cast<CallInst>(AssumeVH);
710 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
711 "Got assumption for the wrong function!");
712 if (Q.ExclInvs.count(I))
715 // Warning: This loop can end up being somewhat performance sensetive.
716 // We're running this loop for once for each value queried resulting in a
717 // runtime of ~O(#assumes * #values).
719 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
720 "must be an assume intrinsic");
722 Value *Arg = I->getArgOperand(0);
724 if (Arg == V && isValidAssumeForContext(I, Q)) {
725 assert(BitWidth == 1 && "assume operand is not i1?");
726 KnownZero.clearAllBits();
727 KnownOne.setAllBits();
731 // The remaining tests are all recursive, so bail out if we hit the limit.
732 if (Depth == MaxDepth)
736 auto m_V = m_CombineOr(m_Specific(V),
737 m_CombineOr(m_PtrToInt(m_Specific(V)),
738 m_BitCast(m_Specific(V))));
740 CmpInst::Predicate Pred;
743 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
744 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
745 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
746 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
747 KnownZero |= RHSKnownZero;
748 KnownOne |= RHSKnownOne;
750 } else if (match(Arg,
751 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
752 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
753 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
754 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
755 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
756 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
758 // For those bits in the mask that are known to be one, we can propagate
759 // known bits from the RHS to V.
760 KnownZero |= RHSKnownZero & MaskKnownOne;
761 KnownOne |= RHSKnownOne & MaskKnownOne;
762 // assume(~(v & b) = a)
763 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
765 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
766 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
767 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
768 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
769 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
771 // For those bits in the mask that are known to be one, we can propagate
772 // inverted known bits from the RHS to V.
773 KnownZero |= RHSKnownOne & MaskKnownOne;
774 KnownOne |= RHSKnownZero & MaskKnownOne;
776 } else if (match(Arg,
777 m_c_ICmp(Pred, m_c_Or(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 BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
782 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
784 // For those bits in B that are known to be zero, we can propagate known
785 // bits from the RHS to V.
786 KnownZero |= RHSKnownZero & BKnownZero;
787 KnownOne |= RHSKnownOne & BKnownZero;
788 // assume(~(v | b) = a)
789 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(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 BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
795 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
797 // For those bits in B that are known to be zero, we can propagate
798 // inverted known bits from the RHS to V.
799 KnownZero |= RHSKnownOne & BKnownZero;
800 KnownOne |= RHSKnownZero & BKnownZero;
802 } else if (match(Arg,
803 m_c_ICmp(Pred, m_c_Xor(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. For those bits in B that are known to be one,
812 // we can propagate inverted known bits from the RHS to V.
813 KnownZero |= RHSKnownZero & BKnownZero;
814 KnownOne |= RHSKnownOne & BKnownZero;
815 KnownZero |= RHSKnownOne & BKnownOne;
816 KnownOne |= RHSKnownZero & BKnownOne;
817 // assume(~(v ^ b) = a)
818 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
820 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
821 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
822 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
823 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
824 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
826 // For those bits in B that are known to be zero, we can propagate
827 // inverted known bits from the RHS to V. For those bits in B that are
828 // known to be one, we can propagate known bits from the RHS to V.
829 KnownZero |= RHSKnownOne & BKnownZero;
830 KnownOne |= RHSKnownZero & BKnownZero;
831 KnownZero |= RHSKnownZero & BKnownOne;
832 KnownOne |= RHSKnownOne & BKnownOne;
833 // assume(v << c = a)
834 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
836 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
837 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
838 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
839 // For those bits in RHS that are known, we can propagate them to known
840 // bits in V shifted to the right by C.
841 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
842 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
843 // assume(~(v << c) = a)
844 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
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 // For those bits in RHS that are known, we can propagate them inverted
850 // to known bits in V shifted to the right by C.
851 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
852 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
853 // assume(v >> c = a)
854 } else if (match(Arg,
855 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
856 m_AShr(m_V, m_ConstantInt(C))),
858 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
859 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
860 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
861 // For those bits in RHS that are known, we can propagate them to known
862 // bits in V shifted to the right by C.
863 KnownZero |= RHSKnownZero << C->getZExtValue();
864 KnownOne |= RHSKnownOne << C->getZExtValue();
865 // assume(~(v >> c) = a)
866 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
867 m_LShr(m_V, m_ConstantInt(C)),
868 m_AShr(m_V, m_ConstantInt(C)))),
870 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
871 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
872 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
873 // For those bits in RHS that are known, we can propagate them inverted
874 // to known bits in V shifted to the right by C.
875 KnownZero |= RHSKnownOne << C->getZExtValue();
876 KnownOne |= RHSKnownZero << C->getZExtValue();
877 // assume(v >=_s c) where c is non-negative
878 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
879 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
880 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
881 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
883 if (RHSKnownZero.isNegative()) {
884 // We know that the sign bit is zero.
885 KnownZero |= APInt::getSignBit(BitWidth);
887 // assume(v >_s c) where c is at least -1.
888 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
889 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
890 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
891 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
893 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
894 // We know that the sign bit is zero.
895 KnownZero |= APInt::getSignBit(BitWidth);
897 // assume(v <=_s c) where c is negative
898 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
899 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
900 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
901 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
903 if (RHSKnownOne.isNegative()) {
904 // We know that the sign bit is one.
905 KnownOne |= APInt::getSignBit(BitWidth);
907 // assume(v <_s c) where c is non-positive
908 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
909 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
910 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
911 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
913 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
914 // We know that the sign bit is one.
915 KnownOne |= APInt::getSignBit(BitWidth);
918 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
919 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
920 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
921 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
923 // Whatever high bits in c are zero are known to be zero.
925 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
927 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
928 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
929 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
930 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
932 // Whatever high bits in c are zero are known to be zero (if c is a power
933 // of 2, then one more).
934 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
936 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
939 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
944 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
945 APInt &KnownOne, const DataLayout &DL,
946 unsigned Depth, const Query &Q) {
947 unsigned BitWidth = KnownZero.getBitWidth();
949 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
950 switch (I->getOpcode()) {
952 case Instruction::Load:
953 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
954 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
956 case Instruction::And: {
957 // If either the LHS or the RHS are Zero, the result is zero.
958 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
959 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
961 // Output known-1 bits are only known if set in both the LHS & RHS.
962 KnownOne &= KnownOne2;
963 // Output known-0 are known to be clear if zero in either the LHS | RHS.
964 KnownZero |= KnownZero2;
967 case Instruction::Or: {
968 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
969 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
971 // Output known-0 bits are only known if clear in both the LHS & RHS.
972 KnownZero &= KnownZero2;
973 // Output known-1 are known to be set if set in either the LHS | RHS.
974 KnownOne |= KnownOne2;
977 case Instruction::Xor: {
978 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
979 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
981 // Output known-0 bits are known if clear or set in both the LHS & RHS.
982 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
983 // Output known-1 are known to be set if set in only one of the LHS, RHS.
984 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
985 KnownZero = KnownZeroOut;
988 case Instruction::Mul: {
989 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
990 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
991 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
994 case Instruction::UDiv: {
995 // For the purposes of computing leading zeros we can conservatively
996 // treat a udiv as a logical right shift by the power of 2 known to
997 // be less than the denominator.
998 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
999 unsigned LeadZ = KnownZero2.countLeadingOnes();
1001 KnownOne2.clearAllBits();
1002 KnownZero2.clearAllBits();
1003 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1004 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1005 if (RHSUnknownLeadingOnes != BitWidth)
1006 LeadZ = std::min(BitWidth,
1007 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1009 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1012 case Instruction::Select:
1013 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1014 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1016 // Only known if known in both the LHS and RHS.
1017 KnownOne &= KnownOne2;
1018 KnownZero &= KnownZero2;
1020 case Instruction::FPTrunc:
1021 case Instruction::FPExt:
1022 case Instruction::FPToUI:
1023 case Instruction::FPToSI:
1024 case Instruction::SIToFP:
1025 case Instruction::UIToFP:
1026 break; // Can't work with floating point.
1027 case Instruction::PtrToInt:
1028 case Instruction::IntToPtr:
1029 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1030 // FALL THROUGH and handle them the same as zext/trunc.
1031 case Instruction::ZExt:
1032 case Instruction::Trunc: {
1033 Type *SrcTy = I->getOperand(0)->getType();
1035 unsigned SrcBitWidth;
1036 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1037 // which fall through here.
1038 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1040 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1041 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1042 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1043 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1044 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1045 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1046 // Any top bits are known to be zero.
1047 if (BitWidth > SrcBitWidth)
1048 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1051 case Instruction::BitCast: {
1052 Type *SrcTy = I->getOperand(0)->getType();
1053 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
1054 // TODO: For now, not handling conversions like:
1055 // (bitcast i64 %x to <2 x i32>)
1056 !I->getType()->isVectorTy()) {
1057 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1062 case Instruction::SExt: {
1063 // Compute the bits in the result that are not present in the input.
1064 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1066 KnownZero = KnownZero.trunc(SrcBitWidth);
1067 KnownOne = KnownOne.trunc(SrcBitWidth);
1068 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1069 KnownZero = KnownZero.zext(BitWidth);
1070 KnownOne = KnownOne.zext(BitWidth);
1072 // If the sign bit of the input is known set or clear, then we know the
1073 // top bits of the result.
1074 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1075 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1076 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1077 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1080 case Instruction::Shl:
1081 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1082 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1083 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1084 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1085 KnownZero <<= ShiftAmt;
1086 KnownOne <<= ShiftAmt;
1087 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1090 case Instruction::LShr:
1091 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1092 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1093 // Compute the new bits that are at the top now.
1094 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1096 // Unsigned shift right.
1097 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1098 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1099 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1100 // high bits known zero.
1101 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1104 case Instruction::AShr:
1105 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1106 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1107 // Compute the new bits that are at the top now.
1108 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1110 // Signed shift right.
1111 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1112 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1113 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1115 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1116 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1117 KnownZero |= HighBits;
1118 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1119 KnownOne |= HighBits;
1122 case Instruction::Sub: {
1123 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1124 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1125 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1129 case Instruction::Add: {
1130 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1131 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1132 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1136 case Instruction::SRem:
1137 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1138 APInt RA = Rem->getValue().abs();
1139 if (RA.isPowerOf2()) {
1140 APInt LowBits = RA - 1;
1141 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1144 // The low bits of the first operand are unchanged by the srem.
1145 KnownZero = KnownZero2 & LowBits;
1146 KnownOne = KnownOne2 & LowBits;
1148 // If the first operand is non-negative or has all low bits zero, then
1149 // the upper bits are all zero.
1150 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1151 KnownZero |= ~LowBits;
1153 // If the first operand is negative and not all low bits are zero, then
1154 // the upper bits are all one.
1155 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1156 KnownOne |= ~LowBits;
1158 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1162 // The sign bit is the LHS's sign bit, except when the result of the
1163 // remainder is zero.
1164 if (KnownZero.isNonNegative()) {
1165 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1166 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1168 // If it's known zero, our sign bit is also zero.
1169 if (LHSKnownZero.isNegative())
1170 KnownZero.setBit(BitWidth - 1);
1174 case Instruction::URem: {
1175 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1176 APInt RA = Rem->getValue();
1177 if (RA.isPowerOf2()) {
1178 APInt LowBits = (RA - 1);
1179 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1181 KnownZero |= ~LowBits;
1182 KnownOne &= LowBits;
1187 // Since the result is less than or equal to either operand, any leading
1188 // zero bits in either operand must also exist in the result.
1189 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1190 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1192 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1193 KnownZero2.countLeadingOnes());
1194 KnownOne.clearAllBits();
1195 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1199 case Instruction::Alloca: {
1200 AllocaInst *AI = cast<AllocaInst>(I);
1201 unsigned Align = AI->getAlignment();
1203 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1206 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1209 case Instruction::GetElementPtr: {
1210 // Analyze all of the subscripts of this getelementptr instruction
1211 // to determine if we can prove known low zero bits.
1212 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1213 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1215 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1217 gep_type_iterator GTI = gep_type_begin(I);
1218 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1219 Value *Index = I->getOperand(i);
1220 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1221 // Handle struct member offset arithmetic.
1223 // Handle case when index is vector zeroinitializer
1224 Constant *CIndex = cast<Constant>(Index);
1225 if (CIndex->isZeroValue())
1228 if (CIndex->getType()->isVectorTy())
1229 Index = CIndex->getSplatValue();
1231 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1232 const StructLayout *SL = DL.getStructLayout(STy);
1233 uint64_t Offset = SL->getElementOffset(Idx);
1234 TrailZ = std::min<unsigned>(TrailZ,
1235 countTrailingZeros(Offset));
1237 // Handle array index arithmetic.
1238 Type *IndexedTy = GTI.getIndexedType();
1239 if (!IndexedTy->isSized()) {
1243 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1244 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1245 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1246 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1248 TrailZ = std::min(TrailZ,
1249 unsigned(countTrailingZeros(TypeSize) +
1250 LocalKnownZero.countTrailingOnes()));
1254 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1257 case Instruction::PHI: {
1258 PHINode *P = cast<PHINode>(I);
1259 // Handle the case of a simple two-predecessor recurrence PHI.
1260 // There's a lot more that could theoretically be done here, but
1261 // this is sufficient to catch some interesting cases.
1262 if (P->getNumIncomingValues() == 2) {
1263 for (unsigned i = 0; i != 2; ++i) {
1264 Value *L = P->getIncomingValue(i);
1265 Value *R = P->getIncomingValue(!i);
1266 Operator *LU = dyn_cast<Operator>(L);
1269 unsigned Opcode = LU->getOpcode();
1270 // Check for operations that have the property that if
1271 // both their operands have low zero bits, the result
1272 // will have low zero bits.
1273 if (Opcode == Instruction::Add ||
1274 Opcode == Instruction::Sub ||
1275 Opcode == Instruction::And ||
1276 Opcode == Instruction::Or ||
1277 Opcode == Instruction::Mul) {
1278 Value *LL = LU->getOperand(0);
1279 Value *LR = LU->getOperand(1);
1280 // Find a recurrence.
1287 // Ok, we have a PHI of the form L op= R. Check for low
1289 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1291 // We need to take the minimum number of known bits
1292 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1293 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1295 KnownZero = APInt::getLowBitsSet(BitWidth,
1296 std::min(KnownZero2.countTrailingOnes(),
1297 KnownZero3.countTrailingOnes()));
1303 // Unreachable blocks may have zero-operand PHI nodes.
1304 if (P->getNumIncomingValues() == 0)
1307 // Otherwise take the unions of the known bit sets of the operands,
1308 // taking conservative care to avoid excessive recursion.
1309 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1310 // Skip if every incoming value references to ourself.
1311 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1314 KnownZero = APInt::getAllOnesValue(BitWidth);
1315 KnownOne = APInt::getAllOnesValue(BitWidth);
1316 for (Value *IncValue : P->incoming_values()) {
1317 // Skip direct self references.
1318 if (IncValue == P) continue;
1320 KnownZero2 = APInt(BitWidth, 0);
1321 KnownOne2 = APInt(BitWidth, 0);
1322 // Recurse, but cap the recursion to one level, because we don't
1323 // want to waste time spinning around in loops.
1324 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1326 KnownZero &= KnownZero2;
1327 KnownOne &= KnownOne2;
1328 // If all bits have been ruled out, there's no need to check
1330 if (!KnownZero && !KnownOne)
1336 case Instruction::Call:
1337 case Instruction::Invoke:
1338 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1339 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1340 // If a range metadata is attached to this IntrinsicInst, intersect the
1341 // explicit range specified by the metadata and the implicit range of
1343 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1344 switch (II->getIntrinsicID()) {
1346 case Intrinsic::ctlz:
1347 case Intrinsic::cttz: {
1348 unsigned LowBits = Log2_32(BitWidth)+1;
1349 // If this call is undefined for 0, the result will be less than 2^n.
1350 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1352 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1355 case Intrinsic::ctpop: {
1356 unsigned LowBits = Log2_32(BitWidth)+1;
1357 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1360 case Intrinsic::x86_sse42_crc32_64_64:
1361 KnownZero |= APInt::getHighBitsSet(64, 32);
1366 case Instruction::ExtractValue:
1367 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1368 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1369 if (EVI->getNumIndices() != 1) break;
1370 if (EVI->getIndices()[0] == 0) {
1371 switch (II->getIntrinsicID()) {
1373 case Intrinsic::uadd_with_overflow:
1374 case Intrinsic::sadd_with_overflow:
1375 computeKnownBitsAddSub(true, II->getArgOperand(0),
1376 II->getArgOperand(1), false, KnownZero,
1377 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1379 case Intrinsic::usub_with_overflow:
1380 case Intrinsic::ssub_with_overflow:
1381 computeKnownBitsAddSub(false, II->getArgOperand(0),
1382 II->getArgOperand(1), false, KnownZero,
1383 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1385 case Intrinsic::umul_with_overflow:
1386 case Intrinsic::smul_with_overflow:
1387 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1388 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1397 /// Determine which bits of V are known to be either zero or one and return
1398 /// them in the KnownZero/KnownOne bit sets.
1400 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1401 /// we cannot optimize based on the assumption that it is zero without changing
1402 /// it to be an explicit zero. If we don't change it to zero, other code could
1403 /// optimized based on the contradictory assumption that it is non-zero.
1404 /// Because instcombine aggressively folds operations with undef args anyway,
1405 /// this won't lose us code quality.
1407 /// This function is defined on values with integer type, values with pointer
1408 /// type, and vectors of integers. In the case
1409 /// where V is a vector, known zero, and known one values are the
1410 /// same width as the vector element, and the bit is set only if it is true
1411 /// for all of the elements in the vector.
1412 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1413 const DataLayout &DL, unsigned Depth, const Query &Q) {
1414 assert(V && "No Value?");
1415 assert(Depth <= MaxDepth && "Limit Search Depth");
1416 unsigned BitWidth = KnownZero.getBitWidth();
1418 assert((V->getType()->isIntOrIntVectorTy() ||
1419 V->getType()->getScalarType()->isPointerTy()) &&
1420 "Not integer or pointer type!");
1421 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1422 (!V->getType()->isIntOrIntVectorTy() ||
1423 V->getType()->getScalarSizeInBits() == BitWidth) &&
1424 KnownZero.getBitWidth() == BitWidth &&
1425 KnownOne.getBitWidth() == BitWidth &&
1426 "V, KnownOne and KnownZero should have same BitWidth");
1428 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1429 // We know all of the bits for a constant!
1430 KnownOne = CI->getValue();
1431 KnownZero = ~KnownOne;
1434 // Null and aggregate-zero are all-zeros.
1435 if (isa<ConstantPointerNull>(V) ||
1436 isa<ConstantAggregateZero>(V)) {
1437 KnownOne.clearAllBits();
1438 KnownZero = APInt::getAllOnesValue(BitWidth);
1441 // Handle a constant vector by taking the intersection of the known bits of
1442 // each element. There is no real need to handle ConstantVector here, because
1443 // we don't handle undef in any particularly useful way.
1444 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1445 // We know that CDS must be a vector of integers. Take the intersection of
1447 KnownZero.setAllBits(); KnownOne.setAllBits();
1448 APInt Elt(KnownZero.getBitWidth(), 0);
1449 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1450 Elt = CDS->getElementAsInteger(i);
1457 // The address of an aligned GlobalValue has trailing zeros.
1458 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1459 unsigned Align = GO->getAlignment();
1461 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1462 Type *ObjectType = GVar->getType()->getElementType();
1463 if (ObjectType->isSized()) {
1464 // If the object is defined in the current Module, we'll be giving
1465 // it the preferred alignment. Otherwise, we have to assume that it
1466 // may only have the minimum ABI alignment.
1467 if (GVar->isStrongDefinitionForLinker())
1468 Align = DL.getPreferredAlignment(GVar);
1470 Align = DL.getABITypeAlignment(ObjectType);
1475 KnownZero = APInt::getLowBitsSet(BitWidth,
1476 countTrailingZeros(Align));
1478 KnownZero.clearAllBits();
1479 KnownOne.clearAllBits();
1483 if (Argument *A = dyn_cast<Argument>(V)) {
1484 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1486 if (!Align && A->hasStructRetAttr()) {
1487 // An sret parameter has at least the ABI alignment of the return type.
1488 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1489 if (EltTy->isSized())
1490 Align = DL.getABITypeAlignment(EltTy);
1494 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1496 KnownZero.clearAllBits();
1497 KnownOne.clearAllBits();
1499 // Don't give up yet... there might be an assumption that provides more
1501 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1503 // Or a dominating condition for that matter
1504 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1505 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL,
1510 // Start out not knowing anything.
1511 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1513 // Limit search depth.
1514 // All recursive calls that increase depth must come after this.
1515 if (Depth == MaxDepth)
1518 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1519 // the bits of its aliasee.
1520 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1521 if (!GA->mayBeOverridden())
1522 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1526 if (Operator *I = dyn_cast<Operator>(V))
1527 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1528 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1529 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1530 // computeKnownBitsFromOperator.
1532 // Check whether a nearby assume intrinsic can determine some known bits.
1533 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1535 // Check whether there's a dominating condition which implies something about
1536 // this value at the given context.
1537 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1538 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1541 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1544 /// Determine whether the sign bit is known to be zero or one.
1545 /// Convenience wrapper around computeKnownBits.
1546 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1547 const DataLayout &DL, unsigned Depth, const Query &Q) {
1548 unsigned BitWidth = getBitWidth(V->getType(), DL);
1554 APInt ZeroBits(BitWidth, 0);
1555 APInt OneBits(BitWidth, 0);
1556 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1557 KnownOne = OneBits[BitWidth - 1];
1558 KnownZero = ZeroBits[BitWidth - 1];
1561 /// Return true if the given value is known to have exactly one
1562 /// bit set when defined. For vectors return true if every element is known to
1563 /// be a power of two when defined. Supports values with integer or pointer
1564 /// types and vectors of integers.
1565 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1566 const Query &Q, const DataLayout &DL) {
1567 if (Constant *C = dyn_cast<Constant>(V)) {
1568 if (C->isNullValue())
1570 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1571 return CI->getValue().isPowerOf2();
1572 // TODO: Handle vector constants.
1575 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1576 // it is shifted off the end then the result is undefined.
1577 if (match(V, m_Shl(m_One(), m_Value())))
1580 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1581 // bottom. If it is shifted off the bottom then the result is undefined.
1582 if (match(V, m_LShr(m_SignBit(), m_Value())))
1585 // The remaining tests are all recursive, so bail out if we hit the limit.
1586 if (Depth++ == MaxDepth)
1589 Value *X = nullptr, *Y = nullptr;
1590 // A shift of a power of two is a power of two or zero.
1591 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1592 match(V, m_Shr(m_Value(X), m_Value()))))
1593 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1595 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1596 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1598 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1599 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1600 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1602 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1603 // A power of two and'd with anything is a power of two or zero.
1604 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1605 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1607 // X & (-X) is always a power of two or zero.
1608 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1613 // Adding a power-of-two or zero to the same power-of-two or zero yields
1614 // either the original power-of-two, a larger power-of-two or zero.
1615 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1616 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1617 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1618 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1619 match(X, m_And(m_Value(), m_Specific(Y))))
1620 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1622 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1623 match(Y, m_And(m_Value(), m_Specific(X))))
1624 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1627 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1628 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1629 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1631 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1632 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1633 // If i8 V is a power of two or zero:
1634 // ZeroBits: 1 1 1 0 1 1 1 1
1635 // ~ZeroBits: 0 0 0 1 0 0 0 0
1636 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1637 // If OrZero isn't set, we cannot give back a zero result.
1638 // Make sure either the LHS or RHS has a bit set.
1639 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1644 // An exact divide or right shift can only shift off zero bits, so the result
1645 // is a power of two only if the first operand is a power of two and not
1646 // copying a sign bit (sdiv int_min, 2).
1647 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1648 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1649 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1656 /// \brief Test whether a GEP's result is known to be non-null.
1658 /// Uses properties inherent in a GEP to try to determine whether it is known
1661 /// Currently this routine does not support vector GEPs.
1662 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1663 unsigned Depth, const Query &Q) {
1664 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1667 // FIXME: Support vector-GEPs.
1668 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1670 // If the base pointer is non-null, we cannot walk to a null address with an
1671 // inbounds GEP in address space zero.
1672 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1675 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1676 // If so, then the GEP cannot produce a null pointer, as doing so would
1677 // inherently violate the inbounds contract within address space zero.
1678 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1679 GTI != GTE; ++GTI) {
1680 // Struct types are easy -- they must always be indexed by a constant.
1681 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1682 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1683 unsigned ElementIdx = OpC->getZExtValue();
1684 const StructLayout *SL = DL.getStructLayout(STy);
1685 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1686 if (ElementOffset > 0)
1691 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1692 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1695 // Fast path the constant operand case both for efficiency and so we don't
1696 // increment Depth when just zipping down an all-constant GEP.
1697 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1703 // We post-increment Depth here because while isKnownNonZero increments it
1704 // as well, when we pop back up that increment won't persist. We don't want
1705 // to recurse 10k times just because we have 10k GEP operands. We don't
1706 // bail completely out because we want to handle constant GEPs regardless
1708 if (Depth++ >= MaxDepth)
1711 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1718 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1719 /// ensure that the value it's attached to is never Value? 'RangeType' is
1720 /// is the type of the value described by the range.
1721 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1722 const APInt& Value) {
1723 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1724 assert(NumRanges >= 1);
1725 for (unsigned i = 0; i < NumRanges; ++i) {
1726 ConstantInt *Lower =
1727 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1728 ConstantInt *Upper =
1729 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1730 ConstantRange Range(Lower->getValue(), Upper->getValue());
1731 if (Range.contains(Value))
1737 /// Return true if the given value is known to be non-zero when defined.
1738 /// For vectors return true if every element is known to be non-zero when
1739 /// defined. Supports values with integer or pointer type and vectors of
1741 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1743 if (Constant *C = dyn_cast<Constant>(V)) {
1744 if (C->isNullValue())
1746 if (isa<ConstantInt>(C))
1747 // Must be non-zero due to null test above.
1749 // TODO: Handle vectors
1753 if (Instruction* I = dyn_cast<Instruction>(V)) {
1754 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1755 // If the possible ranges don't contain zero, then the value is
1756 // definitely non-zero.
1757 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1758 const APInt ZeroValue(Ty->getBitWidth(), 0);
1759 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1765 // The remaining tests are all recursive, so bail out if we hit the limit.
1766 if (Depth++ >= MaxDepth)
1769 // Check for pointer simplifications.
1770 if (V->getType()->isPointerTy()) {
1771 if (isKnownNonNull(V))
1773 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1774 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1778 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1780 // X | Y != 0 if X != 0 or Y != 0.
1781 Value *X = nullptr, *Y = nullptr;
1782 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1783 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1785 // ext X != 0 if X != 0.
1786 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1787 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1789 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1790 // if the lowest bit is shifted off the end.
1791 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1792 // shl nuw can't remove any non-zero bits.
1793 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1794 if (BO->hasNoUnsignedWrap())
1795 return isKnownNonZero(X, DL, Depth, Q);
1797 APInt KnownZero(BitWidth, 0);
1798 APInt KnownOne(BitWidth, 0);
1799 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1803 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1804 // defined if the sign bit is shifted off the end.
1805 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1806 // shr exact can only shift out zero bits.
1807 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1809 return isKnownNonZero(X, DL, Depth, Q);
1811 bool XKnownNonNegative, XKnownNegative;
1812 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1816 // div exact can only produce a zero if the dividend is zero.
1817 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1818 return isKnownNonZero(X, DL, Depth, Q);
1821 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1822 bool XKnownNonNegative, XKnownNegative;
1823 bool YKnownNonNegative, YKnownNegative;
1824 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1825 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1827 // If X and Y are both non-negative (as signed values) then their sum is not
1828 // zero unless both X and Y are zero.
1829 if (XKnownNonNegative && YKnownNonNegative)
1830 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1833 // If X and Y are both negative (as signed values) then their sum is not
1834 // zero unless both X and Y equal INT_MIN.
1835 if (BitWidth && XKnownNegative && YKnownNegative) {
1836 APInt KnownZero(BitWidth, 0);
1837 APInt KnownOne(BitWidth, 0);
1838 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1839 // The sign bit of X is set. If some other bit is set then X is not equal
1841 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1842 if ((KnownOne & Mask) != 0)
1844 // The sign bit of Y is set. If some other bit is set then Y is not equal
1846 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1847 if ((KnownOne & Mask) != 0)
1851 // The sum of a non-negative number and a power of two is not zero.
1852 if (XKnownNonNegative &&
1853 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1855 if (YKnownNonNegative &&
1856 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1860 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1861 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1862 // If X and Y are non-zero then so is X * Y as long as the multiplication
1863 // does not overflow.
1864 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1865 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
1868 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1869 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1870 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
1871 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
1875 if (!BitWidth) return false;
1876 APInt KnownZero(BitWidth, 0);
1877 APInt KnownOne(BitWidth, 0);
1878 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1879 return KnownOne != 0;
1882 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1883 /// simplify operations downstream. Mask is known to be zero for bits that V
1886 /// This function is defined on values with integer type, values with pointer
1887 /// type, and vectors of integers. In the case
1888 /// where V is a vector, the mask, known zero, and known one values are the
1889 /// same width as the vector element, and the bit is set only if it is true
1890 /// for all of the elements in the vector.
1891 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
1892 unsigned Depth, const Query &Q) {
1893 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1894 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1895 return (KnownZero & Mask) == Mask;
1900 /// Return the number of times the sign bit of the register is replicated into
1901 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1902 /// (itself), but other cases can give us information. For example, immediately
1903 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1904 /// other, so we return 3.
1906 /// 'Op' must have a scalar integer type.
1908 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
1910 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
1912 unsigned FirstAnswer = 1;
1914 // Note that ConstantInt is handled by the general computeKnownBits case
1918 return 1; // Limit search depth.
1920 Operator *U = dyn_cast<Operator>(V);
1921 switch (Operator::getOpcode(V)) {
1923 case Instruction::SExt:
1924 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1925 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
1927 case Instruction::SDiv: {
1928 const APInt *Denominator;
1929 // sdiv X, C -> adds log(C) sign bits.
1930 if (match(U->getOperand(1), m_APInt(Denominator))) {
1932 // Ignore non-positive denominator.
1933 if (!Denominator->isStrictlyPositive())
1936 // Calculate the incoming numerator bits.
1937 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1939 // Add floor(log(C)) bits to the numerator bits.
1940 return std::min(TyBits, NumBits + Denominator->logBase2());
1945 case Instruction::SRem: {
1946 const APInt *Denominator;
1947 // srem X, C -> we know that the result is within [-C+1,C) when C is a
1948 // positive constant. This let us put a lower bound on the number of sign
1950 if (match(U->getOperand(1), m_APInt(Denominator))) {
1952 // Ignore non-positive denominator.
1953 if (!Denominator->isStrictlyPositive())
1956 // Calculate the incoming numerator bits. SRem by a positive constant
1957 // can't lower the number of sign bits.
1959 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1961 // Calculate the leading sign bit constraints by examining the
1962 // denominator. Given that the denominator is positive, there are two
1965 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
1966 // (1 << ceilLogBase2(C)).
1968 // 2. the numerator is negative. Then the result range is (-C,0] and
1969 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
1971 // Thus a lower bound on the number of sign bits is `TyBits -
1972 // ceilLogBase2(C)`.
1974 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
1975 return std::max(NumrBits, ResBits);
1980 case Instruction::AShr: {
1981 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1982 // ashr X, C -> adds C sign bits. Vectors too.
1984 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1985 Tmp += ShAmt->getZExtValue();
1986 if (Tmp > TyBits) Tmp = TyBits;
1990 case Instruction::Shl: {
1992 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1993 // shl destroys sign bits.
1994 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1995 Tmp2 = ShAmt->getZExtValue();
1996 if (Tmp2 >= TyBits || // Bad shift.
1997 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2002 case Instruction::And:
2003 case Instruction::Or:
2004 case Instruction::Xor: // NOT is handled here.
2005 // Logical binary ops preserve the number of sign bits at the worst.
2006 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2008 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2009 FirstAnswer = std::min(Tmp, Tmp2);
2010 // We computed what we know about the sign bits as our first
2011 // answer. Now proceed to the generic code that uses
2012 // computeKnownBits, and pick whichever answer is better.
2016 case Instruction::Select:
2017 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2018 if (Tmp == 1) return 1; // Early out.
2019 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2020 return std::min(Tmp, Tmp2);
2022 case Instruction::Add:
2023 // Add can have at most one carry bit. Thus we know that the output
2024 // is, at worst, one more bit than the inputs.
2025 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2026 if (Tmp == 1) return 1; // Early out.
2028 // Special case decrementing a value (ADD X, -1):
2029 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2030 if (CRHS->isAllOnesValue()) {
2031 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2032 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2035 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2037 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2040 // If we are subtracting one from a positive number, there is no carry
2041 // out of the result.
2042 if (KnownZero.isNegative())
2046 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2047 if (Tmp2 == 1) return 1;
2048 return std::min(Tmp, Tmp2)-1;
2050 case Instruction::Sub:
2051 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2052 if (Tmp2 == 1) return 1;
2055 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2056 if (CLHS->isNullValue()) {
2057 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2058 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2060 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2062 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2065 // If the input is known to be positive (the sign bit is known clear),
2066 // the output of the NEG has the same number of sign bits as the input.
2067 if (KnownZero.isNegative())
2070 // Otherwise, we treat this like a SUB.
2073 // Sub can have at most one carry bit. Thus we know that the output
2074 // is, at worst, one more bit than the inputs.
2075 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2076 if (Tmp == 1) return 1; // Early out.
2077 return std::min(Tmp, Tmp2)-1;
2079 case Instruction::PHI: {
2080 PHINode *PN = cast<PHINode>(U);
2081 unsigned NumIncomingValues = PN->getNumIncomingValues();
2082 // Don't analyze large in-degree PHIs.
2083 if (NumIncomingValues > 4) break;
2084 // Unreachable blocks may have zero-operand PHI nodes.
2085 if (NumIncomingValues == 0) break;
2087 // Take the minimum of all incoming values. This can't infinitely loop
2088 // because of our depth threshold.
2089 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2090 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2091 if (Tmp == 1) return Tmp;
2093 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2098 case Instruction::Trunc:
2099 // FIXME: it's tricky to do anything useful for this, but it is an important
2100 // case for targets like X86.
2104 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2105 // use this information.
2106 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2108 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2110 if (KnownZero.isNegative()) { // sign bit is 0
2112 } else if (KnownOne.isNegative()) { // sign bit is 1;
2119 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2120 // the number of identical bits in the top of the input value.
2122 Mask <<= Mask.getBitWidth()-TyBits;
2123 // Return # leading zeros. We use 'min' here in case Val was zero before
2124 // shifting. We don't want to return '64' as for an i32 "0".
2125 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2128 /// This function computes the integer multiple of Base that equals V.
2129 /// If successful, it returns true and returns the multiple in
2130 /// Multiple. If unsuccessful, it returns false. It looks
2131 /// through SExt instructions only if LookThroughSExt is true.
2132 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2133 bool LookThroughSExt, unsigned Depth) {
2134 const unsigned MaxDepth = 6;
2136 assert(V && "No Value?");
2137 assert(Depth <= MaxDepth && "Limit Search Depth");
2138 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2140 Type *T = V->getType();
2142 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2152 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2153 Constant *BaseVal = ConstantInt::get(T, Base);
2154 if (CO && CO == BaseVal) {
2156 Multiple = ConstantInt::get(T, 1);
2160 if (CI && CI->getZExtValue() % Base == 0) {
2161 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2165 if (Depth == MaxDepth) return false; // Limit search depth.
2167 Operator *I = dyn_cast<Operator>(V);
2168 if (!I) return false;
2170 switch (I->getOpcode()) {
2172 case Instruction::SExt:
2173 if (!LookThroughSExt) return false;
2174 // otherwise fall through to ZExt
2175 case Instruction::ZExt:
2176 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2177 LookThroughSExt, Depth+1);
2178 case Instruction::Shl:
2179 case Instruction::Mul: {
2180 Value *Op0 = I->getOperand(0);
2181 Value *Op1 = I->getOperand(1);
2183 if (I->getOpcode() == Instruction::Shl) {
2184 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2185 if (!Op1CI) return false;
2186 // Turn Op0 << Op1 into Op0 * 2^Op1
2187 APInt Op1Int = Op1CI->getValue();
2188 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2189 APInt API(Op1Int.getBitWidth(), 0);
2190 API.setBit(BitToSet);
2191 Op1 = ConstantInt::get(V->getContext(), API);
2194 Value *Mul0 = nullptr;
2195 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2196 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2197 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2198 if (Op1C->getType()->getPrimitiveSizeInBits() <
2199 MulC->getType()->getPrimitiveSizeInBits())
2200 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2201 if (Op1C->getType()->getPrimitiveSizeInBits() >
2202 MulC->getType()->getPrimitiveSizeInBits())
2203 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2205 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2206 Multiple = ConstantExpr::getMul(MulC, Op1C);
2210 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2211 if (Mul0CI->getValue() == 1) {
2212 // V == Base * Op1, so return Op1
2218 Value *Mul1 = nullptr;
2219 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2220 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2221 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2222 if (Op0C->getType()->getPrimitiveSizeInBits() <
2223 MulC->getType()->getPrimitiveSizeInBits())
2224 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2225 if (Op0C->getType()->getPrimitiveSizeInBits() >
2226 MulC->getType()->getPrimitiveSizeInBits())
2227 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2229 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2230 Multiple = ConstantExpr::getMul(MulC, Op0C);
2234 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2235 if (Mul1CI->getValue() == 1) {
2236 // V == Base * Op0, so return Op0
2244 // We could not determine if V is a multiple of Base.
2248 /// Return true if we can prove that the specified FP value is never equal to
2251 /// NOTE: this function will need to be revisited when we support non-default
2254 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2255 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2256 return !CFP->getValueAPF().isNegZero();
2258 // FIXME: Magic number! At the least, this should be given a name because it's
2259 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2260 // expose it as a parameter, so it can be used for testing / experimenting.
2262 return false; // Limit search depth.
2264 const Operator *I = dyn_cast<Operator>(V);
2265 if (!I) return false;
2267 // Check if the nsz fast-math flag is set
2268 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2269 if (FPO->hasNoSignedZeros())
2272 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2273 if (I->getOpcode() == Instruction::FAdd)
2274 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2275 if (CFP->isNullValue())
2278 // sitofp and uitofp turn into +0.0 for zero.
2279 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2282 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2283 // sqrt(-0.0) = -0.0, no other negative results are possible.
2284 if (II->getIntrinsicID() == Intrinsic::sqrt)
2285 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2287 if (const CallInst *CI = dyn_cast<CallInst>(I))
2288 if (const Function *F = CI->getCalledFunction()) {
2289 if (F->isDeclaration()) {
2291 if (F->getName() == "abs") return true;
2292 // fabs[lf](x) != -0.0
2293 if (F->getName() == "fabs") return true;
2294 if (F->getName() == "fabsf") return true;
2295 if (F->getName() == "fabsl") return true;
2296 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2297 F->getName() == "sqrtl")
2298 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2305 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2306 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2307 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2309 // FIXME: Magic number! At the least, this should be given a name because it's
2310 // used similarly in CannotBeNegativeZero(). A better fix may be to
2311 // expose it as a parameter, so it can be used for testing / experimenting.
2313 return false; // Limit search depth.
2315 const Operator *I = dyn_cast<Operator>(V);
2316 if (!I) return false;
2318 switch (I->getOpcode()) {
2320 case Instruction::FMul:
2321 // x*x is always non-negative or a NaN.
2322 if (I->getOperand(0) == I->getOperand(1))
2325 case Instruction::FAdd:
2326 case Instruction::FDiv:
2327 case Instruction::FRem:
2328 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2329 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2330 case Instruction::FPExt:
2331 case Instruction::FPTrunc:
2332 // Widening/narrowing never change sign.
2333 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2334 case Instruction::Call:
2335 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2336 switch (II->getIntrinsicID()) {
2338 case Intrinsic::exp:
2339 case Intrinsic::exp2:
2340 case Intrinsic::fabs:
2341 case Intrinsic::sqrt:
2343 case Intrinsic::powi:
2344 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2345 // powi(x,n) is non-negative if n is even.
2346 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2349 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2350 case Intrinsic::fma:
2351 case Intrinsic::fmuladd:
2352 // x*x+y is non-negative if y is non-negative.
2353 return I->getOperand(0) == I->getOperand(1) &&
2354 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2361 /// If the specified value can be set by repeating the same byte in memory,
2362 /// return the i8 value that it is represented with. This is
2363 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2364 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2365 /// byte store (e.g. i16 0x1234), return null.
2366 Value *llvm::isBytewiseValue(Value *V) {
2367 // All byte-wide stores are splatable, even of arbitrary variables.
2368 if (V->getType()->isIntegerTy(8)) return V;
2370 // Handle 'null' ConstantArrayZero etc.
2371 if (Constant *C = dyn_cast<Constant>(V))
2372 if (C->isNullValue())
2373 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2375 // Constant float and double values can be handled as integer values if the
2376 // corresponding integer value is "byteable". An important case is 0.0.
2377 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2378 if (CFP->getType()->isFloatTy())
2379 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2380 if (CFP->getType()->isDoubleTy())
2381 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2382 // Don't handle long double formats, which have strange constraints.
2385 // We can handle constant integers that are multiple of 8 bits.
2386 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2387 if (CI->getBitWidth() % 8 == 0) {
2388 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2390 if (!CI->getValue().isSplat(8))
2392 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2396 // A ConstantDataArray/Vector is splatable if all its members are equal and
2398 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2399 Value *Elt = CA->getElementAsConstant(0);
2400 Value *Val = isBytewiseValue(Elt);
2404 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2405 if (CA->getElementAsConstant(I) != Elt)
2411 // Conceptually, we could handle things like:
2412 // %a = zext i8 %X to i16
2413 // %b = shl i16 %a, 8
2414 // %c = or i16 %a, %b
2415 // but until there is an example that actually needs this, it doesn't seem
2416 // worth worrying about.
2421 // This is the recursive version of BuildSubAggregate. It takes a few different
2422 // arguments. Idxs is the index within the nested struct From that we are
2423 // looking at now (which is of type IndexedType). IdxSkip is the number of
2424 // indices from Idxs that should be left out when inserting into the resulting
2425 // struct. To is the result struct built so far, new insertvalue instructions
2427 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2428 SmallVectorImpl<unsigned> &Idxs,
2430 Instruction *InsertBefore) {
2431 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2433 // Save the original To argument so we can modify it
2435 // General case, the type indexed by Idxs is a struct
2436 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2437 // Process each struct element recursively
2440 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2444 // Couldn't find any inserted value for this index? Cleanup
2445 while (PrevTo != OrigTo) {
2446 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2447 PrevTo = Del->getAggregateOperand();
2448 Del->eraseFromParent();
2450 // Stop processing elements
2454 // If we successfully found a value for each of our subaggregates
2458 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2459 // the struct's elements had a value that was inserted directly. In the latter
2460 // case, perhaps we can't determine each of the subelements individually, but
2461 // we might be able to find the complete struct somewhere.
2463 // Find the value that is at that particular spot
2464 Value *V = FindInsertedValue(From, Idxs);
2469 // Insert the value in the new (sub) aggregrate
2470 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2471 "tmp", InsertBefore);
2474 // This helper takes a nested struct and extracts a part of it (which is again a
2475 // struct) into a new value. For example, given the struct:
2476 // { a, { b, { c, d }, e } }
2477 // and the indices "1, 1" this returns
2480 // It does this by inserting an insertvalue for each element in the resulting
2481 // struct, as opposed to just inserting a single struct. This will only work if
2482 // each of the elements of the substruct are known (ie, inserted into From by an
2483 // insertvalue instruction somewhere).
2485 // All inserted insertvalue instructions are inserted before InsertBefore
2486 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2487 Instruction *InsertBefore) {
2488 assert(InsertBefore && "Must have someplace to insert!");
2489 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2491 Value *To = UndefValue::get(IndexedType);
2492 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2493 unsigned IdxSkip = Idxs.size();
2495 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2498 /// Given an aggregrate and an sequence of indices, see if
2499 /// the scalar value indexed is already around as a register, for example if it
2500 /// were inserted directly into the aggregrate.
2502 /// If InsertBefore is not null, this function will duplicate (modified)
2503 /// insertvalues when a part of a nested struct is extracted.
2504 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2505 Instruction *InsertBefore) {
2506 // Nothing to index? Just return V then (this is useful at the end of our
2508 if (idx_range.empty())
2510 // We have indices, so V should have an indexable type.
2511 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2512 "Not looking at a struct or array?");
2513 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2514 "Invalid indices for type?");
2516 if (Constant *C = dyn_cast<Constant>(V)) {
2517 C = C->getAggregateElement(idx_range[0]);
2518 if (!C) return nullptr;
2519 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2522 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2523 // Loop the indices for the insertvalue instruction in parallel with the
2524 // requested indices
2525 const unsigned *req_idx = idx_range.begin();
2526 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2527 i != e; ++i, ++req_idx) {
2528 if (req_idx == idx_range.end()) {
2529 // We can't handle this without inserting insertvalues
2533 // The requested index identifies a part of a nested aggregate. Handle
2534 // this specially. For example,
2535 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2536 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2537 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2538 // This can be changed into
2539 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2540 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2541 // which allows the unused 0,0 element from the nested struct to be
2543 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2547 // This insert value inserts something else than what we are looking for.
2548 // See if the (aggregrate) value inserted into has the value we are
2549 // looking for, then.
2551 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2554 // If we end up here, the indices of the insertvalue match with those
2555 // requested (though possibly only partially). Now we recursively look at
2556 // the inserted value, passing any remaining indices.
2557 return FindInsertedValue(I->getInsertedValueOperand(),
2558 makeArrayRef(req_idx, idx_range.end()),
2562 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2563 // If we're extracting a value from an aggregrate that was extracted from
2564 // something else, we can extract from that something else directly instead.
2565 // However, we will need to chain I's indices with the requested indices.
2567 // Calculate the number of indices required
2568 unsigned size = I->getNumIndices() + idx_range.size();
2569 // Allocate some space to put the new indices in
2570 SmallVector<unsigned, 5> Idxs;
2572 // Add indices from the extract value instruction
2573 Idxs.append(I->idx_begin(), I->idx_end());
2575 // Add requested indices
2576 Idxs.append(idx_range.begin(), idx_range.end());
2578 assert(Idxs.size() == size
2579 && "Number of indices added not correct?");
2581 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2583 // Otherwise, we don't know (such as, extracting from a function return value
2584 // or load instruction)
2588 /// Analyze the specified pointer to see if it can be expressed as a base
2589 /// pointer plus a constant offset. Return the base and offset to the caller.
2590 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2591 const DataLayout &DL) {
2592 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2593 APInt ByteOffset(BitWidth, 0);
2595 if (Ptr->getType()->isVectorTy())
2598 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2599 APInt GEPOffset(BitWidth, 0);
2600 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2603 ByteOffset += GEPOffset;
2605 Ptr = GEP->getPointerOperand();
2606 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2607 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2608 Ptr = cast<Operator>(Ptr)->getOperand(0);
2609 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2610 if (GA->mayBeOverridden())
2612 Ptr = GA->getAliasee();
2617 Offset = ByteOffset.getSExtValue();
2622 /// This function computes the length of a null-terminated C string pointed to
2623 /// by V. If successful, it returns true and returns the string in Str.
2624 /// If unsuccessful, it returns false.
2625 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2626 uint64_t Offset, bool TrimAtNul) {
2629 // Look through bitcast instructions and geps.
2630 V = V->stripPointerCasts();
2632 // If the value is a GEP instruction or constant expression, treat it as an
2634 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2635 // Make sure the GEP has exactly three arguments.
2636 if (GEP->getNumOperands() != 3)
2639 // Make sure the index-ee is a pointer to array of i8.
2640 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2641 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2642 if (!AT || !AT->getElementType()->isIntegerTy(8))
2645 // Check to make sure that the first operand of the GEP is an integer and
2646 // has value 0 so that we are sure we're indexing into the initializer.
2647 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2648 if (!FirstIdx || !FirstIdx->isZero())
2651 // If the second index isn't a ConstantInt, then this is a variable index
2652 // into the array. If this occurs, we can't say anything meaningful about
2654 uint64_t StartIdx = 0;
2655 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2656 StartIdx = CI->getZExtValue();
2659 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2663 // The GEP instruction, constant or instruction, must reference a global
2664 // variable that is a constant and is initialized. The referenced constant
2665 // initializer is the array that we'll use for optimization.
2666 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2667 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2670 // Handle the all-zeros case
2671 if (GV->getInitializer()->isNullValue()) {
2672 // This is a degenerate case. The initializer is constant zero so the
2673 // length of the string must be zero.
2678 // Must be a Constant Array
2679 const ConstantDataArray *Array =
2680 dyn_cast<ConstantDataArray>(GV->getInitializer());
2681 if (!Array || !Array->isString())
2684 // Get the number of elements in the array
2685 uint64_t NumElts = Array->getType()->getArrayNumElements();
2687 // Start out with the entire array in the StringRef.
2688 Str = Array->getAsString();
2690 if (Offset > NumElts)
2693 // Skip over 'offset' bytes.
2694 Str = Str.substr(Offset);
2697 // Trim off the \0 and anything after it. If the array is not nul
2698 // terminated, we just return the whole end of string. The client may know
2699 // some other way that the string is length-bound.
2700 Str = Str.substr(0, Str.find('\0'));
2705 // These next two are very similar to the above, but also look through PHI
2707 // TODO: See if we can integrate these two together.
2709 /// If we can compute the length of the string pointed to by
2710 /// the specified pointer, return 'len+1'. If we can't, return 0.
2711 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2712 // Look through noop bitcast instructions.
2713 V = V->stripPointerCasts();
2715 // If this is a PHI node, there are two cases: either we have already seen it
2717 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2718 if (!PHIs.insert(PN).second)
2719 return ~0ULL; // already in the set.
2721 // If it was new, see if all the input strings are the same length.
2722 uint64_t LenSoFar = ~0ULL;
2723 for (Value *IncValue : PN->incoming_values()) {
2724 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2725 if (Len == 0) return 0; // Unknown length -> unknown.
2727 if (Len == ~0ULL) continue;
2729 if (Len != LenSoFar && LenSoFar != ~0ULL)
2730 return 0; // Disagree -> unknown.
2734 // Success, all agree.
2738 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2739 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2740 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2741 if (Len1 == 0) return 0;
2742 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2743 if (Len2 == 0) return 0;
2744 if (Len1 == ~0ULL) return Len2;
2745 if (Len2 == ~0ULL) return Len1;
2746 if (Len1 != Len2) return 0;
2750 // Otherwise, see if we can read the string.
2752 if (!getConstantStringInfo(V, StrData))
2755 return StrData.size()+1;
2758 /// If we can compute the length of the string pointed to by
2759 /// the specified pointer, return 'len+1'. If we can't, return 0.
2760 uint64_t llvm::GetStringLength(Value *V) {
2761 if (!V->getType()->isPointerTy()) return 0;
2763 SmallPtrSet<PHINode*, 32> PHIs;
2764 uint64_t Len = GetStringLengthH(V, PHIs);
2765 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2766 // an empty string as a length.
2767 return Len == ~0ULL ? 1 : Len;
2770 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2771 /// previous iteration of the loop was referring to the same object as \p PN.
2772 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2773 // Find the loop-defined value.
2774 Loop *L = LI->getLoopFor(PN->getParent());
2775 if (PN->getNumIncomingValues() != 2)
2778 // Find the value from previous iteration.
2779 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2780 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2781 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2782 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2785 // If a new pointer is loaded in the loop, the pointer references a different
2786 // object in every iteration. E.g.:
2790 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2791 if (!L->isLoopInvariant(Load->getPointerOperand()))
2796 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2797 unsigned MaxLookup) {
2798 if (!V->getType()->isPointerTy())
2800 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2801 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2802 V = GEP->getPointerOperand();
2803 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2804 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2805 V = cast<Operator>(V)->getOperand(0);
2806 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2807 if (GA->mayBeOverridden())
2809 V = GA->getAliasee();
2811 // See if InstructionSimplify knows any relevant tricks.
2812 if (Instruction *I = dyn_cast<Instruction>(V))
2813 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2814 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
2821 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2826 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
2827 const DataLayout &DL, LoopInfo *LI,
2828 unsigned MaxLookup) {
2829 SmallPtrSet<Value *, 4> Visited;
2830 SmallVector<Value *, 4> Worklist;
2831 Worklist.push_back(V);
2833 Value *P = Worklist.pop_back_val();
2834 P = GetUnderlyingObject(P, DL, MaxLookup);
2836 if (!Visited.insert(P).second)
2839 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2840 Worklist.push_back(SI->getTrueValue());
2841 Worklist.push_back(SI->getFalseValue());
2845 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2846 // If this PHI changes the underlying object in every iteration of the
2847 // loop, don't look through it. Consider:
2850 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
2854 // Prev is tracking Curr one iteration behind so they refer to different
2855 // underlying objects.
2856 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
2857 isSameUnderlyingObjectInLoop(PN, LI))
2858 for (Value *IncValue : PN->incoming_values())
2859 Worklist.push_back(IncValue);
2863 Objects.push_back(P);
2864 } while (!Worklist.empty());
2867 /// Return true if the only users of this pointer are lifetime markers.
2868 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2869 for (const User *U : V->users()) {
2870 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2871 if (!II) return false;
2873 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2874 II->getIntrinsicID() != Intrinsic::lifetime_end)
2880 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
2881 Type *Ty, const DataLayout &DL,
2882 const Instruction *CtxI,
2883 const DominatorTree *DT,
2884 const TargetLibraryInfo *TLI) {
2885 assert(Offset.isNonNegative() && "offset can't be negative");
2886 assert(Ty->isSized() && "must be sized");
2888 APInt DerefBytes(Offset.getBitWidth(), 0);
2889 bool CheckForNonNull = false;
2890 if (const Argument *A = dyn_cast<Argument>(BV)) {
2891 DerefBytes = A->getDereferenceableBytes();
2892 if (!DerefBytes.getBoolValue()) {
2893 DerefBytes = A->getDereferenceableOrNullBytes();
2894 CheckForNonNull = true;
2896 } else if (auto CS = ImmutableCallSite(BV)) {
2897 DerefBytes = CS.getDereferenceableBytes(0);
2898 if (!DerefBytes.getBoolValue()) {
2899 DerefBytes = CS.getDereferenceableOrNullBytes(0);
2900 CheckForNonNull = true;
2902 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
2903 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
2904 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2905 DerefBytes = CI->getLimitedValue();
2907 if (!DerefBytes.getBoolValue()) {
2909 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
2910 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2911 DerefBytes = CI->getLimitedValue();
2913 CheckForNonNull = true;
2917 if (DerefBytes.getBoolValue())
2918 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
2919 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
2925 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
2926 const Instruction *CtxI,
2927 const DominatorTree *DT,
2928 const TargetLibraryInfo *TLI) {
2929 Type *VTy = V->getType();
2930 Type *Ty = VTy->getPointerElementType();
2934 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
2935 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
2938 /// Return true if Value is always a dereferenceable pointer.
2940 /// Test if V is always a pointer to allocated and suitably aligned memory for
2941 /// a simple load or store.
2942 static bool isDereferenceablePointer(const Value *V, const DataLayout &DL,
2943 const Instruction *CtxI,
2944 const DominatorTree *DT,
2945 const TargetLibraryInfo *TLI,
2946 SmallPtrSetImpl<const Value *> &Visited) {
2947 // Note that it is not safe to speculate into a malloc'd region because
2948 // malloc may return null.
2950 // These are obviously ok.
2951 if (isa<AllocaInst>(V)) return true;
2953 // It's not always safe to follow a bitcast, for example:
2954 // bitcast i8* (alloca i8) to i32*
2955 // would result in a 4-byte load from a 1-byte alloca. However,
2956 // if we're casting from a pointer from a type of larger size
2957 // to a type of smaller size (or the same size), and the alignment
2958 // is at least as large as for the resulting pointer type, then
2959 // we can look through the bitcast.
2960 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
2961 Type *STy = BC->getSrcTy()->getPointerElementType(),
2962 *DTy = BC->getDestTy()->getPointerElementType();
2963 if (STy->isSized() && DTy->isSized() &&
2964 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
2965 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
2966 return isDereferenceablePointer(BC->getOperand(0), DL, CtxI,
2970 // Global variables which can't collapse to null are ok.
2971 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2972 return !GV->hasExternalWeakLinkage();
2974 // byval arguments are okay.
2975 if (const Argument *A = dyn_cast<Argument>(V))
2976 if (A->hasByValAttr())
2979 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
2982 // For GEPs, determine if the indexing lands within the allocated object.
2983 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2984 Type *VTy = GEP->getType();
2985 Type *Ty = VTy->getPointerElementType();
2986 const Value *Base = GEP->getPointerOperand();
2988 // Conservatively require that the base pointer be fully dereferenceable.
2989 if (!Visited.insert(Base).second)
2991 if (!isDereferenceablePointer(Base, DL, CtxI,
2995 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
2996 if (!GEP->accumulateConstantOffset(DL, Offset))
2999 // Check if the load is within the bounds of the underlying object.
3000 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3001 Type *BaseType = Base->getType()->getPointerElementType();
3002 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType));
3005 // For gc.relocate, look through relocations
3006 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3007 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3008 GCRelocateOperands RelocateInst(I);
3009 return isDereferenceablePointer(RelocateInst.getDerivedPtr(), DL, CtxI,
3013 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3014 return isDereferenceablePointer(ASC->getOperand(0), DL, CtxI,
3017 // If we don't know, assume the worst.
3021 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3022 const Instruction *CtxI,
3023 const DominatorTree *DT,
3024 const TargetLibraryInfo *TLI) {
3025 // When dereferenceability information is provided by a dereferenceable
3026 // attribute, we know exactly how many bytes are dereferenceable. If we can
3027 // determine the exact offset to the attributed variable, we can use that
3028 // information here.
3029 Type *VTy = V->getType();
3030 Type *Ty = VTy->getPointerElementType();
3031 if (Ty->isSized()) {
3032 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3033 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3035 if (Offset.isNonNegative())
3036 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL,
3041 SmallPtrSet<const Value *, 32> Visited;
3042 return ::isDereferenceablePointer(V, DL, CtxI, DT, TLI, Visited);
3045 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3046 const Instruction *CtxI,
3047 const DominatorTree *DT,
3048 const TargetLibraryInfo *TLI) {
3049 const Operator *Inst = dyn_cast<Operator>(V);
3053 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3054 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3058 switch (Inst->getOpcode()) {
3061 case Instruction::UDiv:
3062 case Instruction::URem: {
3063 // x / y is undefined if y == 0.
3065 if (match(Inst->getOperand(1), m_APInt(V)))
3069 case Instruction::SDiv:
3070 case Instruction::SRem: {
3071 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3072 const APInt *Numerator, *Denominator;
3073 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3075 // We cannot hoist this division if the denominator is 0.
3076 if (*Denominator == 0)
3078 // It's safe to hoist if the denominator is not 0 or -1.
3079 if (*Denominator != -1)
3081 // At this point we know that the denominator is -1. It is safe to hoist as
3082 // long we know that the numerator is not INT_MIN.
3083 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3084 return !Numerator->isMinSignedValue();
3085 // The numerator *might* be MinSignedValue.
3088 case Instruction::Load: {
3089 const LoadInst *LI = cast<LoadInst>(Inst);
3090 if (!LI->isUnordered() ||
3091 // Speculative load may create a race that did not exist in the source.
3092 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
3094 const DataLayout &DL = LI->getModule()->getDataLayout();
3095 return isDereferenceablePointer(LI->getPointerOperand(), DL, CtxI, DT, TLI);
3097 case Instruction::Call: {
3098 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3099 switch (II->getIntrinsicID()) {
3100 // These synthetic intrinsics have no side-effects and just mark
3101 // information about their operands.
3102 // FIXME: There are other no-op synthetic instructions that potentially
3103 // should be considered at least *safe* to speculate...
3104 case Intrinsic::dbg_declare:
3105 case Intrinsic::dbg_value:
3108 case Intrinsic::bswap:
3109 case Intrinsic::ctlz:
3110 case Intrinsic::ctpop:
3111 case Intrinsic::cttz:
3112 case Intrinsic::objectsize:
3113 case Intrinsic::sadd_with_overflow:
3114 case Intrinsic::smul_with_overflow:
3115 case Intrinsic::ssub_with_overflow:
3116 case Intrinsic::uadd_with_overflow:
3117 case Intrinsic::umul_with_overflow:
3118 case Intrinsic::usub_with_overflow:
3120 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3121 // errno like libm sqrt would.
3122 case Intrinsic::sqrt:
3123 case Intrinsic::fma:
3124 case Intrinsic::fmuladd:
3125 case Intrinsic::fabs:
3126 case Intrinsic::minnum:
3127 case Intrinsic::maxnum:
3129 // TODO: some fp intrinsics are marked as having the same error handling
3130 // as libm. They're safe to speculate when they won't error.
3131 // TODO: are convert_{from,to}_fp16 safe?
3132 // TODO: can we list target-specific intrinsics here?
3136 return false; // The called function could have undefined behavior or
3137 // side-effects, even if marked readnone nounwind.
3139 case Instruction::VAArg:
3140 case Instruction::Alloca:
3141 case Instruction::Invoke:
3142 case Instruction::PHI:
3143 case Instruction::Store:
3144 case Instruction::Ret:
3145 case Instruction::Br:
3146 case Instruction::IndirectBr:
3147 case Instruction::Switch:
3148 case Instruction::Unreachable:
3149 case Instruction::Fence:
3150 case Instruction::LandingPad:
3151 case Instruction::AtomicRMW:
3152 case Instruction::AtomicCmpXchg:
3153 case Instruction::Resume:
3154 return false; // Misc instructions which have effects
3158 /// Return true if we know that the specified value is never null.
3159 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3160 // Alloca never returns null, malloc might.
3161 if (isa<AllocaInst>(V)) return true;
3163 // A byval, inalloca, or nonnull argument is never null.
3164 if (const Argument *A = dyn_cast<Argument>(V))
3165 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3167 // Global values are not null unless extern weak.
3168 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3169 return !GV->hasExternalWeakLinkage();
3171 // A Load tagged w/nonnull metadata is never null.
3172 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3173 return LI->getMetadata(LLVMContext::MD_nonnull);
3175 if (auto CS = ImmutableCallSite(V))
3176 if (CS.isReturnNonNull())
3179 // operator new never returns null.
3180 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3186 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3187 const Instruction *CtxI,
3188 const DominatorTree *DT) {
3189 unsigned NumUsesExplored = 0;
3190 for (auto U : V->users()) {
3191 // Avoid massive lists
3192 if (NumUsesExplored >= DomConditionsMaxUses)
3195 // Consider only compare instructions uniquely controlling a branch
3196 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3200 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3203 for (auto *CmpU : Cmp->users()) {
3204 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3208 assert(BI->isConditional() && "uses a comparison!");
3210 BasicBlock *NonNullSuccessor = nullptr;
3211 CmpInst::Predicate Pred;
3213 if (match(const_cast<ICmpInst*>(Cmp),
3214 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3215 if (Pred == ICmpInst::ICMP_EQ)
3216 NonNullSuccessor = BI->getSuccessor(1);
3217 else if (Pred == ICmpInst::ICMP_NE)
3218 NonNullSuccessor = BI->getSuccessor(0);
3221 if (NonNullSuccessor) {
3222 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3223 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3232 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3233 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3234 if (isKnownNonNull(V, TLI))
3237 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3240 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3241 const DataLayout &DL,
3242 AssumptionCache *AC,
3243 const Instruction *CxtI,
3244 const DominatorTree *DT) {
3245 // Multiplying n * m significant bits yields a result of n + m significant
3246 // bits. If the total number of significant bits does not exceed the
3247 // result bit width (minus 1), there is no overflow.
3248 // This means if we have enough leading zero bits in the operands
3249 // we can guarantee that the result does not overflow.
3250 // Ref: "Hacker's Delight" by Henry Warren
3251 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3252 APInt LHSKnownZero(BitWidth, 0);
3253 APInt LHSKnownOne(BitWidth, 0);
3254 APInt RHSKnownZero(BitWidth, 0);
3255 APInt RHSKnownOne(BitWidth, 0);
3256 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3258 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3260 // Note that underestimating the number of zero bits gives a more
3261 // conservative answer.
3262 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3263 RHSKnownZero.countLeadingOnes();
3264 // First handle the easy case: if we have enough zero bits there's
3265 // definitely no overflow.
3266 if (ZeroBits >= BitWidth)
3267 return OverflowResult::NeverOverflows;
3269 // Get the largest possible values for each operand.
3270 APInt LHSMax = ~LHSKnownZero;
3271 APInt RHSMax = ~RHSKnownZero;
3273 // We know the multiply operation doesn't overflow if the maximum values for
3274 // each operand will not overflow after we multiply them together.
3276 LHSMax.umul_ov(RHSMax, MaxOverflow);
3278 return OverflowResult::NeverOverflows;
3280 // We know it always overflows if multiplying the smallest possible values for
3281 // the operands also results in overflow.
3283 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3285 return OverflowResult::AlwaysOverflows;
3287 return OverflowResult::MayOverflow;
3290 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3291 const DataLayout &DL,
3292 AssumptionCache *AC,
3293 const Instruction *CxtI,
3294 const DominatorTree *DT) {
3295 bool LHSKnownNonNegative, LHSKnownNegative;
3296 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3298 if (LHSKnownNonNegative || LHSKnownNegative) {
3299 bool RHSKnownNonNegative, RHSKnownNegative;
3300 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3303 if (LHSKnownNegative && RHSKnownNegative) {
3304 // The sign bit is set in both cases: this MUST overflow.
3305 // Create a simple add instruction, and insert it into the struct.
3306 return OverflowResult::AlwaysOverflows;
3309 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3310 // The sign bit is clear in both cases: this CANNOT overflow.
3311 // Create a simple add instruction, and insert it into the struct.
3312 return OverflowResult::NeverOverflows;
3316 return OverflowResult::MayOverflow;
3319 static SelectPatternFlavor matchSelectPattern(ICmpInst::Predicate Pred,
3320 Value *CmpLHS, Value *CmpRHS,
3321 Value *TrueVal, Value *FalseVal,
3322 Value *&LHS, Value *&RHS) {
3326 // (icmp X, Y) ? X : Y
3327 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3329 default: return SPF_UNKNOWN; // Equality.
3330 case ICmpInst::ICMP_UGT:
3331 case ICmpInst::ICMP_UGE: return SPF_UMAX;
3332 case ICmpInst::ICMP_SGT:
3333 case ICmpInst::ICMP_SGE: return SPF_SMAX;
3334 case ICmpInst::ICMP_ULT:
3335 case ICmpInst::ICMP_ULE: return SPF_UMIN;
3336 case ICmpInst::ICMP_SLT:
3337 case ICmpInst::ICMP_SLE: return SPF_SMIN;
3341 // (icmp X, Y) ? Y : X
3342 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3344 default: return SPF_UNKNOWN; // Equality.
3345 case ICmpInst::ICMP_UGT:
3346 case ICmpInst::ICMP_UGE: return SPF_UMIN;
3347 case ICmpInst::ICMP_SGT:
3348 case ICmpInst::ICMP_SGE: return SPF_SMIN;
3349 case ICmpInst::ICMP_ULT:
3350 case ICmpInst::ICMP_ULE: return SPF_UMAX;
3351 case ICmpInst::ICMP_SLT:
3352 case ICmpInst::ICMP_SLE: return SPF_SMAX;
3356 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3357 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3358 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3360 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3361 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3362 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3363 return (CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS;
3366 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3367 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3368 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3369 return (CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS;
3373 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3374 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3375 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3376 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3377 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3385 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3390 static Constant *lookThroughCast(ICmpInst *CmpI, Value *V1, Value *V2,
3391 Instruction::CastOps *CastOp) {
3392 CastInst *CI = dyn_cast<CastInst>(V1);
3393 Constant *C = dyn_cast<Constant>(V2);
3396 *CastOp = CI->getOpcode();
3398 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3399 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3400 // This is only valid if the truncated value can be sign-extended
3401 // back to the original value.
3402 if (ConstantExpr::getSExt(T, C->getType()) == C)
3406 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3407 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3409 if (isa<TruncInst>(CI))
3410 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3415 SelectPatternFlavor llvm::matchSelectPattern(Value *V,
3416 Value *&LHS, Value *&RHS,
3417 Instruction::CastOps *CastOp) {
3418 SelectInst *SI = dyn_cast<SelectInst>(V);
3419 if (!SI) return SPF_UNKNOWN;
3421 ICmpInst *CmpI = dyn_cast<ICmpInst>(SI->getCondition());
3422 if (!CmpI) return SPF_UNKNOWN;
3424 ICmpInst::Predicate Pred = CmpI->getPredicate();
3425 Value *CmpLHS = CmpI->getOperand(0);
3426 Value *CmpRHS = CmpI->getOperand(1);
3427 Value *TrueVal = SI->getTrueValue();
3428 Value *FalseVal = SI->getFalseValue();
3431 if (CmpI->isEquality())
3434 // Deal with type mismatches.
3435 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
3436 if (Constant *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
3437 return ::matchSelectPattern(Pred, CmpLHS, CmpRHS,
3438 cast<CastInst>(TrueVal)->getOperand(0), C,
3440 if (Constant *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
3441 return ::matchSelectPattern(Pred, CmpLHS, CmpRHS,
3442 C, cast<CastInst>(FalseVal)->getOperand(0),
3445 return ::matchSelectPattern(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal,