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 computeKnownBits(RHS, KnownZero, KnownOne, DL, Depth + 1, Q);
557 computeKnownBits(LHS, KnownZero, KnownOne, DL, Depth + 1, Q);
559 llvm_unreachable("missing use?");
561 case ICmpInst::ICMP_ULE:
563 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
564 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
565 // The known zero bits carry over
566 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
567 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
570 case ICmpInst::ICMP_ULT:
572 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
573 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
574 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
575 // power of 2, then one more).
576 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
577 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
579 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
585 /// Compute known bits in 'V' from conditions which are known to be true along
586 /// all paths leading to the context instruction. In particular, look for
587 /// cases where one branch of an interesting condition dominates the context
588 /// instruction. This does not do general dataflow.
589 /// NOTE: This code is EXPERIMENTAL and currently off by default.
590 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
592 const DataLayout &DL,
595 // Need both the dominator tree and the query location to do anything useful
596 if (!Q.DT || !Q.CxtI)
598 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
600 // Avoid useless work
601 if (auto VI = dyn_cast<Instruction>(V))
602 if (VI->getParent() == Cxt->getParent())
605 // Note: We currently implement two options. It's not clear which of these
606 // will survive long term, we need data for that.
607 // Option 1 - Try walking the dominator tree looking for conditions which
608 // might apply. This works well for local conditions (loop guards, etc..),
609 // but not as well for things far from the context instruction (presuming a
610 // low max blocks explored). If we can set an high enough limit, this would
612 // Option 2 - We restrict out search to those conditions which are uses of
613 // the value we're interested in. This is independent of dom structure,
614 // but is slightly less powerful without looking through lots of use chains.
615 // It does handle conditions far from the context instruction (e.g. early
616 // function exits on entry) really well though.
618 // Option 1 - Search the dom tree
619 unsigned NumBlocksExplored = 0;
620 BasicBlock *Current = Cxt->getParent();
622 // Stop searching if we've gone too far up the chain
623 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
627 if (!Q.DT->getNode(Current)->getIDom())
629 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
631 // found function entry
634 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
635 if (!BI || BI->isUnconditional())
637 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
641 // We're looking for conditions that are guaranteed to hold at the context
642 // instruction. Finding a condition where one path dominates the context
643 // isn't enough because both the true and false cases could merge before
644 // the context instruction we're actually interested in. Instead, we need
645 // to ensure that the taken *edge* dominates the context instruction.
646 BasicBlock *BB0 = BI->getSuccessor(0);
647 BasicBlockEdge Edge(BI->getParent(), BB0);
648 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
651 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
655 // Option 2 - Search the other uses of V
656 unsigned NumUsesExplored = 0;
657 for (auto U : V->users()) {
658 // Avoid massive lists
659 if (NumUsesExplored >= DomConditionsMaxUses)
662 // Consider only compare instructions uniquely controlling a branch
663 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
667 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
670 for (auto *CmpU : Cmp->users()) {
671 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
672 if (!BI || BI->isUnconditional())
674 // We're looking for conditions that are guaranteed to hold at the
675 // context instruction. Finding a condition where one path dominates
676 // the context isn't enough because both the true and false cases could
677 // merge before the context instruction we're actually interested in.
678 // Instead, we need to ensure that the taken *edge* dominates the context
680 BasicBlock *BB0 = BI->getSuccessor(0);
681 BasicBlockEdge Edge(BI->getParent(), BB0);
682 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
685 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
691 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
692 APInt &KnownOne, const DataLayout &DL,
693 unsigned Depth, const Query &Q) {
694 // Use of assumptions is context-sensitive. If we don't have a context, we
696 if (!Q.AC || !Q.CxtI)
699 unsigned BitWidth = KnownZero.getBitWidth();
701 for (auto &AssumeVH : Q.AC->assumptions()) {
704 CallInst *I = cast<CallInst>(AssumeVH);
705 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
706 "Got assumption for the wrong function!");
707 if (Q.ExclInvs.count(I))
710 // Warning: This loop can end up being somewhat performance sensetive.
711 // We're running this loop for once for each value queried resulting in a
712 // runtime of ~O(#assumes * #values).
714 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
715 "must be an assume intrinsic");
717 Value *Arg = I->getArgOperand(0);
719 if (Arg == V && isValidAssumeForContext(I, Q)) {
720 assert(BitWidth == 1 && "assume operand is not i1?");
721 KnownZero.clearAllBits();
722 KnownOne.setAllBits();
726 // The remaining tests are all recursive, so bail out if we hit the limit.
727 if (Depth == MaxDepth)
731 auto m_V = m_CombineOr(m_Specific(V),
732 m_CombineOr(m_PtrToInt(m_Specific(V)),
733 m_BitCast(m_Specific(V))));
735 CmpInst::Predicate Pred;
738 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
739 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
740 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
741 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
742 KnownZero |= RHSKnownZero;
743 KnownOne |= RHSKnownOne;
745 } else if (match(Arg,
746 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
747 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
748 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
749 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
750 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
751 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
753 // For those bits in the mask that are known to be one, we can propagate
754 // known bits from the RHS to V.
755 KnownZero |= RHSKnownZero & MaskKnownOne;
756 KnownOne |= RHSKnownOne & MaskKnownOne;
757 // assume(~(v & b) = a)
758 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
760 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
761 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
762 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
763 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
764 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
766 // For those bits in the mask that are known to be one, we can propagate
767 // inverted known bits from the RHS to V.
768 KnownZero |= RHSKnownOne & MaskKnownOne;
769 KnownOne |= RHSKnownZero & MaskKnownOne;
771 } else if (match(Arg,
772 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
773 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
774 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
775 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
776 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
777 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
779 // For those bits in B that are known to be zero, we can propagate known
780 // bits from the RHS to V.
781 KnownZero |= RHSKnownZero & BKnownZero;
782 KnownOne |= RHSKnownOne & BKnownZero;
783 // assume(~(v | b) = a)
784 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
786 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
787 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
788 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
789 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
790 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
792 // For those bits in B that are known to be zero, we can propagate
793 // inverted known bits from the RHS to V.
794 KnownZero |= RHSKnownOne & BKnownZero;
795 KnownOne |= RHSKnownZero & BKnownZero;
797 } else if (match(Arg,
798 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
799 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
800 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
801 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
802 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
803 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
805 // For those bits in B that are known to be zero, we can propagate known
806 // bits from the RHS to V. For those bits in B that are known to be one,
807 // we can propagate inverted known bits from the RHS to V.
808 KnownZero |= RHSKnownZero & BKnownZero;
809 KnownOne |= RHSKnownOne & BKnownZero;
810 KnownZero |= RHSKnownOne & BKnownOne;
811 KnownOne |= RHSKnownZero & BKnownOne;
812 // assume(~(v ^ b) = a)
813 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
815 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
816 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
817 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
818 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
819 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
821 // For those bits in B that are known to be zero, we can propagate
822 // inverted known bits from the RHS to V. For those bits in B that are
823 // known to be one, we can propagate known bits from the RHS to V.
824 KnownZero |= RHSKnownOne & BKnownZero;
825 KnownOne |= RHSKnownZero & BKnownZero;
826 KnownZero |= RHSKnownZero & BKnownOne;
827 KnownOne |= RHSKnownOne & BKnownOne;
828 // assume(v << c = a)
829 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
831 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
832 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
833 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
834 // For those bits in RHS that are known, we can propagate them to known
835 // bits in V shifted to the right by C.
836 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
837 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
838 // assume(~(v << c) = a)
839 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
841 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
842 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
843 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
844 // For those bits in RHS that are known, we can propagate them inverted
845 // to known bits in V shifted to the right by C.
846 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
847 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
848 // assume(v >> c = a)
849 } else if (match(Arg,
850 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
851 m_AShr(m_V, m_ConstantInt(C))),
853 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
854 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
855 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
856 // For those bits in RHS that are known, we can propagate them to known
857 // bits in V shifted to the right by C.
858 KnownZero |= RHSKnownZero << C->getZExtValue();
859 KnownOne |= RHSKnownOne << C->getZExtValue();
860 // assume(~(v >> c) = a)
861 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
862 m_LShr(m_V, m_ConstantInt(C)),
863 m_AShr(m_V, m_ConstantInt(C)))),
865 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
866 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
867 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
868 // For those bits in RHS that are known, we can propagate them inverted
869 // to known bits in V shifted to the right by C.
870 KnownZero |= RHSKnownOne << C->getZExtValue();
871 KnownOne |= RHSKnownZero << C->getZExtValue();
872 // assume(v >=_s c) where c is non-negative
873 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
874 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
875 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
876 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
878 if (RHSKnownZero.isNegative()) {
879 // We know that the sign bit is zero.
880 KnownZero |= APInt::getSignBit(BitWidth);
882 // assume(v >_s c) where c is at least -1.
883 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
884 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
885 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
886 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
888 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
889 // We know that the sign bit is zero.
890 KnownZero |= APInt::getSignBit(BitWidth);
892 // assume(v <=_s c) where c is negative
893 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
894 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
895 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
896 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
898 if (RHSKnownOne.isNegative()) {
899 // We know that the sign bit is one.
900 KnownOne |= APInt::getSignBit(BitWidth);
902 // assume(v <_s c) where c is non-positive
903 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
904 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
905 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
906 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
908 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
909 // We know that the sign bit is one.
910 KnownOne |= APInt::getSignBit(BitWidth);
913 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
914 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
915 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
916 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
918 // Whatever high bits in c are zero are known to be zero.
920 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
922 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
923 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
924 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
925 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
927 // Whatever high bits in c are zero are known to be zero (if c is a power
928 // of 2, then one more).
929 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
931 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
934 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
939 /// Determine which bits of V are known to be either zero or one and return
940 /// them in the KnownZero/KnownOne bit sets.
942 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
943 /// we cannot optimize based on the assumption that it is zero without changing
944 /// it to be an explicit zero. If we don't change it to zero, other code could
945 /// optimized based on the contradictory assumption that it is non-zero.
946 /// Because instcombine aggressively folds operations with undef args anyway,
947 /// this won't lose us code quality.
949 /// This function is defined on values with integer type, values with pointer
950 /// type, and vectors of integers. In the case
951 /// where V is a vector, known zero, and known one values are the
952 /// same width as the vector element, and the bit is set only if it is true
953 /// for all of the elements in the vector.
954 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
955 const DataLayout &DL, unsigned Depth, const Query &Q) {
956 assert(V && "No Value?");
957 assert(Depth <= MaxDepth && "Limit Search Depth");
958 unsigned BitWidth = KnownZero.getBitWidth();
960 assert((V->getType()->isIntOrIntVectorTy() ||
961 V->getType()->getScalarType()->isPointerTy()) &&
962 "Not integer or pointer type!");
963 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
964 (!V->getType()->isIntOrIntVectorTy() ||
965 V->getType()->getScalarSizeInBits() == BitWidth) &&
966 KnownZero.getBitWidth() == BitWidth &&
967 KnownOne.getBitWidth() == BitWidth &&
968 "V, KnownOne and KnownZero should have same BitWidth");
970 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
971 // We know all of the bits for a constant!
972 KnownOne = CI->getValue();
973 KnownZero = ~KnownOne;
976 // Null and aggregate-zero are all-zeros.
977 if (isa<ConstantPointerNull>(V) ||
978 isa<ConstantAggregateZero>(V)) {
979 KnownOne.clearAllBits();
980 KnownZero = APInt::getAllOnesValue(BitWidth);
983 // Handle a constant vector by taking the intersection of the known bits of
984 // each element. There is no real need to handle ConstantVector here, because
985 // we don't handle undef in any particularly useful way.
986 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
987 // We know that CDS must be a vector of integers. Take the intersection of
989 KnownZero.setAllBits(); KnownOne.setAllBits();
990 APInt Elt(KnownZero.getBitWidth(), 0);
991 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
992 Elt = CDS->getElementAsInteger(i);
999 // The address of an aligned GlobalValue has trailing zeros.
1000 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1001 unsigned Align = GO->getAlignment();
1003 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1004 Type *ObjectType = GVar->getType()->getElementType();
1005 if (ObjectType->isSized()) {
1006 // If the object is defined in the current Module, we'll be giving
1007 // it the preferred alignment. Otherwise, we have to assume that it
1008 // may only have the minimum ABI alignment.
1009 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
1010 Align = DL.getPreferredAlignment(GVar);
1012 Align = DL.getABITypeAlignment(ObjectType);
1017 KnownZero = APInt::getLowBitsSet(BitWidth,
1018 countTrailingZeros(Align));
1020 KnownZero.clearAllBits();
1021 KnownOne.clearAllBits();
1025 if (Argument *A = dyn_cast<Argument>(V)) {
1026 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1028 if (!Align && A->hasStructRetAttr()) {
1029 // An sret parameter has at least the ABI alignment of the return type.
1030 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1031 if (EltTy->isSized())
1032 Align = DL.getABITypeAlignment(EltTy);
1036 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1038 KnownZero.clearAllBits();
1039 KnownOne.clearAllBits();
1041 // Don't give up yet... there might be an assumption that provides more
1043 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1045 // Or a dominating condition for that matter
1046 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1047 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL,
1052 // Start out not knowing anything.
1053 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1055 // Limit search depth.
1056 // All recursive calls that increase depth must come after this.
1057 if (Depth == MaxDepth)
1060 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1061 // the bits of its aliasee.
1062 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1063 if (!GA->mayBeOverridden())
1064 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1068 // Check whether a nearby assume intrinsic can determine some known bits.
1069 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1071 // Check whether there's a dominating condition which implies something about
1072 // this value at the given context.
1073 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1074 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1077 Operator *I = dyn_cast<Operator>(V);
1080 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
1081 switch (I->getOpcode()) {
1083 case Instruction::Load:
1084 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
1085 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1087 case Instruction::And: {
1088 // If either the LHS or the RHS are Zero, the result is zero.
1089 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1090 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1092 // Output known-1 bits are only known if set in both the LHS & RHS.
1093 KnownOne &= KnownOne2;
1094 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1095 KnownZero |= KnownZero2;
1098 case Instruction::Or: {
1099 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1100 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1102 // Output known-0 bits are only known if clear in both the LHS & RHS.
1103 KnownZero &= KnownZero2;
1104 // Output known-1 are known to be set if set in either the LHS | RHS.
1105 KnownOne |= KnownOne2;
1108 case Instruction::Xor: {
1109 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1110 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1112 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1113 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1114 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1115 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1116 KnownZero = KnownZeroOut;
1119 case Instruction::Mul: {
1120 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1121 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1122 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1125 case Instruction::UDiv: {
1126 // For the purposes of computing leading zeros we can conservatively
1127 // treat a udiv as a logical right shift by the power of 2 known to
1128 // be less than the denominator.
1129 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1130 unsigned LeadZ = KnownZero2.countLeadingOnes();
1132 KnownOne2.clearAllBits();
1133 KnownZero2.clearAllBits();
1134 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1135 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1136 if (RHSUnknownLeadingOnes != BitWidth)
1137 LeadZ = std::min(BitWidth,
1138 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1140 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1143 case Instruction::Select:
1144 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1145 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1147 // Only known if known in both the LHS and RHS.
1148 KnownOne &= KnownOne2;
1149 KnownZero &= KnownZero2;
1151 case Instruction::FPTrunc:
1152 case Instruction::FPExt:
1153 case Instruction::FPToUI:
1154 case Instruction::FPToSI:
1155 case Instruction::SIToFP:
1156 case Instruction::UIToFP:
1157 break; // Can't work with floating point.
1158 case Instruction::PtrToInt:
1159 case Instruction::IntToPtr:
1160 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1161 // FALL THROUGH and handle them the same as zext/trunc.
1162 case Instruction::ZExt:
1163 case Instruction::Trunc: {
1164 Type *SrcTy = I->getOperand(0)->getType();
1166 unsigned SrcBitWidth;
1167 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1168 // which fall through here.
1169 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1171 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1172 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1173 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1174 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1175 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1176 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1177 // Any top bits are known to be zero.
1178 if (BitWidth > SrcBitWidth)
1179 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1182 case Instruction::BitCast: {
1183 Type *SrcTy = I->getOperand(0)->getType();
1184 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
1185 // TODO: For now, not handling conversions like:
1186 // (bitcast i64 %x to <2 x i32>)
1187 !I->getType()->isVectorTy()) {
1188 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1193 case Instruction::SExt: {
1194 // Compute the bits in the result that are not present in the input.
1195 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1197 KnownZero = KnownZero.trunc(SrcBitWidth);
1198 KnownOne = KnownOne.trunc(SrcBitWidth);
1199 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1200 KnownZero = KnownZero.zext(BitWidth);
1201 KnownOne = KnownOne.zext(BitWidth);
1203 // If the sign bit of the input is known set or clear, then we know the
1204 // top bits of the result.
1205 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1206 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1207 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1208 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1211 case Instruction::Shl:
1212 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1213 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1214 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1215 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1216 KnownZero <<= ShiftAmt;
1217 KnownOne <<= ShiftAmt;
1218 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1221 case Instruction::LShr:
1222 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1223 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1224 // Compute the new bits that are at the top now.
1225 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1227 // Unsigned shift right.
1228 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1229 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1230 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1231 // high bits known zero.
1232 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1235 case Instruction::AShr:
1236 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1237 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1238 // Compute the new bits that are at the top now.
1239 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1241 // Signed shift right.
1242 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1243 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1244 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1246 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1247 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1248 KnownZero |= HighBits;
1249 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1250 KnownOne |= HighBits;
1253 case Instruction::Sub: {
1254 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1255 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1256 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1260 case Instruction::Add: {
1261 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1262 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1263 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1267 case Instruction::SRem:
1268 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1269 APInt RA = Rem->getValue().abs();
1270 if (RA.isPowerOf2()) {
1271 APInt LowBits = RA - 1;
1272 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1275 // The low bits of the first operand are unchanged by the srem.
1276 KnownZero = KnownZero2 & LowBits;
1277 KnownOne = KnownOne2 & LowBits;
1279 // If the first operand is non-negative or has all low bits zero, then
1280 // the upper bits are all zero.
1281 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1282 KnownZero |= ~LowBits;
1284 // If the first operand is negative and not all low bits are zero, then
1285 // the upper bits are all one.
1286 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1287 KnownOne |= ~LowBits;
1289 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1293 // The sign bit is the LHS's sign bit, except when the result of the
1294 // remainder is zero.
1295 if (KnownZero.isNonNegative()) {
1296 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1297 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1299 // If it's known zero, our sign bit is also zero.
1300 if (LHSKnownZero.isNegative())
1301 KnownZero.setBit(BitWidth - 1);
1305 case Instruction::URem: {
1306 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1307 APInt RA = Rem->getValue();
1308 if (RA.isPowerOf2()) {
1309 APInt LowBits = (RA - 1);
1310 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1312 KnownZero |= ~LowBits;
1313 KnownOne &= LowBits;
1318 // Since the result is less than or equal to either operand, any leading
1319 // zero bits in either operand must also exist in the result.
1320 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1321 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1323 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1324 KnownZero2.countLeadingOnes());
1325 KnownOne.clearAllBits();
1326 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1330 case Instruction::Alloca: {
1331 AllocaInst *AI = cast<AllocaInst>(V);
1332 unsigned Align = AI->getAlignment();
1334 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1337 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1340 case Instruction::GetElementPtr: {
1341 // Analyze all of the subscripts of this getelementptr instruction
1342 // to determine if we can prove known low zero bits.
1343 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1344 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1346 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1348 gep_type_iterator GTI = gep_type_begin(I);
1349 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1350 Value *Index = I->getOperand(i);
1351 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1352 // Handle struct member offset arithmetic.
1354 // Handle case when index is vector zeroinitializer
1355 Constant *CIndex = cast<Constant>(Index);
1356 if (CIndex->isZeroValue())
1359 if (CIndex->getType()->isVectorTy())
1360 Index = CIndex->getSplatValue();
1362 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1363 const StructLayout *SL = DL.getStructLayout(STy);
1364 uint64_t Offset = SL->getElementOffset(Idx);
1365 TrailZ = std::min<unsigned>(TrailZ,
1366 countTrailingZeros(Offset));
1368 // Handle array index arithmetic.
1369 Type *IndexedTy = GTI.getIndexedType();
1370 if (!IndexedTy->isSized()) {
1374 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1375 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1376 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1377 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1379 TrailZ = std::min(TrailZ,
1380 unsigned(countTrailingZeros(TypeSize) +
1381 LocalKnownZero.countTrailingOnes()));
1385 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1388 case Instruction::PHI: {
1389 PHINode *P = cast<PHINode>(I);
1390 // Handle the case of a simple two-predecessor recurrence PHI.
1391 // There's a lot more that could theoretically be done here, but
1392 // this is sufficient to catch some interesting cases.
1393 if (P->getNumIncomingValues() == 2) {
1394 for (unsigned i = 0; i != 2; ++i) {
1395 Value *L = P->getIncomingValue(i);
1396 Value *R = P->getIncomingValue(!i);
1397 Operator *LU = dyn_cast<Operator>(L);
1400 unsigned Opcode = LU->getOpcode();
1401 // Check for operations that have the property that if
1402 // both their operands have low zero bits, the result
1403 // will have low zero bits.
1404 if (Opcode == Instruction::Add ||
1405 Opcode == Instruction::Sub ||
1406 Opcode == Instruction::And ||
1407 Opcode == Instruction::Or ||
1408 Opcode == Instruction::Mul) {
1409 Value *LL = LU->getOperand(0);
1410 Value *LR = LU->getOperand(1);
1411 // Find a recurrence.
1418 // Ok, we have a PHI of the form L op= R. Check for low
1420 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1422 // We need to take the minimum number of known bits
1423 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1424 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1426 KnownZero = APInt::getLowBitsSet(BitWidth,
1427 std::min(KnownZero2.countTrailingOnes(),
1428 KnownZero3.countTrailingOnes()));
1434 // Unreachable blocks may have zero-operand PHI nodes.
1435 if (P->getNumIncomingValues() == 0)
1438 // Otherwise take the unions of the known bit sets of the operands,
1439 // taking conservative care to avoid excessive recursion.
1440 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1441 // Skip if every incoming value references to ourself.
1442 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1445 KnownZero = APInt::getAllOnesValue(BitWidth);
1446 KnownOne = APInt::getAllOnesValue(BitWidth);
1447 for (Value *IncValue : P->incoming_values()) {
1448 // Skip direct self references.
1449 if (IncValue == P) continue;
1451 KnownZero2 = APInt(BitWidth, 0);
1452 KnownOne2 = APInt(BitWidth, 0);
1453 // Recurse, but cap the recursion to one level, because we don't
1454 // want to waste time spinning around in loops.
1455 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1457 KnownZero &= KnownZero2;
1458 KnownOne &= KnownOne2;
1459 // If all bits have been ruled out, there's no need to check
1461 if (!KnownZero && !KnownOne)
1467 case Instruction::Call:
1468 case Instruction::Invoke:
1469 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1470 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1471 // If a range metadata is attached to this IntrinsicInst, intersect the
1472 // explicit range specified by the metadata and the implicit range of
1474 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1475 switch (II->getIntrinsicID()) {
1477 case Intrinsic::ctlz:
1478 case Intrinsic::cttz: {
1479 unsigned LowBits = Log2_32(BitWidth)+1;
1480 // If this call is undefined for 0, the result will be less than 2^n.
1481 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1483 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1486 case Intrinsic::ctpop: {
1487 unsigned LowBits = Log2_32(BitWidth)+1;
1488 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1491 case Intrinsic::x86_sse42_crc32_64_64:
1492 KnownZero |= APInt::getHighBitsSet(64, 32);
1497 case Instruction::ExtractValue:
1498 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1499 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1500 if (EVI->getNumIndices() != 1) break;
1501 if (EVI->getIndices()[0] == 0) {
1502 switch (II->getIntrinsicID()) {
1504 case Intrinsic::uadd_with_overflow:
1505 case Intrinsic::sadd_with_overflow:
1506 computeKnownBitsAddSub(true, II->getArgOperand(0),
1507 II->getArgOperand(1), false, KnownZero,
1508 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1510 case Intrinsic::usub_with_overflow:
1511 case Intrinsic::ssub_with_overflow:
1512 computeKnownBitsAddSub(false, II->getArgOperand(0),
1513 II->getArgOperand(1), false, KnownZero,
1514 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1516 case Intrinsic::umul_with_overflow:
1517 case Intrinsic::smul_with_overflow:
1518 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1519 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1527 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1530 /// Determine whether the sign bit is known to be zero or one.
1531 /// Convenience wrapper around computeKnownBits.
1532 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1533 const DataLayout &DL, unsigned Depth, const Query &Q) {
1534 unsigned BitWidth = getBitWidth(V->getType(), DL);
1540 APInt ZeroBits(BitWidth, 0);
1541 APInt OneBits(BitWidth, 0);
1542 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1543 KnownOne = OneBits[BitWidth - 1];
1544 KnownZero = ZeroBits[BitWidth - 1];
1547 /// Return true if the given value is known to have exactly one
1548 /// bit set when defined. For vectors return true if every element is known to
1549 /// be a power of two when defined. Supports values with integer or pointer
1550 /// types and vectors of integers.
1551 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1552 const Query &Q, const DataLayout &DL) {
1553 if (Constant *C = dyn_cast<Constant>(V)) {
1554 if (C->isNullValue())
1556 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1557 return CI->getValue().isPowerOf2();
1558 // TODO: Handle vector constants.
1561 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1562 // it is shifted off the end then the result is undefined.
1563 if (match(V, m_Shl(m_One(), m_Value())))
1566 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1567 // bottom. If it is shifted off the bottom then the result is undefined.
1568 if (match(V, m_LShr(m_SignBit(), m_Value())))
1571 // The remaining tests are all recursive, so bail out if we hit the limit.
1572 if (Depth++ == MaxDepth)
1575 Value *X = nullptr, *Y = nullptr;
1576 // A shift of a power of two is a power of two or zero.
1577 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1578 match(V, m_Shr(m_Value(X), m_Value()))))
1579 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1581 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1582 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1584 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1585 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1586 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1588 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1589 // A power of two and'd with anything is a power of two or zero.
1590 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1591 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1593 // X & (-X) is always a power of two or zero.
1594 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1599 // Adding a power-of-two or zero to the same power-of-two or zero yields
1600 // either the original power-of-two, a larger power-of-two or zero.
1601 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1602 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1603 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1604 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1605 match(X, m_And(m_Value(), m_Specific(Y))))
1606 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1608 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1609 match(Y, m_And(m_Value(), m_Specific(X))))
1610 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1613 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1614 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1615 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1617 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1618 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1619 // If i8 V is a power of two or zero:
1620 // ZeroBits: 1 1 1 0 1 1 1 1
1621 // ~ZeroBits: 0 0 0 1 0 0 0 0
1622 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1623 // If OrZero isn't set, we cannot give back a zero result.
1624 // Make sure either the LHS or RHS has a bit set.
1625 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1630 // An exact divide or right shift can only shift off zero bits, so the result
1631 // is a power of two only if the first operand is a power of two and not
1632 // copying a sign bit (sdiv int_min, 2).
1633 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1634 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1635 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1642 /// \brief Test whether a GEP's result is known to be non-null.
1644 /// Uses properties inherent in a GEP to try to determine whether it is known
1647 /// Currently this routine does not support vector GEPs.
1648 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1649 unsigned Depth, const Query &Q) {
1650 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1653 // FIXME: Support vector-GEPs.
1654 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1656 // If the base pointer is non-null, we cannot walk to a null address with an
1657 // inbounds GEP in address space zero.
1658 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1661 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1662 // If so, then the GEP cannot produce a null pointer, as doing so would
1663 // inherently violate the inbounds contract within address space zero.
1664 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1665 GTI != GTE; ++GTI) {
1666 // Struct types are easy -- they must always be indexed by a constant.
1667 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1668 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1669 unsigned ElementIdx = OpC->getZExtValue();
1670 const StructLayout *SL = DL.getStructLayout(STy);
1671 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1672 if (ElementOffset > 0)
1677 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1678 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1681 // Fast path the constant operand case both for efficiency and so we don't
1682 // increment Depth when just zipping down an all-constant GEP.
1683 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1689 // We post-increment Depth here because while isKnownNonZero increments it
1690 // as well, when we pop back up that increment won't persist. We don't want
1691 // to recurse 10k times just because we have 10k GEP operands. We don't
1692 // bail completely out because we want to handle constant GEPs regardless
1694 if (Depth++ >= MaxDepth)
1697 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1704 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1705 /// ensure that the value it's attached to is never Value? 'RangeType' is
1706 /// is the type of the value described by the range.
1707 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1708 const APInt& Value) {
1709 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1710 assert(NumRanges >= 1);
1711 for (unsigned i = 0; i < NumRanges; ++i) {
1712 ConstantInt *Lower =
1713 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1714 ConstantInt *Upper =
1715 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1716 ConstantRange Range(Lower->getValue(), Upper->getValue());
1717 if (Range.contains(Value))
1723 /// Return true if the given value is known to be non-zero when defined.
1724 /// For vectors return true if every element is known to be non-zero when
1725 /// defined. Supports values with integer or pointer type and vectors of
1727 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1729 if (Constant *C = dyn_cast<Constant>(V)) {
1730 if (C->isNullValue())
1732 if (isa<ConstantInt>(C))
1733 // Must be non-zero due to null test above.
1735 // TODO: Handle vectors
1739 if (Instruction* I = dyn_cast<Instruction>(V)) {
1740 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1741 // If the possible ranges don't contain zero, then the value is
1742 // definitely non-zero.
1743 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1744 const APInt ZeroValue(Ty->getBitWidth(), 0);
1745 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1751 // The remaining tests are all recursive, so bail out if we hit the limit.
1752 if (Depth++ >= MaxDepth)
1755 // Check for pointer simplifications.
1756 if (V->getType()->isPointerTy()) {
1757 if (isKnownNonNull(V))
1759 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1760 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1764 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1766 // X | Y != 0 if X != 0 or Y != 0.
1767 Value *X = nullptr, *Y = nullptr;
1768 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1769 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1771 // ext X != 0 if X != 0.
1772 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1773 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1775 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1776 // if the lowest bit is shifted off the end.
1777 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1778 // shl nuw can't remove any non-zero bits.
1779 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1780 if (BO->hasNoUnsignedWrap())
1781 return isKnownNonZero(X, DL, Depth, Q);
1783 APInt KnownZero(BitWidth, 0);
1784 APInt KnownOne(BitWidth, 0);
1785 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1789 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1790 // defined if the sign bit is shifted off the end.
1791 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1792 // shr exact can only shift out zero bits.
1793 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1795 return isKnownNonZero(X, DL, Depth, Q);
1797 bool XKnownNonNegative, XKnownNegative;
1798 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1802 // div exact can only produce a zero if the dividend is zero.
1803 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1804 return isKnownNonZero(X, DL, Depth, Q);
1807 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1808 bool XKnownNonNegative, XKnownNegative;
1809 bool YKnownNonNegative, YKnownNegative;
1810 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1811 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1813 // If X and Y are both non-negative (as signed values) then their sum is not
1814 // zero unless both X and Y are zero.
1815 if (XKnownNonNegative && YKnownNonNegative)
1816 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1819 // If X and Y are both negative (as signed values) then their sum is not
1820 // zero unless both X and Y equal INT_MIN.
1821 if (BitWidth && XKnownNegative && YKnownNegative) {
1822 APInt KnownZero(BitWidth, 0);
1823 APInt KnownOne(BitWidth, 0);
1824 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1825 // The sign bit of X is set. If some other bit is set then X is not equal
1827 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1828 if ((KnownOne & Mask) != 0)
1830 // The sign bit of Y is set. If some other bit is set then Y is not equal
1832 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1833 if ((KnownOne & Mask) != 0)
1837 // The sum of a non-negative number and a power of two is not zero.
1838 if (XKnownNonNegative &&
1839 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
1841 if (YKnownNonNegative &&
1842 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
1846 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1847 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1848 // If X and Y are non-zero then so is X * Y as long as the multiplication
1849 // does not overflow.
1850 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1851 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
1854 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1855 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1856 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
1857 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
1861 if (!BitWidth) return false;
1862 APInt KnownZero(BitWidth, 0);
1863 APInt KnownOne(BitWidth, 0);
1864 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1865 return KnownOne != 0;
1868 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1869 /// simplify operations downstream. Mask is known to be zero for bits that V
1872 /// This function is defined on values with integer type, values with pointer
1873 /// type, and vectors of integers. In the case
1874 /// where V is a vector, the mask, known zero, and known one values are the
1875 /// same width as the vector element, and the bit is set only if it is true
1876 /// for all of the elements in the vector.
1877 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
1878 unsigned Depth, const Query &Q) {
1879 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1880 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
1881 return (KnownZero & Mask) == Mask;
1886 /// Return the number of times the sign bit of the register is replicated into
1887 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1888 /// (itself), but other cases can give us information. For example, immediately
1889 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1890 /// other, so we return 3.
1892 /// 'Op' must have a scalar integer type.
1894 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
1896 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
1898 unsigned FirstAnswer = 1;
1900 // Note that ConstantInt is handled by the general computeKnownBits case
1904 return 1; // Limit search depth.
1906 Operator *U = dyn_cast<Operator>(V);
1907 switch (Operator::getOpcode(V)) {
1909 case Instruction::SExt:
1910 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1911 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
1913 case Instruction::SDiv: {
1914 const APInt *Denominator;
1915 // sdiv X, C -> adds log(C) sign bits.
1916 if (match(U->getOperand(1), m_APInt(Denominator))) {
1918 // Ignore non-positive denominator.
1919 if (!Denominator->isStrictlyPositive())
1922 // Calculate the incoming numerator bits.
1923 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1925 // Add floor(log(C)) bits to the numerator bits.
1926 return std::min(TyBits, NumBits + Denominator->logBase2());
1931 case Instruction::SRem: {
1932 const APInt *Denominator;
1933 // srem X, C -> we know that the result is within [-C+1,C) when C is a
1934 // positive constant. This let us put a lower bound on the number of sign
1936 if (match(U->getOperand(1), m_APInt(Denominator))) {
1938 // Ignore non-positive denominator.
1939 if (!Denominator->isStrictlyPositive())
1942 // Calculate the incoming numerator bits. SRem by a positive constant
1943 // can't lower the number of sign bits.
1945 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1947 // Calculate the leading sign bit constraints by examining the
1948 // denominator. Given that the denominator is positive, there are two
1951 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
1952 // (1 << ceilLogBase2(C)).
1954 // 2. the numerator is negative. Then the result range is (-C,0] and
1955 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
1957 // Thus a lower bound on the number of sign bits is `TyBits -
1958 // ceilLogBase2(C)`.
1960 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
1961 return std::max(NumrBits, ResBits);
1966 case Instruction::AShr: {
1967 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1968 // ashr X, C -> adds C sign bits. Vectors too.
1970 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1971 Tmp += ShAmt->getZExtValue();
1972 if (Tmp > TyBits) Tmp = TyBits;
1976 case Instruction::Shl: {
1978 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1979 // shl destroys sign bits.
1980 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1981 Tmp2 = ShAmt->getZExtValue();
1982 if (Tmp2 >= TyBits || // Bad shift.
1983 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1988 case Instruction::And:
1989 case Instruction::Or:
1990 case Instruction::Xor: // NOT is handled here.
1991 // Logical binary ops preserve the number of sign bits at the worst.
1992 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
1994 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
1995 FirstAnswer = std::min(Tmp, Tmp2);
1996 // We computed what we know about the sign bits as our first
1997 // answer. Now proceed to the generic code that uses
1998 // computeKnownBits, and pick whichever answer is better.
2002 case Instruction::Select:
2003 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2004 if (Tmp == 1) return 1; // Early out.
2005 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2006 return std::min(Tmp, Tmp2);
2008 case Instruction::Add:
2009 // Add can have at most one carry bit. Thus we know that the output
2010 // is, at worst, one more bit than the inputs.
2011 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2012 if (Tmp == 1) return 1; // Early out.
2014 // Special case decrementing a value (ADD X, -1):
2015 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2016 if (CRHS->isAllOnesValue()) {
2017 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2018 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2021 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2023 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2026 // If we are subtracting one from a positive number, there is no carry
2027 // out of the result.
2028 if (KnownZero.isNegative())
2032 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2033 if (Tmp2 == 1) return 1;
2034 return std::min(Tmp, Tmp2)-1;
2036 case Instruction::Sub:
2037 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2038 if (Tmp2 == 1) return 1;
2041 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2042 if (CLHS->isNullValue()) {
2043 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2044 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2046 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2048 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2051 // If the input is known to be positive (the sign bit is known clear),
2052 // the output of the NEG has the same number of sign bits as the input.
2053 if (KnownZero.isNegative())
2056 // Otherwise, we treat this like a SUB.
2059 // Sub can have at most one carry bit. Thus we know that the output
2060 // is, at worst, one more bit than the inputs.
2061 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2062 if (Tmp == 1) return 1; // Early out.
2063 return std::min(Tmp, Tmp2)-1;
2065 case Instruction::PHI: {
2066 PHINode *PN = cast<PHINode>(U);
2067 unsigned NumIncomingValues = PN->getNumIncomingValues();
2068 // Don't analyze large in-degree PHIs.
2069 if (NumIncomingValues > 4) break;
2070 // Unreachable blocks may have zero-operand PHI nodes.
2071 if (NumIncomingValues == 0) break;
2073 // Take the minimum of all incoming values. This can't infinitely loop
2074 // because of our depth threshold.
2075 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2076 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2077 if (Tmp == 1) return Tmp;
2079 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2084 case Instruction::Trunc:
2085 // FIXME: it's tricky to do anything useful for this, but it is an important
2086 // case for targets like X86.
2090 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2091 // use this information.
2092 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2094 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2096 if (KnownZero.isNegative()) { // sign bit is 0
2098 } else if (KnownOne.isNegative()) { // sign bit is 1;
2105 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2106 // the number of identical bits in the top of the input value.
2108 Mask <<= Mask.getBitWidth()-TyBits;
2109 // Return # leading zeros. We use 'min' here in case Val was zero before
2110 // shifting. We don't want to return '64' as for an i32 "0".
2111 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2114 /// This function computes the integer multiple of Base that equals V.
2115 /// If successful, it returns true and returns the multiple in
2116 /// Multiple. If unsuccessful, it returns false. It looks
2117 /// through SExt instructions only if LookThroughSExt is true.
2118 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2119 bool LookThroughSExt, unsigned Depth) {
2120 const unsigned MaxDepth = 6;
2122 assert(V && "No Value?");
2123 assert(Depth <= MaxDepth && "Limit Search Depth");
2124 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2126 Type *T = V->getType();
2128 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2138 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2139 Constant *BaseVal = ConstantInt::get(T, Base);
2140 if (CO && CO == BaseVal) {
2142 Multiple = ConstantInt::get(T, 1);
2146 if (CI && CI->getZExtValue() % Base == 0) {
2147 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2151 if (Depth == MaxDepth) return false; // Limit search depth.
2153 Operator *I = dyn_cast<Operator>(V);
2154 if (!I) return false;
2156 switch (I->getOpcode()) {
2158 case Instruction::SExt:
2159 if (!LookThroughSExt) return false;
2160 // otherwise fall through to ZExt
2161 case Instruction::ZExt:
2162 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2163 LookThroughSExt, Depth+1);
2164 case Instruction::Shl:
2165 case Instruction::Mul: {
2166 Value *Op0 = I->getOperand(0);
2167 Value *Op1 = I->getOperand(1);
2169 if (I->getOpcode() == Instruction::Shl) {
2170 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2171 if (!Op1CI) return false;
2172 // Turn Op0 << Op1 into Op0 * 2^Op1
2173 APInt Op1Int = Op1CI->getValue();
2174 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2175 APInt API(Op1Int.getBitWidth(), 0);
2176 API.setBit(BitToSet);
2177 Op1 = ConstantInt::get(V->getContext(), API);
2180 Value *Mul0 = nullptr;
2181 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2182 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2183 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2184 if (Op1C->getType()->getPrimitiveSizeInBits() <
2185 MulC->getType()->getPrimitiveSizeInBits())
2186 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2187 if (Op1C->getType()->getPrimitiveSizeInBits() >
2188 MulC->getType()->getPrimitiveSizeInBits())
2189 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2191 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2192 Multiple = ConstantExpr::getMul(MulC, Op1C);
2196 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2197 if (Mul0CI->getValue() == 1) {
2198 // V == Base * Op1, so return Op1
2204 Value *Mul1 = nullptr;
2205 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2206 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2207 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2208 if (Op0C->getType()->getPrimitiveSizeInBits() <
2209 MulC->getType()->getPrimitiveSizeInBits())
2210 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2211 if (Op0C->getType()->getPrimitiveSizeInBits() >
2212 MulC->getType()->getPrimitiveSizeInBits())
2213 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2215 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2216 Multiple = ConstantExpr::getMul(MulC, Op0C);
2220 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2221 if (Mul1CI->getValue() == 1) {
2222 // V == Base * Op0, so return Op0
2230 // We could not determine if V is a multiple of Base.
2234 /// Return true if we can prove that the specified FP value is never equal to
2237 /// NOTE: this function will need to be revisited when we support non-default
2240 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2241 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2242 return !CFP->getValueAPF().isNegZero();
2244 // FIXME: Magic number! At the least, this should be given a name because it's
2245 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2246 // expose it as a parameter, so it can be used for testing / experimenting.
2248 return false; // Limit search depth.
2250 const Operator *I = dyn_cast<Operator>(V);
2251 if (!I) return false;
2253 // Check if the nsz fast-math flag is set
2254 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2255 if (FPO->hasNoSignedZeros())
2258 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2259 if (I->getOpcode() == Instruction::FAdd)
2260 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2261 if (CFP->isNullValue())
2264 // sitofp and uitofp turn into +0.0 for zero.
2265 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2268 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2269 // sqrt(-0.0) = -0.0, no other negative results are possible.
2270 if (II->getIntrinsicID() == Intrinsic::sqrt)
2271 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2273 if (const CallInst *CI = dyn_cast<CallInst>(I))
2274 if (const Function *F = CI->getCalledFunction()) {
2275 if (F->isDeclaration()) {
2277 if (F->getName() == "abs") return true;
2278 // fabs[lf](x) != -0.0
2279 if (F->getName() == "fabs") return true;
2280 if (F->getName() == "fabsf") return true;
2281 if (F->getName() == "fabsl") return true;
2282 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2283 F->getName() == "sqrtl")
2284 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2291 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2292 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2293 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2295 // FIXME: Magic number! At the least, this should be given a name because it's
2296 // used similarly in CannotBeNegativeZero(). A better fix may be to
2297 // expose it as a parameter, so it can be used for testing / experimenting.
2299 return false; // Limit search depth.
2301 const Operator *I = dyn_cast<Operator>(V);
2302 if (!I) return false;
2304 switch (I->getOpcode()) {
2306 case Instruction::FMul:
2307 // x*x is always non-negative or a NaN.
2308 if (I->getOperand(0) == I->getOperand(1))
2311 case Instruction::FAdd:
2312 case Instruction::FDiv:
2313 case Instruction::FRem:
2314 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2315 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2316 case Instruction::FPExt:
2317 case Instruction::FPTrunc:
2318 // Widening/narrowing never change sign.
2319 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2320 case Instruction::Call:
2321 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2322 switch (II->getIntrinsicID()) {
2324 case Intrinsic::exp:
2325 case Intrinsic::exp2:
2326 case Intrinsic::fabs:
2327 case Intrinsic::sqrt:
2329 case Intrinsic::powi:
2330 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2331 // powi(x,n) is non-negative if n is even.
2332 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2335 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2336 case Intrinsic::fma:
2337 case Intrinsic::fmuladd:
2338 // x*x+y is non-negative if y is non-negative.
2339 return I->getOperand(0) == I->getOperand(1) &&
2340 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2347 /// If the specified value can be set by repeating the same byte in memory,
2348 /// return the i8 value that it is represented with. This is
2349 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2350 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2351 /// byte store (e.g. i16 0x1234), return null.
2352 Value *llvm::isBytewiseValue(Value *V) {
2353 // All byte-wide stores are splatable, even of arbitrary variables.
2354 if (V->getType()->isIntegerTy(8)) return V;
2356 // Handle 'null' ConstantArrayZero etc.
2357 if (Constant *C = dyn_cast<Constant>(V))
2358 if (C->isNullValue())
2359 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2361 // Constant float and double values can be handled as integer values if the
2362 // corresponding integer value is "byteable". An important case is 0.0.
2363 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2364 if (CFP->getType()->isFloatTy())
2365 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2366 if (CFP->getType()->isDoubleTy())
2367 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2368 // Don't handle long double formats, which have strange constraints.
2371 // We can handle constant integers that are multiple of 8 bits.
2372 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2373 if (CI->getBitWidth() % 8 == 0) {
2374 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2376 if (!CI->getValue().isSplat(8))
2378 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2382 // A ConstantDataArray/Vector is splatable if all its members are equal and
2384 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2385 Value *Elt = CA->getElementAsConstant(0);
2386 Value *Val = isBytewiseValue(Elt);
2390 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2391 if (CA->getElementAsConstant(I) != Elt)
2397 // Conceptually, we could handle things like:
2398 // %a = zext i8 %X to i16
2399 // %b = shl i16 %a, 8
2400 // %c = or i16 %a, %b
2401 // but until there is an example that actually needs this, it doesn't seem
2402 // worth worrying about.
2407 // This is the recursive version of BuildSubAggregate. It takes a few different
2408 // arguments. Idxs is the index within the nested struct From that we are
2409 // looking at now (which is of type IndexedType). IdxSkip is the number of
2410 // indices from Idxs that should be left out when inserting into the resulting
2411 // struct. To is the result struct built so far, new insertvalue instructions
2413 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2414 SmallVectorImpl<unsigned> &Idxs,
2416 Instruction *InsertBefore) {
2417 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2419 // Save the original To argument so we can modify it
2421 // General case, the type indexed by Idxs is a struct
2422 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2423 // Process each struct element recursively
2426 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2430 // Couldn't find any inserted value for this index? Cleanup
2431 while (PrevTo != OrigTo) {
2432 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2433 PrevTo = Del->getAggregateOperand();
2434 Del->eraseFromParent();
2436 // Stop processing elements
2440 // If we successfully found a value for each of our subaggregates
2444 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2445 // the struct's elements had a value that was inserted directly. In the latter
2446 // case, perhaps we can't determine each of the subelements individually, but
2447 // we might be able to find the complete struct somewhere.
2449 // Find the value that is at that particular spot
2450 Value *V = FindInsertedValue(From, Idxs);
2455 // Insert the value in the new (sub) aggregrate
2456 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2457 "tmp", InsertBefore);
2460 // This helper takes a nested struct and extracts a part of it (which is again a
2461 // struct) into a new value. For example, given the struct:
2462 // { a, { b, { c, d }, e } }
2463 // and the indices "1, 1" this returns
2466 // It does this by inserting an insertvalue for each element in the resulting
2467 // struct, as opposed to just inserting a single struct. This will only work if
2468 // each of the elements of the substruct are known (ie, inserted into From by an
2469 // insertvalue instruction somewhere).
2471 // All inserted insertvalue instructions are inserted before InsertBefore
2472 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2473 Instruction *InsertBefore) {
2474 assert(InsertBefore && "Must have someplace to insert!");
2475 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2477 Value *To = UndefValue::get(IndexedType);
2478 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2479 unsigned IdxSkip = Idxs.size();
2481 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2484 /// Given an aggregrate and an sequence of indices, see if
2485 /// the scalar value indexed is already around as a register, for example if it
2486 /// were inserted directly into the aggregrate.
2488 /// If InsertBefore is not null, this function will duplicate (modified)
2489 /// insertvalues when a part of a nested struct is extracted.
2490 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2491 Instruction *InsertBefore) {
2492 // Nothing to index? Just return V then (this is useful at the end of our
2494 if (idx_range.empty())
2496 // We have indices, so V should have an indexable type.
2497 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2498 "Not looking at a struct or array?");
2499 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2500 "Invalid indices for type?");
2502 if (Constant *C = dyn_cast<Constant>(V)) {
2503 C = C->getAggregateElement(idx_range[0]);
2504 if (!C) return nullptr;
2505 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2508 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2509 // Loop the indices for the insertvalue instruction in parallel with the
2510 // requested indices
2511 const unsigned *req_idx = idx_range.begin();
2512 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2513 i != e; ++i, ++req_idx) {
2514 if (req_idx == idx_range.end()) {
2515 // We can't handle this without inserting insertvalues
2519 // The requested index identifies a part of a nested aggregate. Handle
2520 // this specially. For example,
2521 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2522 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2523 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2524 // This can be changed into
2525 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2526 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2527 // which allows the unused 0,0 element from the nested struct to be
2529 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2533 // This insert value inserts something else than what we are looking for.
2534 // See if the (aggregrate) value inserted into has the value we are
2535 // looking for, then.
2537 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2540 // If we end up here, the indices of the insertvalue match with those
2541 // requested (though possibly only partially). Now we recursively look at
2542 // the inserted value, passing any remaining indices.
2543 return FindInsertedValue(I->getInsertedValueOperand(),
2544 makeArrayRef(req_idx, idx_range.end()),
2548 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2549 // If we're extracting a value from an aggregrate that was extracted from
2550 // something else, we can extract from that something else directly instead.
2551 // However, we will need to chain I's indices with the requested indices.
2553 // Calculate the number of indices required
2554 unsigned size = I->getNumIndices() + idx_range.size();
2555 // Allocate some space to put the new indices in
2556 SmallVector<unsigned, 5> Idxs;
2558 // Add indices from the extract value instruction
2559 Idxs.append(I->idx_begin(), I->idx_end());
2561 // Add requested indices
2562 Idxs.append(idx_range.begin(), idx_range.end());
2564 assert(Idxs.size() == size
2565 && "Number of indices added not correct?");
2567 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2569 // Otherwise, we don't know (such as, extracting from a function return value
2570 // or load instruction)
2574 /// Analyze the specified pointer to see if it can be expressed as a base
2575 /// pointer plus a constant offset. Return the base and offset to the caller.
2576 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2577 const DataLayout &DL) {
2578 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2579 APInt ByteOffset(BitWidth, 0);
2581 if (Ptr->getType()->isVectorTy())
2584 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2585 APInt GEPOffset(BitWidth, 0);
2586 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2589 ByteOffset += GEPOffset;
2591 Ptr = GEP->getPointerOperand();
2592 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2593 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2594 Ptr = cast<Operator>(Ptr)->getOperand(0);
2595 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2596 if (GA->mayBeOverridden())
2598 Ptr = GA->getAliasee();
2603 Offset = ByteOffset.getSExtValue();
2608 /// This function computes the length of a null-terminated C string pointed to
2609 /// by V. If successful, it returns true and returns the string in Str.
2610 /// If unsuccessful, it returns false.
2611 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2612 uint64_t Offset, bool TrimAtNul) {
2615 // Look through bitcast instructions and geps.
2616 V = V->stripPointerCasts();
2618 // If the value is a GEP instruction or constant expression, treat it as an
2620 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2621 // Make sure the GEP has exactly three arguments.
2622 if (GEP->getNumOperands() != 3)
2625 // Make sure the index-ee is a pointer to array of i8.
2626 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2627 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2628 if (!AT || !AT->getElementType()->isIntegerTy(8))
2631 // Check to make sure that the first operand of the GEP is an integer and
2632 // has value 0 so that we are sure we're indexing into the initializer.
2633 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2634 if (!FirstIdx || !FirstIdx->isZero())
2637 // If the second index isn't a ConstantInt, then this is a variable index
2638 // into the array. If this occurs, we can't say anything meaningful about
2640 uint64_t StartIdx = 0;
2641 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2642 StartIdx = CI->getZExtValue();
2645 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2649 // The GEP instruction, constant or instruction, must reference a global
2650 // variable that is a constant and is initialized. The referenced constant
2651 // initializer is the array that we'll use for optimization.
2652 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2653 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2656 // Handle the all-zeros case
2657 if (GV->getInitializer()->isNullValue()) {
2658 // This is a degenerate case. The initializer is constant zero so the
2659 // length of the string must be zero.
2664 // Must be a Constant Array
2665 const ConstantDataArray *Array =
2666 dyn_cast<ConstantDataArray>(GV->getInitializer());
2667 if (!Array || !Array->isString())
2670 // Get the number of elements in the array
2671 uint64_t NumElts = Array->getType()->getArrayNumElements();
2673 // Start out with the entire array in the StringRef.
2674 Str = Array->getAsString();
2676 if (Offset > NumElts)
2679 // Skip over 'offset' bytes.
2680 Str = Str.substr(Offset);
2683 // Trim off the \0 and anything after it. If the array is not nul
2684 // terminated, we just return the whole end of string. The client may know
2685 // some other way that the string is length-bound.
2686 Str = Str.substr(0, Str.find('\0'));
2691 // These next two are very similar to the above, but also look through PHI
2693 // TODO: See if we can integrate these two together.
2695 /// If we can compute the length of the string pointed to by
2696 /// the specified pointer, return 'len+1'. If we can't, return 0.
2697 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2698 // Look through noop bitcast instructions.
2699 V = V->stripPointerCasts();
2701 // If this is a PHI node, there are two cases: either we have already seen it
2703 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2704 if (!PHIs.insert(PN).second)
2705 return ~0ULL; // already in the set.
2707 // If it was new, see if all the input strings are the same length.
2708 uint64_t LenSoFar = ~0ULL;
2709 for (Value *IncValue : PN->incoming_values()) {
2710 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2711 if (Len == 0) return 0; // Unknown length -> unknown.
2713 if (Len == ~0ULL) continue;
2715 if (Len != LenSoFar && LenSoFar != ~0ULL)
2716 return 0; // Disagree -> unknown.
2720 // Success, all agree.
2724 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2725 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2726 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2727 if (Len1 == 0) return 0;
2728 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2729 if (Len2 == 0) return 0;
2730 if (Len1 == ~0ULL) return Len2;
2731 if (Len2 == ~0ULL) return Len1;
2732 if (Len1 != Len2) return 0;
2736 // Otherwise, see if we can read the string.
2738 if (!getConstantStringInfo(V, StrData))
2741 return StrData.size()+1;
2744 /// If we can compute the length of the string pointed to by
2745 /// the specified pointer, return 'len+1'. If we can't, return 0.
2746 uint64_t llvm::GetStringLength(Value *V) {
2747 if (!V->getType()->isPointerTy()) return 0;
2749 SmallPtrSet<PHINode*, 32> PHIs;
2750 uint64_t Len = GetStringLengthH(V, PHIs);
2751 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2752 // an empty string as a length.
2753 return Len == ~0ULL ? 1 : Len;
2756 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2757 /// previous iteration of the loop was referring to the same object as \p PN.
2758 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2759 // Find the loop-defined value.
2760 Loop *L = LI->getLoopFor(PN->getParent());
2761 if (PN->getNumIncomingValues() != 2)
2764 // Find the value from previous iteration.
2765 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2766 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2767 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2768 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2771 // If a new pointer is loaded in the loop, the pointer references a different
2772 // object in every iteration. E.g.:
2776 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
2777 if (!L->isLoopInvariant(Load->getPointerOperand()))
2782 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
2783 unsigned MaxLookup) {
2784 if (!V->getType()->isPointerTy())
2786 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2787 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2788 V = GEP->getPointerOperand();
2789 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2790 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2791 V = cast<Operator>(V)->getOperand(0);
2792 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2793 if (GA->mayBeOverridden())
2795 V = GA->getAliasee();
2797 // See if InstructionSimplify knows any relevant tricks.
2798 if (Instruction *I = dyn_cast<Instruction>(V))
2799 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2800 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
2807 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2812 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
2813 const DataLayout &DL, LoopInfo *LI,
2814 unsigned MaxLookup) {
2815 SmallPtrSet<Value *, 4> Visited;
2816 SmallVector<Value *, 4> Worklist;
2817 Worklist.push_back(V);
2819 Value *P = Worklist.pop_back_val();
2820 P = GetUnderlyingObject(P, DL, MaxLookup);
2822 if (!Visited.insert(P).second)
2825 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2826 Worklist.push_back(SI->getTrueValue());
2827 Worklist.push_back(SI->getFalseValue());
2831 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2832 // If this PHI changes the underlying object in every iteration of the
2833 // loop, don't look through it. Consider:
2836 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
2840 // Prev is tracking Curr one iteration behind so they refer to different
2841 // underlying objects.
2842 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
2843 isSameUnderlyingObjectInLoop(PN, LI))
2844 for (Value *IncValue : PN->incoming_values())
2845 Worklist.push_back(IncValue);
2849 Objects.push_back(P);
2850 } while (!Worklist.empty());
2853 /// Return true if the only users of this pointer are lifetime markers.
2854 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2855 for (const User *U : V->users()) {
2856 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2857 if (!II) return false;
2859 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2860 II->getIntrinsicID() != Intrinsic::lifetime_end)
2866 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
2867 Type *Ty, const DataLayout &DL,
2868 const Instruction *CtxI,
2869 const DominatorTree *DT,
2870 const TargetLibraryInfo *TLI) {
2871 assert(Offset.isNonNegative() && "offset can't be negative");
2872 assert(Ty->isSized() && "must be sized");
2874 APInt DerefBytes(Offset.getBitWidth(), 0);
2875 bool CheckForNonNull = false;
2876 if (const Argument *A = dyn_cast<Argument>(BV)) {
2877 DerefBytes = A->getDereferenceableBytes();
2878 if (!DerefBytes.getBoolValue()) {
2879 DerefBytes = A->getDereferenceableOrNullBytes();
2880 CheckForNonNull = true;
2882 } else if (auto CS = ImmutableCallSite(BV)) {
2883 DerefBytes = CS.getDereferenceableBytes(0);
2884 if (!DerefBytes.getBoolValue()) {
2885 DerefBytes = CS.getDereferenceableOrNullBytes(0);
2886 CheckForNonNull = true;
2888 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
2889 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
2890 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2891 DerefBytes = CI->getLimitedValue();
2893 if (!DerefBytes.getBoolValue()) {
2895 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
2896 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
2897 DerefBytes = CI->getLimitedValue();
2899 CheckForNonNull = true;
2903 if (DerefBytes.getBoolValue())
2904 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
2905 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
2911 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
2912 const Instruction *CtxI,
2913 const DominatorTree *DT,
2914 const TargetLibraryInfo *TLI) {
2915 Type *VTy = V->getType();
2916 Type *Ty = VTy->getPointerElementType();
2920 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
2921 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
2924 /// Return true if Value is always a dereferenceable pointer.
2926 /// Test if V is always a pointer to allocated and suitably aligned memory for
2927 /// a simple load or store.
2928 static bool isDereferenceablePointer(const Value *V, const DataLayout &DL,
2929 const Instruction *CtxI,
2930 const DominatorTree *DT,
2931 const TargetLibraryInfo *TLI,
2932 SmallPtrSetImpl<const Value *> &Visited) {
2933 // Note that it is not safe to speculate into a malloc'd region because
2934 // malloc may return null.
2936 // These are obviously ok.
2937 if (isa<AllocaInst>(V)) return true;
2939 // It's not always safe to follow a bitcast, for example:
2940 // bitcast i8* (alloca i8) to i32*
2941 // would result in a 4-byte load from a 1-byte alloca. However,
2942 // if we're casting from a pointer from a type of larger size
2943 // to a type of smaller size (or the same size), and the alignment
2944 // is at least as large as for the resulting pointer type, then
2945 // we can look through the bitcast.
2946 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
2947 Type *STy = BC->getSrcTy()->getPointerElementType(),
2948 *DTy = BC->getDestTy()->getPointerElementType();
2949 if (STy->isSized() && DTy->isSized() &&
2950 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
2951 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
2952 return isDereferenceablePointer(BC->getOperand(0), DL, CtxI,
2956 // Global variables which can't collapse to null are ok.
2957 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2958 return !GV->hasExternalWeakLinkage();
2960 // byval arguments are okay.
2961 if (const Argument *A = dyn_cast<Argument>(V))
2962 if (A->hasByValAttr())
2965 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
2968 // For GEPs, determine if the indexing lands within the allocated object.
2969 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2970 // Conservatively require that the base pointer be fully dereferenceable.
2971 if (!Visited.insert(GEP->getOperand(0)).second)
2973 if (!isDereferenceablePointer(GEP->getOperand(0), DL, CtxI,
2976 // Check the indices.
2977 gep_type_iterator GTI = gep_type_begin(GEP);
2978 for (User::const_op_iterator I = GEP->op_begin()+1,
2979 E = GEP->op_end(); I != E; ++I) {
2982 // Struct indices can't be out of bounds.
2983 if (isa<StructType>(Ty))
2985 ConstantInt *CI = dyn_cast<ConstantInt>(Index);
2988 // Zero is always ok.
2991 // Check to see that it's within the bounds of an array.
2992 ArrayType *ATy = dyn_cast<ArrayType>(Ty);
2995 if (CI->getValue().getActiveBits() > 64)
2997 if (CI->getZExtValue() >= ATy->getNumElements())
3000 // Indices check out; this is dereferenceable.
3004 // For gc.relocate, look through relocations
3005 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3006 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3007 GCRelocateOperands RelocateInst(I);
3008 return isDereferenceablePointer(RelocateInst.getDerivedPtr(), DL, CtxI,
3012 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3013 return isDereferenceablePointer(ASC->getOperand(0), DL, CtxI,
3016 // If we don't know, assume the worst.
3020 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3021 const Instruction *CtxI,
3022 const DominatorTree *DT,
3023 const TargetLibraryInfo *TLI) {
3024 // When dereferenceability information is provided by a dereferenceable
3025 // attribute, we know exactly how many bytes are dereferenceable. If we can
3026 // determine the exact offset to the attributed variable, we can use that
3027 // information here.
3028 Type *VTy = V->getType();
3029 Type *Ty = VTy->getPointerElementType();
3030 if (Ty->isSized()) {
3031 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3032 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3034 if (Offset.isNonNegative())
3035 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL,
3040 SmallPtrSet<const Value *, 32> Visited;
3041 return ::isDereferenceablePointer(V, DL, CtxI, DT, TLI, Visited);
3044 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3045 const Instruction *CtxI,
3046 const DominatorTree *DT,
3047 const TargetLibraryInfo *TLI) {
3048 const Operator *Inst = dyn_cast<Operator>(V);
3052 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3053 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3057 switch (Inst->getOpcode()) {
3060 case Instruction::UDiv:
3061 case Instruction::URem: {
3062 // x / y is undefined if y == 0.
3064 if (match(Inst->getOperand(1), m_APInt(V)))
3068 case Instruction::SDiv:
3069 case Instruction::SRem: {
3070 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3071 const APInt *Numerator, *Denominator;
3072 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3074 // We cannot hoist this division if the denominator is 0.
3075 if (*Denominator == 0)
3077 // It's safe to hoist if the denominator is not 0 or -1.
3078 if (*Denominator != -1)
3080 // At this point we know that the denominator is -1. It is safe to hoist as
3081 // long we know that the numerator is not INT_MIN.
3082 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3083 return !Numerator->isMinSignedValue();
3084 // The numerator *might* be MinSignedValue.
3087 case Instruction::Load: {
3088 const LoadInst *LI = cast<LoadInst>(Inst);
3089 if (!LI->isUnordered() ||
3090 // Speculative load may create a race that did not exist in the source.
3091 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
3093 const DataLayout &DL = LI->getModule()->getDataLayout();
3094 return isDereferenceablePointer(LI->getPointerOperand(), DL, CtxI, DT, TLI);
3096 case Instruction::Call: {
3097 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3098 switch (II->getIntrinsicID()) {
3099 // These synthetic intrinsics have no side-effects and just mark
3100 // information about their operands.
3101 // FIXME: There are other no-op synthetic instructions that potentially
3102 // should be considered at least *safe* to speculate...
3103 case Intrinsic::dbg_declare:
3104 case Intrinsic::dbg_value:
3107 case Intrinsic::bswap:
3108 case Intrinsic::ctlz:
3109 case Intrinsic::ctpop:
3110 case Intrinsic::cttz:
3111 case Intrinsic::objectsize:
3112 case Intrinsic::sadd_with_overflow:
3113 case Intrinsic::smul_with_overflow:
3114 case Intrinsic::ssub_with_overflow:
3115 case Intrinsic::uadd_with_overflow:
3116 case Intrinsic::umul_with_overflow:
3117 case Intrinsic::usub_with_overflow:
3119 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3120 // errno like libm sqrt would.
3121 case Intrinsic::sqrt:
3122 case Intrinsic::fma:
3123 case Intrinsic::fmuladd:
3124 case Intrinsic::fabs:
3125 case Intrinsic::minnum:
3126 case Intrinsic::maxnum:
3128 // TODO: some fp intrinsics are marked as having the same error handling
3129 // as libm. They're safe to speculate when they won't error.
3130 // TODO: are convert_{from,to}_fp16 safe?
3131 // TODO: can we list target-specific intrinsics here?
3135 return false; // The called function could have undefined behavior or
3136 // side-effects, even if marked readnone nounwind.
3138 case Instruction::VAArg:
3139 case Instruction::Alloca:
3140 case Instruction::Invoke:
3141 case Instruction::PHI:
3142 case Instruction::Store:
3143 case Instruction::Ret:
3144 case Instruction::Br:
3145 case Instruction::IndirectBr:
3146 case Instruction::Switch:
3147 case Instruction::Unreachable:
3148 case Instruction::Fence:
3149 case Instruction::LandingPad:
3150 case Instruction::AtomicRMW:
3151 case Instruction::AtomicCmpXchg:
3152 case Instruction::Resume:
3153 return false; // Misc instructions which have effects
3157 /// Return true if we know that the specified value is never null.
3158 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3159 // Alloca never returns null, malloc might.
3160 if (isa<AllocaInst>(V)) return true;
3162 // A byval, inalloca, or nonnull argument is never null.
3163 if (const Argument *A = dyn_cast<Argument>(V))
3164 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3166 // Global values are not null unless extern weak.
3167 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3168 return !GV->hasExternalWeakLinkage();
3170 // A Load tagged w/nonnull metadata is never null.
3171 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3172 return LI->getMetadata(LLVMContext::MD_nonnull);
3174 if (auto CS = ImmutableCallSite(V))
3175 if (CS.isReturnNonNull())
3178 // operator new never returns null.
3179 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3185 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3186 const Instruction *CtxI,
3187 const DominatorTree *DT) {
3188 unsigned NumUsesExplored = 0;
3189 for (auto U : V->users()) {
3190 // Avoid massive lists
3191 if (NumUsesExplored >= DomConditionsMaxUses)
3194 // Consider only compare instructions uniquely controlling a branch
3195 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3199 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3202 for (auto *CmpU : Cmp->users()) {
3203 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3207 assert(BI->isConditional() && "uses a comparison!");
3209 BasicBlock *NonNullSuccessor = nullptr;
3210 CmpInst::Predicate Pred;
3212 if (match(const_cast<ICmpInst*>(Cmp),
3213 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3214 if (Pred == ICmpInst::ICMP_EQ)
3215 NonNullSuccessor = BI->getSuccessor(1);
3216 else if (Pred == ICmpInst::ICMP_NE)
3217 NonNullSuccessor = BI->getSuccessor(0);
3220 if (NonNullSuccessor) {
3221 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3222 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3231 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3232 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3233 if (isKnownNonNull(V, TLI))
3236 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3239 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3240 const DataLayout &DL,
3241 AssumptionCache *AC,
3242 const Instruction *CxtI,
3243 const DominatorTree *DT) {
3244 // Multiplying n * m significant bits yields a result of n + m significant
3245 // bits. If the total number of significant bits does not exceed the
3246 // result bit width (minus 1), there is no overflow.
3247 // This means if we have enough leading zero bits in the operands
3248 // we can guarantee that the result does not overflow.
3249 // Ref: "Hacker's Delight" by Henry Warren
3250 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3251 APInt LHSKnownZero(BitWidth, 0);
3252 APInt LHSKnownOne(BitWidth, 0);
3253 APInt RHSKnownZero(BitWidth, 0);
3254 APInt RHSKnownOne(BitWidth, 0);
3255 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3257 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3259 // Note that underestimating the number of zero bits gives a more
3260 // conservative answer.
3261 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3262 RHSKnownZero.countLeadingOnes();
3263 // First handle the easy case: if we have enough zero bits there's
3264 // definitely no overflow.
3265 if (ZeroBits >= BitWidth)
3266 return OverflowResult::NeverOverflows;
3268 // Get the largest possible values for each operand.
3269 APInt LHSMax = ~LHSKnownZero;
3270 APInt RHSMax = ~RHSKnownZero;
3272 // We know the multiply operation doesn't overflow if the maximum values for
3273 // each operand will not overflow after we multiply them together.
3275 LHSMax.umul_ov(RHSMax, MaxOverflow);
3277 return OverflowResult::NeverOverflows;
3279 // We know it always overflows if multiplying the smallest possible values for
3280 // the operands also results in overflow.
3282 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3284 return OverflowResult::AlwaysOverflows;
3286 return OverflowResult::MayOverflow;
3289 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3290 const DataLayout &DL,
3291 AssumptionCache *AC,
3292 const Instruction *CxtI,
3293 const DominatorTree *DT) {
3294 bool LHSKnownNonNegative, LHSKnownNegative;
3295 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3297 if (LHSKnownNonNegative || LHSKnownNegative) {
3298 bool RHSKnownNonNegative, RHSKnownNegative;
3299 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3302 if (LHSKnownNegative && RHSKnownNegative) {
3303 // The sign bit is set in both cases: this MUST overflow.
3304 // Create a simple add instruction, and insert it into the struct.
3305 return OverflowResult::AlwaysOverflows;
3308 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3309 // The sign bit is clear in both cases: this CANNOT overflow.
3310 // Create a simple add instruction, and insert it into the struct.
3311 return OverflowResult::NeverOverflows;
3315 return OverflowResult::MayOverflow;
3318 static SelectPatternFlavor matchSelectPattern(ICmpInst::Predicate Pred,
3319 Value *CmpLHS, Value *CmpRHS,
3320 Value *TrueVal, Value *FalseVal,
3321 Value *&LHS, Value *&RHS) {
3325 // (icmp X, Y) ? X : Y
3326 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3328 default: return SPF_UNKNOWN; // Equality.
3329 case ICmpInst::ICMP_UGT:
3330 case ICmpInst::ICMP_UGE: return SPF_UMAX;
3331 case ICmpInst::ICMP_SGT:
3332 case ICmpInst::ICMP_SGE: return SPF_SMAX;
3333 case ICmpInst::ICMP_ULT:
3334 case ICmpInst::ICMP_ULE: return SPF_UMIN;
3335 case ICmpInst::ICMP_SLT:
3336 case ICmpInst::ICMP_SLE: return SPF_SMIN;
3340 // (icmp X, Y) ? Y : X
3341 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3343 default: return SPF_UNKNOWN; // Equality.
3344 case ICmpInst::ICMP_UGT:
3345 case ICmpInst::ICMP_UGE: return SPF_UMIN;
3346 case ICmpInst::ICMP_SGT:
3347 case ICmpInst::ICMP_SGE: return SPF_SMIN;
3348 case ICmpInst::ICMP_ULT:
3349 case ICmpInst::ICMP_ULE: return SPF_UMAX;
3350 case ICmpInst::ICMP_SLT:
3351 case ICmpInst::ICMP_SLE: return SPF_SMAX;
3355 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3356 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3357 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3359 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3360 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3361 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3362 return (CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS;
3365 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3366 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3367 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3368 return (CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS;
3372 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3373 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3374 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3375 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3376 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3384 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3389 static Constant *lookThroughCast(ICmpInst *CmpI, Value *V1, Value *V2,
3390 Instruction::CastOps *CastOp) {
3391 CastInst *CI = dyn_cast<CastInst>(V1);
3392 Constant *C = dyn_cast<Constant>(V2);
3395 *CastOp = CI->getOpcode();
3397 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3398 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3399 // This is only valid if the truncated value can be sign-extended
3400 // back to the original value.
3401 if (ConstantExpr::getSExt(T, C->getType()) == C)
3405 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3406 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3408 if (isa<TruncInst>(CI))
3409 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3414 SelectPatternFlavor llvm::matchSelectPattern(Value *V,
3415 Value *&LHS, Value *&RHS,
3416 Instruction::CastOps *CastOp) {
3417 SelectInst *SI = dyn_cast<SelectInst>(V);
3418 if (!SI) return SPF_UNKNOWN;
3420 ICmpInst *CmpI = dyn_cast<ICmpInst>(SI->getCondition());
3421 if (!CmpI) return SPF_UNKNOWN;
3423 ICmpInst::Predicate Pred = CmpI->getPredicate();
3424 Value *CmpLHS = CmpI->getOperand(0);
3425 Value *CmpRHS = CmpI->getOperand(1);
3426 Value *TrueVal = SI->getTrueValue();
3427 Value *FalseVal = SI->getFalseValue();
3430 if (CmpI->isEquality())
3433 // Deal with type mismatches.
3434 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
3435 if (Constant *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
3436 return ::matchSelectPattern(Pred, CmpLHS, CmpRHS,
3437 cast<CastInst>(TrueVal)->getOperand(0), C,
3439 if (Constant *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
3440 return ::matchSelectPattern(Pred, CmpLHS, CmpRHS,
3441 C, cast<CastInst>(FalseVal)->getOperand(0),
3444 return ::matchSelectPattern(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal,