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/Optional.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/MemoryBuiltins.h"
21 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/IR/CallSite.h"
23 #include "llvm/IR/ConstantRange.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/GetElementPtrTypeIterator.h"
28 #include "llvm/IR/GlobalAlias.h"
29 #include "llvm/IR/GlobalVariable.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/Metadata.h"
34 #include "llvm/IR/Operator.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/MathExtras.h"
41 using namespace llvm::PatternMatch;
43 const unsigned MaxDepth = 6;
45 /// Enable an experimental feature to leverage information about dominating
46 /// conditions to compute known bits. The individual options below control how
47 /// hard we search. The defaults are chosen to be fairly aggressive. If you
48 /// run into compile time problems when testing, scale them back and report
50 static cl::opt<bool> EnableDomConditions("value-tracking-dom-conditions",
51 cl::Hidden, cl::init(false));
53 // This is expensive, so we only do it for the top level query value.
54 // (TODO: evaluate cost vs profit, consider higher thresholds)
55 static cl::opt<unsigned> DomConditionsMaxDepth("dom-conditions-max-depth",
56 cl::Hidden, cl::init(1));
58 /// How many dominating blocks should be scanned looking for dominating
60 static cl::opt<unsigned> DomConditionsMaxDomBlocks("dom-conditions-dom-blocks",
64 // Controls the number of uses of the value searched for possible
65 // dominating comparisons.
66 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
67 cl::Hidden, cl::init(20));
69 // If true, don't consider only compares whose only use is a branch.
70 static cl::opt<bool> DomConditionsSingleCmpUse("dom-conditions-single-cmp-use",
71 cl::Hidden, cl::init(false));
73 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
74 /// 0). For vector types, returns the element type's bitwidth.
75 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
76 if (unsigned BitWidth = Ty->getScalarSizeInBits())
79 return DL.getPointerTypeSizeInBits(Ty);
82 // Many of these functions have internal versions that take an assumption
83 // exclusion set. This is because of the potential for mutual recursion to
84 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
85 // classic case of this is assume(x = y), which will attempt to determine
86 // bits in x from bits in y, which will attempt to determine bits in y from
87 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
88 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
89 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
90 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
93 // Simplifying using an assume can only be done in a particular control-flow
94 // context (the context instruction provides that context). If an assume and
95 // the context instruction are not in the same block then the DT helps in
96 // figuring out if we can use it.
100 const Instruction *CxtI;
101 const DominatorTree *DT;
103 Query(AssumptionCache *AC = nullptr, const Instruction *CxtI = nullptr,
104 const DominatorTree *DT = nullptr)
105 : AC(AC), CxtI(CxtI), DT(DT) {}
107 Query(const Query &Q, const Value *NewExcl)
108 : ExclInvs(Q.ExclInvs), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT) {
109 ExclInvs.insert(NewExcl);
112 } // end anonymous namespace
114 // Given the provided Value and, potentially, a context instruction, return
115 // the preferred context instruction (if any).
116 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
117 // If we've been provided with a context instruction, then use that (provided
118 // it has been inserted).
119 if (CxtI && CxtI->getParent())
122 // If the value is really an already-inserted instruction, then use that.
123 CxtI = dyn_cast<Instruction>(V);
124 if (CxtI && CxtI->getParent())
130 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
131 const DataLayout &DL, unsigned Depth,
134 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
135 const DataLayout &DL, unsigned Depth,
136 AssumptionCache *AC, const Instruction *CxtI,
137 const DominatorTree *DT) {
138 ::computeKnownBits(V, KnownZero, KnownOne, DL, Depth,
139 Query(AC, safeCxtI(V, CxtI), DT));
142 bool llvm::haveNoCommonBitsSet(Value *LHS, Value *RHS, const DataLayout &DL,
143 AssumptionCache *AC, const Instruction *CxtI,
144 const DominatorTree *DT) {
145 assert(LHS->getType() == RHS->getType() &&
146 "LHS and RHS should have the same type");
147 assert(LHS->getType()->isIntOrIntVectorTy() &&
148 "LHS and RHS should be integers");
149 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
150 APInt LHSKnownZero(IT->getBitWidth(), 0), LHSKnownOne(IT->getBitWidth(), 0);
151 APInt RHSKnownZero(IT->getBitWidth(), 0), RHSKnownOne(IT->getBitWidth(), 0);
152 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, 0, AC, CxtI, DT);
153 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, 0, AC, CxtI, DT);
154 return (LHSKnownZero | RHSKnownZero).isAllOnesValue();
157 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
158 const DataLayout &DL, unsigned Depth,
161 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
162 const DataLayout &DL, unsigned Depth,
163 AssumptionCache *AC, const Instruction *CxtI,
164 const DominatorTree *DT) {
165 ::ComputeSignBit(V, KnownZero, KnownOne, DL, Depth,
166 Query(AC, safeCxtI(V, CxtI), DT));
169 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
170 const Query &Q, const DataLayout &DL);
172 bool llvm::isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, bool OrZero,
173 unsigned Depth, AssumptionCache *AC,
174 const Instruction *CxtI,
175 const DominatorTree *DT) {
176 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
177 Query(AC, safeCxtI(V, CxtI), DT), DL);
180 static bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
183 bool llvm::isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
184 AssumptionCache *AC, const Instruction *CxtI,
185 const DominatorTree *DT) {
186 return ::isKnownNonZero(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
189 bool llvm::isKnownNonNegative(Value *V, const DataLayout &DL, unsigned Depth,
190 AssumptionCache *AC, const Instruction *CxtI,
191 const DominatorTree *DT) {
192 bool NonNegative, Negative;
193 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
197 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
200 bool llvm::isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
201 AssumptionCache *AC, const Instruction *CxtI,
202 const DominatorTree *DT) {
203 return ::isKnownNonEqual(V1, V2, DL, Query(AC,
204 safeCxtI(V1, safeCxtI(V2, CxtI)),
208 static bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
209 unsigned Depth, const Query &Q);
211 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
212 unsigned Depth, AssumptionCache *AC,
213 const Instruction *CxtI, const DominatorTree *DT) {
214 return ::MaskedValueIsZero(V, Mask, DL, Depth,
215 Query(AC, safeCxtI(V, CxtI), DT));
218 static unsigned ComputeNumSignBits(Value *V, const DataLayout &DL,
219 unsigned Depth, const Query &Q);
221 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL,
222 unsigned Depth, AssumptionCache *AC,
223 const Instruction *CxtI,
224 const DominatorTree *DT) {
225 return ::ComputeNumSignBits(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT));
228 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
229 APInt &KnownZero, APInt &KnownOne,
230 APInt &KnownZero2, APInt &KnownOne2,
231 const DataLayout &DL, unsigned Depth,
234 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
235 // We know that the top bits of C-X are clear if X contains less bits
236 // than C (i.e. no wrap-around can happen). For example, 20-X is
237 // positive if we can prove that X is >= 0 and < 16.
238 if (!CLHS->getValue().isNegative()) {
239 unsigned BitWidth = KnownZero.getBitWidth();
240 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
241 // NLZ can't be BitWidth with no sign bit
242 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
243 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
245 // If all of the MaskV bits are known to be zero, then we know the
246 // output top bits are zero, because we now know that the output is
248 if ((KnownZero2 & MaskV) == MaskV) {
249 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
250 // Top bits known zero.
251 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
257 unsigned BitWidth = KnownZero.getBitWidth();
259 // If an initial sequence of bits in the result is not needed, the
260 // corresponding bits in the operands are not needed.
261 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
262 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, DL, Depth + 1, Q);
263 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q);
265 // Carry in a 1 for a subtract, rather than a 0.
266 APInt CarryIn(BitWidth, 0);
268 // Sum = LHS + ~RHS + 1
269 std::swap(KnownZero2, KnownOne2);
273 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
274 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
276 // Compute known bits of the carry.
277 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
278 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
280 // Compute set of known bits (where all three relevant bits are known).
281 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
282 APInt RHSKnown = KnownZero2 | KnownOne2;
283 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
284 APInt Known = LHSKnown & RHSKnown & CarryKnown;
286 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
287 "known bits of sum differ");
289 // Compute known bits of the result.
290 KnownZero = ~PossibleSumOne & Known;
291 KnownOne = PossibleSumOne & Known;
293 // Are we still trying to solve for the sign bit?
294 if (!Known.isNegative()) {
296 // Adding two non-negative numbers, or subtracting a negative number from
297 // a non-negative one, can't wrap into negative.
298 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
299 KnownZero |= APInt::getSignBit(BitWidth);
300 // Adding two negative numbers, or subtracting a non-negative number from
301 // a negative one, can't wrap into non-negative.
302 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
303 KnownOne |= APInt::getSignBit(BitWidth);
308 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
309 APInt &KnownZero, APInt &KnownOne,
310 APInt &KnownZero2, APInt &KnownOne2,
311 const DataLayout &DL, unsigned Depth,
313 unsigned BitWidth = KnownZero.getBitWidth();
314 computeKnownBits(Op1, KnownZero, KnownOne, DL, Depth + 1, Q);
315 computeKnownBits(Op0, KnownZero2, KnownOne2, DL, Depth + 1, Q);
317 bool isKnownNegative = false;
318 bool isKnownNonNegative = false;
319 // If the multiplication is known not to overflow, compute the sign bit.
322 // The product of a number with itself is non-negative.
323 isKnownNonNegative = true;
325 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
326 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
327 bool isKnownNegativeOp1 = KnownOne.isNegative();
328 bool isKnownNegativeOp0 = KnownOne2.isNegative();
329 // The product of two numbers with the same sign is non-negative.
330 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
331 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
332 // The product of a negative number and a non-negative number is either
334 if (!isKnownNonNegative)
335 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
336 isKnownNonZero(Op0, DL, Depth, Q)) ||
337 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
338 isKnownNonZero(Op1, DL, Depth, Q));
342 // If low bits are zero in either operand, output low known-0 bits.
343 // Also compute a conservative estimate for high known-0 bits.
344 // More trickiness is possible, but this is sufficient for the
345 // interesting case of alignment computation.
346 KnownOne.clearAllBits();
347 unsigned TrailZ = KnownZero.countTrailingOnes() +
348 KnownZero2.countTrailingOnes();
349 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
350 KnownZero2.countLeadingOnes(),
351 BitWidth) - BitWidth;
353 TrailZ = std::min(TrailZ, BitWidth);
354 LeadZ = std::min(LeadZ, BitWidth);
355 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
356 APInt::getHighBitsSet(BitWidth, LeadZ);
358 // Only make use of no-wrap flags if we failed to compute the sign bit
359 // directly. This matters if the multiplication always overflows, in
360 // which case we prefer to follow the result of the direct computation,
361 // though as the program is invoking undefined behaviour we can choose
362 // whatever we like here.
363 if (isKnownNonNegative && !KnownOne.isNegative())
364 KnownZero.setBit(BitWidth - 1);
365 else if (isKnownNegative && !KnownZero.isNegative())
366 KnownOne.setBit(BitWidth - 1);
369 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
371 unsigned BitWidth = KnownZero.getBitWidth();
372 unsigned NumRanges = Ranges.getNumOperands() / 2;
373 assert(NumRanges >= 1);
375 // Use the high end of the ranges to find leading zeros.
376 unsigned MinLeadingZeros = BitWidth;
377 for (unsigned i = 0; i < NumRanges; ++i) {
379 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
381 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
382 ConstantRange Range(Lower->getValue(), Upper->getValue());
383 unsigned LeadingZeros = Range.getUnsignedMax().countLeadingZeros();
384 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
387 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
390 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
391 SmallVector<const Value *, 16> WorkSet(1, I);
392 SmallPtrSet<const Value *, 32> Visited;
393 SmallPtrSet<const Value *, 16> EphValues;
395 // The instruction defining an assumption's condition itself is always
396 // considered ephemeral to that assumption (even if it has other
397 // non-ephemeral users). See r246696's test case for an example.
398 if (std::find(I->op_begin(), I->op_end(), E) != I->op_end())
401 while (!WorkSet.empty()) {
402 const Value *V = WorkSet.pop_back_val();
403 if (!Visited.insert(V).second)
406 // If all uses of this value are ephemeral, then so is this value.
407 if (std::all_of(V->user_begin(), V->user_end(),
408 [&](const User *U) { return EphValues.count(U); })) {
413 if (const User *U = dyn_cast<User>(V))
414 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
416 if (isSafeToSpeculativelyExecute(*J))
417 WorkSet.push_back(*J);
425 // Is this an intrinsic that cannot be speculated but also cannot trap?
426 static bool isAssumeLikeIntrinsic(const Instruction *I) {
427 if (const CallInst *CI = dyn_cast<CallInst>(I))
428 if (Function *F = CI->getCalledFunction())
429 switch (F->getIntrinsicID()) {
431 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
432 case Intrinsic::assume:
433 case Intrinsic::dbg_declare:
434 case Intrinsic::dbg_value:
435 case Intrinsic::invariant_start:
436 case Intrinsic::invariant_end:
437 case Intrinsic::lifetime_start:
438 case Intrinsic::lifetime_end:
439 case Intrinsic::objectsize:
440 case Intrinsic::ptr_annotation:
441 case Intrinsic::var_annotation:
448 static bool isValidAssumeForContext(Value *V, const Query &Q) {
449 Instruction *Inv = cast<Instruction>(V);
451 // There are two restrictions on the use of an assume:
452 // 1. The assume must dominate the context (or the control flow must
453 // reach the assume whenever it reaches the context).
454 // 2. The context must not be in the assume's set of ephemeral values
455 // (otherwise we will use the assume to prove that the condition
456 // feeding the assume is trivially true, thus causing the removal of
460 if (Q.DT->dominates(Inv, Q.CxtI)) {
462 } else if (Inv->getParent() == Q.CxtI->getParent()) {
463 // The context comes first, but they're both in the same block. Make sure
464 // there is nothing in between that might interrupt the control flow.
465 for (BasicBlock::const_iterator I =
466 std::next(BasicBlock::const_iterator(Q.CxtI)),
467 IE(Inv); I != IE; ++I)
468 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
471 return !isEphemeralValueOf(Inv, Q.CxtI);
477 // When we don't have a DT, we do a limited search...
478 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
480 } else if (Inv->getParent() == Q.CxtI->getParent()) {
481 // Search forward from the assume until we reach the context (or the end
482 // of the block); the common case is that the assume will come first.
483 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
484 IE = Inv->getParent()->end(); I != IE; ++I)
488 // The context must come first...
489 for (BasicBlock::const_iterator I =
490 std::next(BasicBlock::const_iterator(Q.CxtI)),
491 IE(Inv); I != IE; ++I)
492 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
495 return !isEphemeralValueOf(Inv, Q.CxtI);
501 bool llvm::isValidAssumeForContext(const Instruction *I,
502 const Instruction *CxtI,
503 const DominatorTree *DT) {
504 return ::isValidAssumeForContext(const_cast<Instruction *>(I),
505 Query(nullptr, CxtI, DT));
508 template<typename LHS, typename RHS>
509 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
510 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
511 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
512 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
515 template<typename LHS, typename RHS>
516 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
517 BinaryOp_match<RHS, LHS, Instruction::And>>
518 m_c_And(const LHS &L, const RHS &R) {
519 return m_CombineOr(m_And(L, R), m_And(R, L));
522 template<typename LHS, typename RHS>
523 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
524 BinaryOp_match<RHS, LHS, Instruction::Or>>
525 m_c_Or(const LHS &L, const RHS &R) {
526 return m_CombineOr(m_Or(L, R), m_Or(R, L));
529 template<typename LHS, typename RHS>
530 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
531 BinaryOp_match<RHS, LHS, Instruction::Xor>>
532 m_c_Xor(const LHS &L, const RHS &R) {
533 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
536 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
537 /// true (at the context instruction.) This is mostly a utility function for
538 /// the prototype dominating conditions reasoning below.
539 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
542 const DataLayout &DL,
543 unsigned Depth, const Query &Q) {
544 Value *LHS = Cmp->getOperand(0);
545 Value *RHS = Cmp->getOperand(1);
546 // TODO: We could potentially be more aggressive here. This would be worth
547 // evaluating. If we can, explore commoning this code with the assume
549 if (LHS != V && RHS != V)
552 const unsigned BitWidth = KnownZero.getBitWidth();
554 switch (Cmp->getPredicate()) {
556 // We know nothing from this condition
558 // TODO: implement unsigned bound from below (known one bits)
559 // TODO: common condition check implementations with assumes
560 // TODO: implement other patterns from assume (e.g. V & B == A)
561 case ICmpInst::ICMP_SGT:
563 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
564 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
565 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
566 // We know that the sign bit is zero.
567 KnownZero |= APInt::getSignBit(BitWidth);
571 case ICmpInst::ICMP_EQ:
573 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
575 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
577 computeKnownBits(LHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
579 llvm_unreachable("missing use?");
580 KnownZero |= KnownZeroTemp;
581 KnownOne |= KnownOneTemp;
584 case ICmpInst::ICMP_ULE:
586 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
587 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
588 // The known zero bits carry over
589 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
590 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
593 case ICmpInst::ICMP_ULT:
595 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
596 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
597 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
598 // power of 2, then one more).
599 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
600 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
602 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
608 /// Compute known bits in 'V' from conditions which are known to be true along
609 /// all paths leading to the context instruction. In particular, look for
610 /// cases where one branch of an interesting condition dominates the context
611 /// instruction. This does not do general dataflow.
612 /// NOTE: This code is EXPERIMENTAL and currently off by default.
613 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
615 const DataLayout &DL,
618 // Need both the dominator tree and the query location to do anything useful
619 if (!Q.DT || !Q.CxtI)
621 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
622 // The context instruction might be in a statically unreachable block. If
623 // so, asking dominator queries may yield suprising results. (e.g. the block
624 // may not have a dom tree node)
625 if (!Q.DT->isReachableFromEntry(Cxt->getParent()))
628 // Avoid useless work
629 if (auto VI = dyn_cast<Instruction>(V))
630 if (VI->getParent() == Cxt->getParent())
633 // Note: We currently implement two options. It's not clear which of these
634 // will survive long term, we need data for that.
635 // Option 1 - Try walking the dominator tree looking for conditions which
636 // might apply. This works well for local conditions (loop guards, etc..),
637 // but not as well for things far from the context instruction (presuming a
638 // low max blocks explored). If we can set an high enough limit, this would
640 // Option 2 - We restrict out search to those conditions which are uses of
641 // the value we're interested in. This is independent of dom structure,
642 // but is slightly less powerful without looking through lots of use chains.
643 // It does handle conditions far from the context instruction (e.g. early
644 // function exits on entry) really well though.
646 // Option 1 - Search the dom tree
647 unsigned NumBlocksExplored = 0;
648 BasicBlock *Current = Cxt->getParent();
650 // Stop searching if we've gone too far up the chain
651 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
655 if (!Q.DT->getNode(Current)->getIDom())
657 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
659 // found function entry
662 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
663 if (!BI || BI->isUnconditional())
665 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
669 // We're looking for conditions that are guaranteed to hold at the context
670 // instruction. Finding a condition where one path dominates the context
671 // isn't enough because both the true and false cases could merge before
672 // the context instruction we're actually interested in. Instead, we need
673 // to ensure that the taken *edge* dominates the context instruction. We
674 // know that the edge must be reachable since we started from a reachable
676 BasicBlock *BB0 = BI->getSuccessor(0);
677 BasicBlockEdge Edge(BI->getParent(), BB0);
678 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
681 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
685 // Option 2 - Search the other uses of V
686 unsigned NumUsesExplored = 0;
687 for (auto U : V->users()) {
688 // Avoid massive lists
689 if (NumUsesExplored >= DomConditionsMaxUses)
692 // Consider only compare instructions uniquely controlling a branch
693 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
697 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
700 for (auto *CmpU : Cmp->users()) {
701 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
702 if (!BI || BI->isUnconditional())
704 // We're looking for conditions that are guaranteed to hold at the
705 // context instruction. Finding a condition where one path dominates
706 // the context isn't enough because both the true and false cases could
707 // merge before the context instruction we're actually interested in.
708 // Instead, we need to ensure that the taken *edge* dominates the context
710 BasicBlock *BB0 = BI->getSuccessor(0);
711 BasicBlockEdge Edge(BI->getParent(), BB0);
712 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
715 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
721 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
722 APInt &KnownOne, const DataLayout &DL,
723 unsigned Depth, const Query &Q) {
724 // Use of assumptions is context-sensitive. If we don't have a context, we
726 if (!Q.AC || !Q.CxtI)
729 unsigned BitWidth = KnownZero.getBitWidth();
731 for (auto &AssumeVH : Q.AC->assumptions()) {
734 CallInst *I = cast<CallInst>(AssumeVH);
735 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
736 "Got assumption for the wrong function!");
737 if (Q.ExclInvs.count(I))
740 // Warning: This loop can end up being somewhat performance sensetive.
741 // We're running this loop for once for each value queried resulting in a
742 // runtime of ~O(#assumes * #values).
744 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
745 "must be an assume intrinsic");
747 Value *Arg = I->getArgOperand(0);
749 if (Arg == V && isValidAssumeForContext(I, Q)) {
750 assert(BitWidth == 1 && "assume operand is not i1?");
751 KnownZero.clearAllBits();
752 KnownOne.setAllBits();
756 // The remaining tests are all recursive, so bail out if we hit the limit.
757 if (Depth == MaxDepth)
761 auto m_V = m_CombineOr(m_Specific(V),
762 m_CombineOr(m_PtrToInt(m_Specific(V)),
763 m_BitCast(m_Specific(V))));
765 CmpInst::Predicate Pred;
768 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
769 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
770 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
771 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
772 KnownZero |= RHSKnownZero;
773 KnownOne |= RHSKnownOne;
775 } else if (match(Arg,
776 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
777 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
778 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
779 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
780 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
781 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
783 // For those bits in the mask that are known to be one, we can propagate
784 // known bits from the RHS to V.
785 KnownZero |= RHSKnownZero & MaskKnownOne;
786 KnownOne |= RHSKnownOne & MaskKnownOne;
787 // assume(~(v & b) = a)
788 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
790 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
791 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
792 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
793 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
794 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
796 // For those bits in the mask that are known to be one, we can propagate
797 // inverted known bits from the RHS to V.
798 KnownZero |= RHSKnownOne & MaskKnownOne;
799 KnownOne |= RHSKnownZero & MaskKnownOne;
801 } else if (match(Arg,
802 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
803 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
804 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
805 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
806 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
807 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
809 // For those bits in B that are known to be zero, we can propagate known
810 // bits from the RHS to V.
811 KnownZero |= RHSKnownZero & BKnownZero;
812 KnownOne |= RHSKnownOne & BKnownZero;
813 // assume(~(v | b) = a)
814 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
816 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
817 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
818 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
819 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
820 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
822 // For those bits in B that are known to be zero, we can propagate
823 // inverted known bits from the RHS to V.
824 KnownZero |= RHSKnownOne & BKnownZero;
825 KnownOne |= RHSKnownZero & BKnownZero;
827 } else if (match(Arg,
828 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
829 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
830 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
831 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
832 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
833 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
835 // For those bits in B that are known to be zero, we can propagate known
836 // bits from the RHS to V. For those bits in B that are known to be one,
837 // we can propagate inverted known bits from the RHS to V.
838 KnownZero |= RHSKnownZero & BKnownZero;
839 KnownOne |= RHSKnownOne & BKnownZero;
840 KnownZero |= RHSKnownOne & BKnownOne;
841 KnownOne |= RHSKnownZero & BKnownOne;
842 // assume(~(v ^ b) = a)
843 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
845 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
846 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
847 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
848 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
849 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
851 // For those bits in B that are known to be zero, we can propagate
852 // inverted known bits from the RHS to V. For those bits in B that are
853 // known to be one, we can propagate known bits from the RHS to V.
854 KnownZero |= RHSKnownOne & BKnownZero;
855 KnownOne |= RHSKnownZero & BKnownZero;
856 KnownZero |= RHSKnownZero & BKnownOne;
857 KnownOne |= RHSKnownOne & BKnownOne;
858 // assume(v << c = a)
859 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
861 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
862 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
863 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
864 // For those bits in RHS that are known, we can propagate them to known
865 // bits in V shifted to the right by C.
866 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
867 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
868 // assume(~(v << c) = a)
869 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
871 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
872 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
873 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
874 // For those bits in RHS that are known, we can propagate them inverted
875 // to known bits in V shifted to the right by C.
876 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
877 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
878 // assume(v >> c = a)
879 } else if (match(Arg,
880 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
881 m_AShr(m_V, m_ConstantInt(C))),
883 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
884 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
885 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
886 // For those bits in RHS that are known, we can propagate them to known
887 // bits in V shifted to the right by C.
888 KnownZero |= RHSKnownZero << C->getZExtValue();
889 KnownOne |= RHSKnownOne << C->getZExtValue();
890 // assume(~(v >> c) = a)
891 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
892 m_LShr(m_V, m_ConstantInt(C)),
893 m_AShr(m_V, m_ConstantInt(C)))),
895 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
896 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
897 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
898 // For those bits in RHS that are known, we can propagate them inverted
899 // to known bits in V shifted to the right by C.
900 KnownZero |= RHSKnownOne << C->getZExtValue();
901 KnownOne |= RHSKnownZero << C->getZExtValue();
902 // assume(v >=_s c) where c is non-negative
903 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
904 Pred == ICmpInst::ICMP_SGE && 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.isNegative()) {
909 // We know that the sign bit is zero.
910 KnownZero |= APInt::getSignBit(BitWidth);
912 // assume(v >_s c) where c is at least -1.
913 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
914 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
915 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
916 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
918 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
919 // We know that the sign bit is zero.
920 KnownZero |= APInt::getSignBit(BitWidth);
922 // assume(v <=_s c) where c is negative
923 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
924 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
925 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
926 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
928 if (RHSKnownOne.isNegative()) {
929 // We know that the sign bit is one.
930 KnownOne |= APInt::getSignBit(BitWidth);
932 // assume(v <_s c) where c is non-positive
933 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
934 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
935 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
936 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
938 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
939 // We know that the sign bit is one.
940 KnownOne |= APInt::getSignBit(BitWidth);
943 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
944 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
945 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
946 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
948 // Whatever high bits in c are zero are known to be zero.
950 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
952 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
953 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
954 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
955 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
957 // Whatever high bits in c are zero are known to be zero (if c is a power
958 // of 2, then one more).
959 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
961 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
964 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
969 // Compute known bits from a shift operator, including those with a
970 // non-constant shift amount. KnownZero and KnownOne are the outputs of this
971 // function. KnownZero2 and KnownOne2 are pre-allocated temporaries with the
972 // same bit width as KnownZero and KnownOne. KZF and KOF are operator-specific
973 // functors that, given the known-zero or known-one bits respectively, and a
974 // shift amount, compute the implied known-zero or known-one bits of the shift
975 // operator's result respectively for that shift amount. The results from calling
976 // KZF and KOF are conservatively combined for all permitted shift amounts.
977 template <typename KZFunctor, typename KOFunctor>
978 static void computeKnownBitsFromShiftOperator(Operator *I,
979 APInt &KnownZero, APInt &KnownOne,
980 APInt &KnownZero2, APInt &KnownOne2,
981 const DataLayout &DL, unsigned Depth, const Query &Q,
982 KZFunctor KZF, KOFunctor KOF) {
983 unsigned BitWidth = KnownZero.getBitWidth();
985 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
986 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
988 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
989 KnownZero = KZF(KnownZero, ShiftAmt);
990 KnownOne = KOF(KnownOne, ShiftAmt);
994 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
996 // Note: We cannot use KnownZero.getLimitedValue() here, because if
997 // BitWidth > 64 and any upper bits are known, we'll end up returning the
998 // limit value (which implies all bits are known).
999 uint64_t ShiftAmtKZ = KnownZero.zextOrTrunc(64).getZExtValue();
1000 uint64_t ShiftAmtKO = KnownOne.zextOrTrunc(64).getZExtValue();
1002 // It would be more-clearly correct to use the two temporaries for this
1003 // calculation. Reusing the APInts here to prevent unnecessary allocations.
1004 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1006 // If we know the shifter operand is nonzero, we can sometimes infer more
1007 // known bits. However this is expensive to compute, so be lazy about it and
1008 // only compute it when absolutely necessary.
1009 Optional<bool> ShifterOperandIsNonZero;
1011 // Early exit if we can't constrain any well-defined shift amount.
1012 if (!(ShiftAmtKZ & (BitWidth - 1)) && !(ShiftAmtKO & (BitWidth - 1))) {
1013 ShifterOperandIsNonZero =
1014 isKnownNonZero(I->getOperand(1), DL, Depth + 1, Q);
1015 if (!*ShifterOperandIsNonZero)
1019 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1021 KnownZero = KnownOne = APInt::getAllOnesValue(BitWidth);
1022 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
1023 // Combine the shifted known input bits only for those shift amounts
1024 // compatible with its known constraints.
1025 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
1027 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
1029 // If we know the shifter is nonzero, we may be able to infer more known
1030 // bits. This check is sunk down as far as possible to avoid the expensive
1031 // call to isKnownNonZero if the cheaper checks above fail.
1032 if (ShiftAmt == 0) {
1033 if (!ShifterOperandIsNonZero.hasValue())
1034 ShifterOperandIsNonZero =
1035 isKnownNonZero(I->getOperand(1), DL, Depth + 1, Q);
1036 if (*ShifterOperandIsNonZero)
1040 KnownZero &= KZF(KnownZero2, ShiftAmt);
1041 KnownOne &= KOF(KnownOne2, ShiftAmt);
1044 // If there are no compatible shift amounts, then we've proven that the shift
1045 // amount must be >= the BitWidth, and the result is undefined. We could
1046 // return anything we'd like, but we need to make sure the sets of known bits
1047 // stay disjoint (it should be better for some other code to actually
1048 // propagate the undef than to pick a value here using known bits).
1049 if ((KnownZero & KnownOne) != 0)
1050 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1053 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
1054 APInt &KnownOne, const DataLayout &DL,
1055 unsigned Depth, const Query &Q) {
1056 unsigned BitWidth = KnownZero.getBitWidth();
1058 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
1059 switch (I->getOpcode()) {
1061 case Instruction::Load:
1062 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
1063 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1065 case Instruction::And: {
1066 // If either the LHS or the RHS are Zero, the result is zero.
1067 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1068 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1070 // Output known-1 bits are only known if set in both the LHS & RHS.
1071 KnownOne &= KnownOne2;
1072 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1073 KnownZero |= KnownZero2;
1076 case Instruction::Or: {
1077 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1078 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1080 // Output known-0 bits are only known if clear in both the LHS & RHS.
1081 KnownZero &= KnownZero2;
1082 // Output known-1 are known to be set if set in either the LHS | RHS.
1083 KnownOne |= KnownOne2;
1086 case Instruction::Xor: {
1087 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1088 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1090 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1091 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1092 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1093 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1094 KnownZero = KnownZeroOut;
1097 case Instruction::Mul: {
1098 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1099 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1100 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1103 case Instruction::UDiv: {
1104 // For the purposes of computing leading zeros we can conservatively
1105 // treat a udiv as a logical right shift by the power of 2 known to
1106 // be less than the denominator.
1107 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1108 unsigned LeadZ = KnownZero2.countLeadingOnes();
1110 KnownOne2.clearAllBits();
1111 KnownZero2.clearAllBits();
1112 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1113 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1114 if (RHSUnknownLeadingOnes != BitWidth)
1115 LeadZ = std::min(BitWidth,
1116 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1118 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1121 case Instruction::Select:
1122 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1123 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1125 // Only known if known in both the LHS and RHS.
1126 KnownOne &= KnownOne2;
1127 KnownZero &= KnownZero2;
1129 case Instruction::FPTrunc:
1130 case Instruction::FPExt:
1131 case Instruction::FPToUI:
1132 case Instruction::FPToSI:
1133 case Instruction::SIToFP:
1134 case Instruction::UIToFP:
1135 break; // Can't work with floating point.
1136 case Instruction::PtrToInt:
1137 case Instruction::IntToPtr:
1138 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1139 // FALL THROUGH and handle them the same as zext/trunc.
1140 case Instruction::ZExt:
1141 case Instruction::Trunc: {
1142 Type *SrcTy = I->getOperand(0)->getType();
1144 unsigned SrcBitWidth;
1145 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1146 // which fall through here.
1147 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1149 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1150 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1151 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1152 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1153 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1154 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1155 // Any top bits are known to be zero.
1156 if (BitWidth > SrcBitWidth)
1157 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1160 case Instruction::BitCast: {
1161 Type *SrcTy = I->getOperand(0)->getType();
1162 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy() ||
1163 SrcTy->isFloatingPointTy()) &&
1164 // TODO: For now, not handling conversions like:
1165 // (bitcast i64 %x to <2 x i32>)
1166 !I->getType()->isVectorTy()) {
1167 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1172 case Instruction::SExt: {
1173 // Compute the bits in the result that are not present in the input.
1174 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1176 KnownZero = KnownZero.trunc(SrcBitWidth);
1177 KnownOne = KnownOne.trunc(SrcBitWidth);
1178 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1179 KnownZero = KnownZero.zext(BitWidth);
1180 KnownOne = KnownOne.zext(BitWidth);
1182 // If the sign bit of the input is known set or clear, then we know the
1183 // top bits of the result.
1184 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1185 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1186 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1187 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1190 case Instruction::Shl: {
1191 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1192 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1193 return (KnownZero << ShiftAmt) |
1194 APInt::getLowBitsSet(BitWidth, ShiftAmt); // Low bits known 0.
1197 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1198 return KnownOne << ShiftAmt;
1201 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1202 KnownZero2, KnownOne2, DL, Depth, Q,
1206 case Instruction::LShr: {
1207 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1208 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1209 return APIntOps::lshr(KnownZero, ShiftAmt) |
1210 // High bits known zero.
1211 APInt::getHighBitsSet(BitWidth, ShiftAmt);
1214 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1215 return APIntOps::lshr(KnownOne, ShiftAmt);
1218 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1219 KnownZero2, KnownOne2, DL, Depth, Q,
1223 case Instruction::AShr: {
1224 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1225 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1226 return APIntOps::ashr(KnownZero, ShiftAmt);
1229 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1230 return APIntOps::ashr(KnownOne, ShiftAmt);
1233 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1234 KnownZero2, KnownOne2, DL, Depth, Q,
1238 case Instruction::Sub: {
1239 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1240 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1241 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1245 case Instruction::Add: {
1246 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1247 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1248 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1252 case Instruction::SRem:
1253 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1254 APInt RA = Rem->getValue().abs();
1255 if (RA.isPowerOf2()) {
1256 APInt LowBits = RA - 1;
1257 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1260 // The low bits of the first operand are unchanged by the srem.
1261 KnownZero = KnownZero2 & LowBits;
1262 KnownOne = KnownOne2 & LowBits;
1264 // If the first operand is non-negative or has all low bits zero, then
1265 // the upper bits are all zero.
1266 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1267 KnownZero |= ~LowBits;
1269 // If the first operand is negative and not all low bits are zero, then
1270 // the upper bits are all one.
1271 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1272 KnownOne |= ~LowBits;
1274 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1278 // The sign bit is the LHS's sign bit, except when the result of the
1279 // remainder is zero.
1280 if (KnownZero.isNonNegative()) {
1281 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1282 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1284 // If it's known zero, our sign bit is also zero.
1285 if (LHSKnownZero.isNegative())
1286 KnownZero.setBit(BitWidth - 1);
1290 case Instruction::URem: {
1291 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1292 APInt RA = Rem->getValue();
1293 if (RA.isPowerOf2()) {
1294 APInt LowBits = (RA - 1);
1295 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1297 KnownZero |= ~LowBits;
1298 KnownOne &= LowBits;
1303 // Since the result is less than or equal to either operand, any leading
1304 // zero bits in either operand must also exist in the result.
1305 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1306 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1308 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1309 KnownZero2.countLeadingOnes());
1310 KnownOne.clearAllBits();
1311 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1315 case Instruction::Alloca: {
1316 AllocaInst *AI = cast<AllocaInst>(I);
1317 unsigned Align = AI->getAlignment();
1319 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1322 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1325 case Instruction::GetElementPtr: {
1326 // Analyze all of the subscripts of this getelementptr instruction
1327 // to determine if we can prove known low zero bits.
1328 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1329 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1331 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1333 gep_type_iterator GTI = gep_type_begin(I);
1334 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1335 Value *Index = I->getOperand(i);
1336 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1337 // Handle struct member offset arithmetic.
1339 // Handle case when index is vector zeroinitializer
1340 Constant *CIndex = cast<Constant>(Index);
1341 if (CIndex->isZeroValue())
1344 if (CIndex->getType()->isVectorTy())
1345 Index = CIndex->getSplatValue();
1347 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1348 const StructLayout *SL = DL.getStructLayout(STy);
1349 uint64_t Offset = SL->getElementOffset(Idx);
1350 TrailZ = std::min<unsigned>(TrailZ,
1351 countTrailingZeros(Offset));
1353 // Handle array index arithmetic.
1354 Type *IndexedTy = GTI.getIndexedType();
1355 if (!IndexedTy->isSized()) {
1359 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1360 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1361 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1362 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1364 TrailZ = std::min(TrailZ,
1365 unsigned(countTrailingZeros(TypeSize) +
1366 LocalKnownZero.countTrailingOnes()));
1370 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1373 case Instruction::PHI: {
1374 PHINode *P = cast<PHINode>(I);
1375 // Handle the case of a simple two-predecessor recurrence PHI.
1376 // There's a lot more that could theoretically be done here, but
1377 // this is sufficient to catch some interesting cases.
1378 if (P->getNumIncomingValues() == 2) {
1379 for (unsigned i = 0; i != 2; ++i) {
1380 Value *L = P->getIncomingValue(i);
1381 Value *R = P->getIncomingValue(!i);
1382 Operator *LU = dyn_cast<Operator>(L);
1385 unsigned Opcode = LU->getOpcode();
1386 // Check for operations that have the property that if
1387 // both their operands have low zero bits, the result
1388 // will have low zero bits.
1389 if (Opcode == Instruction::Add ||
1390 Opcode == Instruction::Sub ||
1391 Opcode == Instruction::And ||
1392 Opcode == Instruction::Or ||
1393 Opcode == Instruction::Mul) {
1394 Value *LL = LU->getOperand(0);
1395 Value *LR = LU->getOperand(1);
1396 // Find a recurrence.
1403 // Ok, we have a PHI of the form L op= R. Check for low
1405 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1407 // We need to take the minimum number of known bits
1408 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1409 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1411 KnownZero = APInt::getLowBitsSet(BitWidth,
1412 std::min(KnownZero2.countTrailingOnes(),
1413 KnownZero3.countTrailingOnes()));
1419 // Unreachable blocks may have zero-operand PHI nodes.
1420 if (P->getNumIncomingValues() == 0)
1423 // Otherwise take the unions of the known bit sets of the operands,
1424 // taking conservative care to avoid excessive recursion.
1425 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1426 // Skip if every incoming value references to ourself.
1427 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1430 KnownZero = APInt::getAllOnesValue(BitWidth);
1431 KnownOne = APInt::getAllOnesValue(BitWidth);
1432 for (Value *IncValue : P->incoming_values()) {
1433 // Skip direct self references.
1434 if (IncValue == P) continue;
1436 KnownZero2 = APInt(BitWidth, 0);
1437 KnownOne2 = APInt(BitWidth, 0);
1438 // Recurse, but cap the recursion to one level, because we don't
1439 // want to waste time spinning around in loops.
1440 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1442 KnownZero &= KnownZero2;
1443 KnownOne &= KnownOne2;
1444 // If all bits have been ruled out, there's no need to check
1446 if (!KnownZero && !KnownOne)
1452 case Instruction::Call:
1453 case Instruction::Invoke:
1454 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1455 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1456 // If a range metadata is attached to this IntrinsicInst, intersect the
1457 // explicit range specified by the metadata and the implicit range of
1459 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1460 switch (II->getIntrinsicID()) {
1462 case Intrinsic::bswap:
1463 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1465 KnownZero |= KnownZero2.byteSwap();
1466 KnownOne |= KnownOne2.byteSwap();
1468 case Intrinsic::ctlz:
1469 case Intrinsic::cttz: {
1470 unsigned LowBits = Log2_32(BitWidth)+1;
1471 // If this call is undefined for 0, the result will be less than 2^n.
1472 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1474 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1477 case Intrinsic::ctpop: {
1478 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1480 // We can bound the space the count needs. Also, bits known to be zero
1481 // can't contribute to the population.
1482 unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation();
1483 unsigned LeadingZeros =
1484 APInt(BitWidth, BitsPossiblySet).countLeadingZeros();
1485 assert(LeadingZeros <= BitWidth);
1486 KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros);
1487 KnownOne &= ~KnownZero;
1488 // TODO: we could bound KnownOne using the lower bound on the number
1489 // of bits which might be set provided by popcnt KnownOne2.
1492 case Intrinsic::fabs: {
1493 Type *Ty = II->getType();
1494 APInt SignBit = APInt::getSignBit(Ty->getScalarSizeInBits());
1495 KnownZero |= APInt::getSplat(Ty->getPrimitiveSizeInBits(), SignBit);
1498 case Intrinsic::x86_sse42_crc32_64_64:
1499 KnownZero |= APInt::getHighBitsSet(64, 32);
1504 case Instruction::ExtractValue:
1505 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1506 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1507 if (EVI->getNumIndices() != 1) break;
1508 if (EVI->getIndices()[0] == 0) {
1509 switch (II->getIntrinsicID()) {
1511 case Intrinsic::uadd_with_overflow:
1512 case Intrinsic::sadd_with_overflow:
1513 computeKnownBitsAddSub(true, II->getArgOperand(0),
1514 II->getArgOperand(1), false, KnownZero,
1515 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1517 case Intrinsic::usub_with_overflow:
1518 case Intrinsic::ssub_with_overflow:
1519 computeKnownBitsAddSub(false, II->getArgOperand(0),
1520 II->getArgOperand(1), false, KnownZero,
1521 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1523 case Intrinsic::umul_with_overflow:
1524 case Intrinsic::smul_with_overflow:
1525 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1526 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1535 static unsigned getAlignment(const Value *V, const DataLayout &DL) {
1537 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1538 Align = GO->getAlignment();
1540 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1541 Type *ObjectType = GVar->getType()->getElementType();
1542 if (ObjectType->isSized()) {
1543 // If the object is defined in the current Module, we'll be giving
1544 // it the preferred alignment. Otherwise, we have to assume that it
1545 // may only have the minimum ABI alignment.
1546 if (GVar->isStrongDefinitionForLinker())
1547 Align = DL.getPreferredAlignment(GVar);
1549 Align = DL.getABITypeAlignment(ObjectType);
1553 } else if (const Argument *A = dyn_cast<Argument>(V)) {
1554 Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1556 if (!Align && A->hasStructRetAttr()) {
1557 // An sret parameter has at least the ABI alignment of the return type.
1558 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1559 if (EltTy->isSized())
1560 Align = DL.getABITypeAlignment(EltTy);
1562 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
1563 Align = AI->getAlignment();
1564 else if (auto CS = ImmutableCallSite(V))
1565 Align = CS.getAttributes().getParamAlignment(AttributeSet::ReturnIndex);
1566 else if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1567 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_align)) {
1568 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
1569 Align = CI->getLimitedValue();
1575 /// Determine which bits of V are known to be either zero or one and return
1576 /// them in the KnownZero/KnownOne bit sets.
1578 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1579 /// we cannot optimize based on the assumption that it is zero without changing
1580 /// it to be an explicit zero. If we don't change it to zero, other code could
1581 /// optimized based on the contradictory assumption that it is non-zero.
1582 /// Because instcombine aggressively folds operations with undef args anyway,
1583 /// this won't lose us code quality.
1585 /// This function is defined on values with integer type, values with pointer
1586 /// type, and vectors of integers. In the case
1587 /// where V is a vector, known zero, and known one values are the
1588 /// same width as the vector element, and the bit is set only if it is true
1589 /// for all of the elements in the vector.
1590 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1591 const DataLayout &DL, unsigned Depth, const Query &Q) {
1592 assert(V && "No Value?");
1593 assert(Depth <= MaxDepth && "Limit Search Depth");
1594 unsigned BitWidth = KnownZero.getBitWidth();
1596 assert((V->getType()->isIntOrIntVectorTy() ||
1597 V->getType()->isFPOrFPVectorTy() ||
1598 V->getType()->getScalarType()->isPointerTy()) &&
1599 "Not integer, floating point, or pointer type!");
1600 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1601 (!V->getType()->isIntOrIntVectorTy() ||
1602 V->getType()->getScalarSizeInBits() == BitWidth) &&
1603 KnownZero.getBitWidth() == BitWidth &&
1604 KnownOne.getBitWidth() == BitWidth &&
1605 "V, KnownOne and KnownZero should have same BitWidth");
1607 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1608 // We know all of the bits for a constant!
1609 KnownOne = CI->getValue();
1610 KnownZero = ~KnownOne;
1613 // Null and aggregate-zero are all-zeros.
1614 if (isa<ConstantPointerNull>(V) ||
1615 isa<ConstantAggregateZero>(V)) {
1616 KnownOne.clearAllBits();
1617 KnownZero = APInt::getAllOnesValue(BitWidth);
1620 // Handle a constant vector by taking the intersection of the known bits of
1621 // each element. There is no real need to handle ConstantVector here, because
1622 // we don't handle undef in any particularly useful way.
1623 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1624 // We know that CDS must be a vector of integers. Take the intersection of
1626 KnownZero.setAllBits(); KnownOne.setAllBits();
1627 APInt Elt(KnownZero.getBitWidth(), 0);
1628 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1629 Elt = CDS->getElementAsInteger(i);
1636 // Start out not knowing anything.
1637 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1639 // Limit search depth.
1640 // All recursive calls that increase depth must come after this.
1641 if (Depth == MaxDepth)
1644 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1645 // the bits of its aliasee.
1646 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1647 if (!GA->mayBeOverridden())
1648 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1652 if (Operator *I = dyn_cast<Operator>(V))
1653 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1655 // Aligned pointers have trailing zeros - refine KnownZero set
1656 if (V->getType()->isPointerTy()) {
1657 unsigned Align = getAlignment(V, DL);
1659 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1662 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1663 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1664 // computeKnownBitsFromOperator.
1666 // Check whether a nearby assume intrinsic can determine some known bits.
1667 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1669 // Check whether there's a dominating condition which implies something about
1670 // this value at the given context.
1671 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1672 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1675 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1678 /// Determine whether the sign bit is known to be zero or one.
1679 /// Convenience wrapper around computeKnownBits.
1680 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1681 const DataLayout &DL, unsigned Depth, const Query &Q) {
1682 unsigned BitWidth = getBitWidth(V->getType(), DL);
1688 APInt ZeroBits(BitWidth, 0);
1689 APInt OneBits(BitWidth, 0);
1690 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1691 KnownOne = OneBits[BitWidth - 1];
1692 KnownZero = ZeroBits[BitWidth - 1];
1695 /// Return true if the given value is known to have exactly one
1696 /// bit set when defined. For vectors return true if every element is known to
1697 /// be a power of two when defined. Supports values with integer or pointer
1698 /// types and vectors of integers.
1699 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1700 const Query &Q, const DataLayout &DL) {
1701 if (Constant *C = dyn_cast<Constant>(V)) {
1702 if (C->isNullValue())
1704 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1705 return CI->getValue().isPowerOf2();
1706 // TODO: Handle vector constants.
1709 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1710 // it is shifted off the end then the result is undefined.
1711 if (match(V, m_Shl(m_One(), m_Value())))
1714 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1715 // bottom. If it is shifted off the bottom then the result is undefined.
1716 if (match(V, m_LShr(m_SignBit(), m_Value())))
1719 // The remaining tests are all recursive, so bail out if we hit the limit.
1720 if (Depth++ == MaxDepth)
1723 Value *X = nullptr, *Y = nullptr;
1724 // A shift of a power of two is a power of two or zero.
1725 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1726 match(V, m_Shr(m_Value(X), m_Value()))))
1727 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1729 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1730 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1732 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1733 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1734 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1736 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1737 // A power of two and'd with anything is a power of two or zero.
1738 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1739 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1741 // X & (-X) is always a power of two or zero.
1742 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1747 // Adding a power-of-two or zero to the same power-of-two or zero yields
1748 // either the original power-of-two, a larger power-of-two or zero.
1749 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1750 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1751 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1752 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1753 match(X, m_And(m_Value(), m_Specific(Y))))
1754 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1756 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1757 match(Y, m_And(m_Value(), m_Specific(X))))
1758 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1761 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1762 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1763 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1765 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1766 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1767 // If i8 V is a power of two or zero:
1768 // ZeroBits: 1 1 1 0 1 1 1 1
1769 // ~ZeroBits: 0 0 0 1 0 0 0 0
1770 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1771 // If OrZero isn't set, we cannot give back a zero result.
1772 // Make sure either the LHS or RHS has a bit set.
1773 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1778 // An exact divide or right shift can only shift off zero bits, so the result
1779 // is a power of two only if the first operand is a power of two and not
1780 // copying a sign bit (sdiv int_min, 2).
1781 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1782 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1783 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1790 /// \brief Test whether a GEP's result is known to be non-null.
1792 /// Uses properties inherent in a GEP to try to determine whether it is known
1795 /// Currently this routine does not support vector GEPs.
1796 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1797 unsigned Depth, const Query &Q) {
1798 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1801 // FIXME: Support vector-GEPs.
1802 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1804 // If the base pointer is non-null, we cannot walk to a null address with an
1805 // inbounds GEP in address space zero.
1806 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1809 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1810 // If so, then the GEP cannot produce a null pointer, as doing so would
1811 // inherently violate the inbounds contract within address space zero.
1812 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1813 GTI != GTE; ++GTI) {
1814 // Struct types are easy -- they must always be indexed by a constant.
1815 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1816 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1817 unsigned ElementIdx = OpC->getZExtValue();
1818 const StructLayout *SL = DL.getStructLayout(STy);
1819 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1820 if (ElementOffset > 0)
1825 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1826 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1829 // Fast path the constant operand case both for efficiency and so we don't
1830 // increment Depth when just zipping down an all-constant GEP.
1831 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1837 // We post-increment Depth here because while isKnownNonZero increments it
1838 // as well, when we pop back up that increment won't persist. We don't want
1839 // to recurse 10k times just because we have 10k GEP operands. We don't
1840 // bail completely out because we want to handle constant GEPs regardless
1842 if (Depth++ >= MaxDepth)
1845 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1852 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1853 /// ensure that the value it's attached to is never Value? 'RangeType' is
1854 /// is the type of the value described by the range.
1855 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1856 const APInt& Value) {
1857 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1858 assert(NumRanges >= 1);
1859 for (unsigned i = 0; i < NumRanges; ++i) {
1860 ConstantInt *Lower =
1861 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1862 ConstantInt *Upper =
1863 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1864 ConstantRange Range(Lower->getValue(), Upper->getValue());
1865 if (Range.contains(Value))
1871 /// Return true if the given value is known to be non-zero when defined.
1872 /// For vectors return true if every element is known to be non-zero when
1873 /// defined. Supports values with integer or pointer type and vectors of
1875 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1877 if (Constant *C = dyn_cast<Constant>(V)) {
1878 if (C->isNullValue())
1880 if (isa<ConstantInt>(C))
1881 // Must be non-zero due to null test above.
1883 // TODO: Handle vectors
1887 if (Instruction* I = dyn_cast<Instruction>(V)) {
1888 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1889 // If the possible ranges don't contain zero, then the value is
1890 // definitely non-zero.
1891 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1892 const APInt ZeroValue(Ty->getBitWidth(), 0);
1893 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1899 // The remaining tests are all recursive, so bail out if we hit the limit.
1900 if (Depth++ >= MaxDepth)
1903 // Check for pointer simplifications.
1904 if (V->getType()->isPointerTy()) {
1905 if (isKnownNonNull(V))
1907 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1908 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1912 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1914 // X | Y != 0 if X != 0 or Y != 0.
1915 Value *X = nullptr, *Y = nullptr;
1916 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1917 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1919 // ext X != 0 if X != 0.
1920 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1921 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1923 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1924 // if the lowest bit is shifted off the end.
1925 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1926 // shl nuw can't remove any non-zero bits.
1927 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1928 if (BO->hasNoUnsignedWrap())
1929 return isKnownNonZero(X, DL, Depth, Q);
1931 APInt KnownZero(BitWidth, 0);
1932 APInt KnownOne(BitWidth, 0);
1933 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1937 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1938 // defined if the sign bit is shifted off the end.
1939 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1940 // shr exact can only shift out zero bits.
1941 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1943 return isKnownNonZero(X, DL, Depth, Q);
1945 bool XKnownNonNegative, XKnownNegative;
1946 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1950 // If the shifter operand is a constant, and all of the bits shifted
1951 // out are known to be zero, and X is known non-zero then at least one
1952 // non-zero bit must remain.
1953 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1954 APInt KnownZero(BitWidth, 0);
1955 APInt KnownOne(BitWidth, 0);
1956 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1958 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1959 // Is there a known one in the portion not shifted out?
1960 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1962 // Are all the bits to be shifted out known zero?
1963 if (KnownZero.countTrailingOnes() >= ShiftVal)
1964 return isKnownNonZero(X, DL, Depth, Q);
1967 // div exact can only produce a zero if the dividend is zero.
1968 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1969 return isKnownNonZero(X, DL, Depth, Q);
1972 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1973 bool XKnownNonNegative, XKnownNegative;
1974 bool YKnownNonNegative, YKnownNegative;
1975 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1976 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
1978 // If X and Y are both non-negative (as signed values) then their sum is not
1979 // zero unless both X and Y are zero.
1980 if (XKnownNonNegative && YKnownNonNegative)
1981 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
1984 // If X and Y are both negative (as signed values) then their sum is not
1985 // zero unless both X and Y equal INT_MIN.
1986 if (BitWidth && XKnownNegative && YKnownNegative) {
1987 APInt KnownZero(BitWidth, 0);
1988 APInt KnownOne(BitWidth, 0);
1989 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1990 // The sign bit of X is set. If some other bit is set then X is not equal
1992 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1993 if ((KnownOne & Mask) != 0)
1995 // The sign bit of Y is set. If some other bit is set then Y is not equal
1997 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
1998 if ((KnownOne & Mask) != 0)
2002 // The sum of a non-negative number and a power of two is not zero.
2003 if (XKnownNonNegative &&
2004 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
2006 if (YKnownNonNegative &&
2007 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
2011 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2012 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2013 // If X and Y are non-zero then so is X * Y as long as the multiplication
2014 // does not overflow.
2015 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
2016 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
2019 // (C ? X : Y) != 0 if X != 0 and Y != 0.
2020 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2021 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
2022 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
2026 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
2027 // Try and detect a recurrence that monotonically increases from a
2028 // starting value, as these are common as induction variables.
2029 if (PN->getNumIncomingValues() == 2) {
2030 Value *Start = PN->getIncomingValue(0);
2031 Value *Induction = PN->getIncomingValue(1);
2032 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2033 std::swap(Start, Induction);
2034 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2035 if (!C->isZero() && !C->isNegative()) {
2037 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2038 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2046 if (!BitWidth) return false;
2047 APInt KnownZero(BitWidth, 0);
2048 APInt KnownOne(BitWidth, 0);
2049 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2050 return KnownOne != 0;
2053 /// Return true if V2 == V1 + X, where X is known non-zero.
2054 static bool isAddOfNonZero(Value *V1, Value *V2, const DataLayout &DL,
2056 BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2057 if (!BO || BO->getOpcode() != Instruction::Add)
2059 Value *Op = nullptr;
2060 if (V2 == BO->getOperand(0))
2061 Op = BO->getOperand(1);
2062 else if (V2 == BO->getOperand(1))
2063 Op = BO->getOperand(0);
2066 return isKnownNonZero(Op, DL, 0, Q);
2069 /// Return true if it is known that V1 != V2.
2070 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
2072 if (V1->getType()->isVectorTy() || V1 == V2)
2074 if (V1->getType() != V2->getType())
2075 // We can't look through casts yet.
2077 if (isAddOfNonZero(V1, V2, DL, Q) || isAddOfNonZero(V2, V1, DL, Q))
2080 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
2081 // Are any known bits in V1 contradictory to known bits in V2? If V1
2082 // has a known zero where V2 has a known one, they must not be equal.
2083 auto BitWidth = Ty->getBitWidth();
2084 APInt KnownZero1(BitWidth, 0);
2085 APInt KnownOne1(BitWidth, 0);
2086 computeKnownBits(V1, KnownZero1, KnownOne1, DL, 0, Q);
2087 APInt KnownZero2(BitWidth, 0);
2088 APInt KnownOne2(BitWidth, 0);
2089 computeKnownBits(V2, KnownZero2, KnownOne2, DL, 0, Q);
2091 auto OppositeBits = (KnownZero1 & KnownOne2) | (KnownZero2 & KnownOne1);
2092 if (OppositeBits.getBoolValue())
2098 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2099 /// simplify operations downstream. Mask is known to be zero for bits that V
2102 /// This function is defined on values with integer type, values with pointer
2103 /// type, and vectors of integers. In the case
2104 /// where V is a vector, the mask, known zero, and known one values are the
2105 /// same width as the vector element, and the bit is set only if it is true
2106 /// for all of the elements in the vector.
2107 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
2108 unsigned Depth, const Query &Q) {
2109 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
2110 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2111 return (KnownZero & Mask) == Mask;
2116 /// Return the number of times the sign bit of the register is replicated into
2117 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2118 /// (itself), but other cases can give us information. For example, immediately
2119 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2120 /// other, so we return 3.
2122 /// 'Op' must have a scalar integer type.
2124 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
2126 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
2128 unsigned FirstAnswer = 1;
2130 // Note that ConstantInt is handled by the general computeKnownBits case
2134 return 1; // Limit search depth.
2136 Operator *U = dyn_cast<Operator>(V);
2137 switch (Operator::getOpcode(V)) {
2139 case Instruction::SExt:
2140 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2141 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
2143 case Instruction::SDiv: {
2144 const APInt *Denominator;
2145 // sdiv X, C -> adds log(C) sign bits.
2146 if (match(U->getOperand(1), m_APInt(Denominator))) {
2148 // Ignore non-positive denominator.
2149 if (!Denominator->isStrictlyPositive())
2152 // Calculate the incoming numerator bits.
2153 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2155 // Add floor(log(C)) bits to the numerator bits.
2156 return std::min(TyBits, NumBits + Denominator->logBase2());
2161 case Instruction::SRem: {
2162 const APInt *Denominator;
2163 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2164 // positive constant. This let us put a lower bound on the number of sign
2166 if (match(U->getOperand(1), m_APInt(Denominator))) {
2168 // Ignore non-positive denominator.
2169 if (!Denominator->isStrictlyPositive())
2172 // Calculate the incoming numerator bits. SRem by a positive constant
2173 // can't lower the number of sign bits.
2175 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2177 // Calculate the leading sign bit constraints by examining the
2178 // denominator. Given that the denominator is positive, there are two
2181 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2182 // (1 << ceilLogBase2(C)).
2184 // 2. the numerator is negative. Then the result range is (-C,0] and
2185 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2187 // Thus a lower bound on the number of sign bits is `TyBits -
2188 // ceilLogBase2(C)`.
2190 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2191 return std::max(NumrBits, ResBits);
2196 case Instruction::AShr: {
2197 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2198 // ashr X, C -> adds C sign bits. Vectors too.
2200 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2201 Tmp += ShAmt->getZExtValue();
2202 if (Tmp > TyBits) Tmp = TyBits;
2206 case Instruction::Shl: {
2208 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2209 // shl destroys sign bits.
2210 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2211 Tmp2 = ShAmt->getZExtValue();
2212 if (Tmp2 >= TyBits || // Bad shift.
2213 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2218 case Instruction::And:
2219 case Instruction::Or:
2220 case Instruction::Xor: // NOT is handled here.
2221 // Logical binary ops preserve the number of sign bits at the worst.
2222 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2224 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2225 FirstAnswer = std::min(Tmp, Tmp2);
2226 // We computed what we know about the sign bits as our first
2227 // answer. Now proceed to the generic code that uses
2228 // computeKnownBits, and pick whichever answer is better.
2232 case Instruction::Select:
2233 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2234 if (Tmp == 1) return 1; // Early out.
2235 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2236 return std::min(Tmp, Tmp2);
2238 case Instruction::Add:
2239 // Add can have at most one carry bit. Thus we know that the output
2240 // is, at worst, one more bit than the inputs.
2241 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2242 if (Tmp == 1) return 1; // Early out.
2244 // Special case decrementing a value (ADD X, -1):
2245 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2246 if (CRHS->isAllOnesValue()) {
2247 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2248 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2251 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2253 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2256 // If we are subtracting one from a positive number, there is no carry
2257 // out of the result.
2258 if (KnownZero.isNegative())
2262 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2263 if (Tmp2 == 1) return 1;
2264 return std::min(Tmp, Tmp2)-1;
2266 case Instruction::Sub:
2267 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2268 if (Tmp2 == 1) return 1;
2271 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2272 if (CLHS->isNullValue()) {
2273 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2274 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2276 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2278 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2281 // If the input is known to be positive (the sign bit is known clear),
2282 // the output of the NEG has the same number of sign bits as the input.
2283 if (KnownZero.isNegative())
2286 // Otherwise, we treat this like a SUB.
2289 // Sub can have at most one carry bit. Thus we know that the output
2290 // is, at worst, one more bit than the inputs.
2291 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2292 if (Tmp == 1) return 1; // Early out.
2293 return std::min(Tmp, Tmp2)-1;
2295 case Instruction::PHI: {
2296 PHINode *PN = cast<PHINode>(U);
2297 unsigned NumIncomingValues = PN->getNumIncomingValues();
2298 // Don't analyze large in-degree PHIs.
2299 if (NumIncomingValues > 4) break;
2300 // Unreachable blocks may have zero-operand PHI nodes.
2301 if (NumIncomingValues == 0) break;
2303 // Take the minimum of all incoming values. This can't infinitely loop
2304 // because of our depth threshold.
2305 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2306 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2307 if (Tmp == 1) return Tmp;
2309 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2314 case Instruction::Trunc:
2315 // FIXME: it's tricky to do anything useful for this, but it is an important
2316 // case for targets like X86.
2320 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2321 // use this information.
2322 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2324 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2326 if (KnownZero.isNegative()) { // sign bit is 0
2328 } else if (KnownOne.isNegative()) { // sign bit is 1;
2335 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2336 // the number of identical bits in the top of the input value.
2338 Mask <<= Mask.getBitWidth()-TyBits;
2339 // Return # leading zeros. We use 'min' here in case Val was zero before
2340 // shifting. We don't want to return '64' as for an i32 "0".
2341 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2344 /// This function computes the integer multiple of Base that equals V.
2345 /// If successful, it returns true and returns the multiple in
2346 /// Multiple. If unsuccessful, it returns false. It looks
2347 /// through SExt instructions only if LookThroughSExt is true.
2348 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2349 bool LookThroughSExt, unsigned Depth) {
2350 const unsigned MaxDepth = 6;
2352 assert(V && "No Value?");
2353 assert(Depth <= MaxDepth && "Limit Search Depth");
2354 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2356 Type *T = V->getType();
2358 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2368 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2369 Constant *BaseVal = ConstantInt::get(T, Base);
2370 if (CO && CO == BaseVal) {
2372 Multiple = ConstantInt::get(T, 1);
2376 if (CI && CI->getZExtValue() % Base == 0) {
2377 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2381 if (Depth == MaxDepth) return false; // Limit search depth.
2383 Operator *I = dyn_cast<Operator>(V);
2384 if (!I) return false;
2386 switch (I->getOpcode()) {
2388 case Instruction::SExt:
2389 if (!LookThroughSExt) return false;
2390 // otherwise fall through to ZExt
2391 case Instruction::ZExt:
2392 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2393 LookThroughSExt, Depth+1);
2394 case Instruction::Shl:
2395 case Instruction::Mul: {
2396 Value *Op0 = I->getOperand(0);
2397 Value *Op1 = I->getOperand(1);
2399 if (I->getOpcode() == Instruction::Shl) {
2400 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2401 if (!Op1CI) return false;
2402 // Turn Op0 << Op1 into Op0 * 2^Op1
2403 APInt Op1Int = Op1CI->getValue();
2404 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2405 APInt API(Op1Int.getBitWidth(), 0);
2406 API.setBit(BitToSet);
2407 Op1 = ConstantInt::get(V->getContext(), API);
2410 Value *Mul0 = nullptr;
2411 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2412 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2413 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2414 if (Op1C->getType()->getPrimitiveSizeInBits() <
2415 MulC->getType()->getPrimitiveSizeInBits())
2416 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2417 if (Op1C->getType()->getPrimitiveSizeInBits() >
2418 MulC->getType()->getPrimitiveSizeInBits())
2419 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2421 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2422 Multiple = ConstantExpr::getMul(MulC, Op1C);
2426 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2427 if (Mul0CI->getValue() == 1) {
2428 // V == Base * Op1, so return Op1
2434 Value *Mul1 = nullptr;
2435 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2436 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2437 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2438 if (Op0C->getType()->getPrimitiveSizeInBits() <
2439 MulC->getType()->getPrimitiveSizeInBits())
2440 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2441 if (Op0C->getType()->getPrimitiveSizeInBits() >
2442 MulC->getType()->getPrimitiveSizeInBits())
2443 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2445 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2446 Multiple = ConstantExpr::getMul(MulC, Op0C);
2450 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2451 if (Mul1CI->getValue() == 1) {
2452 // V == Base * Op0, so return Op0
2460 // We could not determine if V is a multiple of Base.
2464 /// Return true if we can prove that the specified FP value is never equal to
2467 /// NOTE: this function will need to be revisited when we support non-default
2470 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2471 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2472 return !CFP->getValueAPF().isNegZero();
2474 // FIXME: Magic number! At the least, this should be given a name because it's
2475 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2476 // expose it as a parameter, so it can be used for testing / experimenting.
2478 return false; // Limit search depth.
2480 const Operator *I = dyn_cast<Operator>(V);
2481 if (!I) return false;
2483 // Check if the nsz fast-math flag is set
2484 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2485 if (FPO->hasNoSignedZeros())
2488 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2489 if (I->getOpcode() == Instruction::FAdd)
2490 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2491 if (CFP->isNullValue())
2494 // sitofp and uitofp turn into +0.0 for zero.
2495 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2498 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2499 // sqrt(-0.0) = -0.0, no other negative results are possible.
2500 if (II->getIntrinsicID() == Intrinsic::sqrt)
2501 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2503 if (const CallInst *CI = dyn_cast<CallInst>(I))
2504 if (const Function *F = CI->getCalledFunction()) {
2505 if (F->isDeclaration()) {
2507 if (F->getName() == "abs") return true;
2508 // fabs[lf](x) != -0.0
2509 if (F->getName() == "fabs") return true;
2510 if (F->getName() == "fabsf") return true;
2511 if (F->getName() == "fabsl") return true;
2512 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2513 F->getName() == "sqrtl")
2514 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2521 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2522 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2523 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2525 // FIXME: Magic number! At the least, this should be given a name because it's
2526 // used similarly in CannotBeNegativeZero(). A better fix may be to
2527 // expose it as a parameter, so it can be used for testing / experimenting.
2529 return false; // Limit search depth.
2531 const Operator *I = dyn_cast<Operator>(V);
2532 if (!I) return false;
2534 switch (I->getOpcode()) {
2536 case Instruction::FMul:
2537 // x*x is always non-negative or a NaN.
2538 if (I->getOperand(0) == I->getOperand(1))
2541 case Instruction::FAdd:
2542 case Instruction::FDiv:
2543 case Instruction::FRem:
2544 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2545 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2546 case Instruction::FPExt:
2547 case Instruction::FPTrunc:
2548 // Widening/narrowing never change sign.
2549 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2550 case Instruction::Call:
2551 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2552 switch (II->getIntrinsicID()) {
2554 case Intrinsic::exp:
2555 case Intrinsic::exp2:
2556 case Intrinsic::fabs:
2557 case Intrinsic::sqrt:
2559 case Intrinsic::powi:
2560 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2561 // powi(x,n) is non-negative if n is even.
2562 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2565 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2566 case Intrinsic::fma:
2567 case Intrinsic::fmuladd:
2568 // x*x+y is non-negative if y is non-negative.
2569 return I->getOperand(0) == I->getOperand(1) &&
2570 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2577 /// If the specified value can be set by repeating the same byte in memory,
2578 /// return the i8 value that it is represented with. This is
2579 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2580 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2581 /// byte store (e.g. i16 0x1234), return null.
2582 Value *llvm::isBytewiseValue(Value *V) {
2583 // All byte-wide stores are splatable, even of arbitrary variables.
2584 if (V->getType()->isIntegerTy(8)) return V;
2586 // Handle 'null' ConstantArrayZero etc.
2587 if (Constant *C = dyn_cast<Constant>(V))
2588 if (C->isNullValue())
2589 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2591 // Constant float and double values can be handled as integer values if the
2592 // corresponding integer value is "byteable". An important case is 0.0.
2593 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2594 if (CFP->getType()->isFloatTy())
2595 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2596 if (CFP->getType()->isDoubleTy())
2597 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2598 // Don't handle long double formats, which have strange constraints.
2601 // We can handle constant integers that are multiple of 8 bits.
2602 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2603 if (CI->getBitWidth() % 8 == 0) {
2604 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2606 if (!CI->getValue().isSplat(8))
2608 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2612 // A ConstantDataArray/Vector is splatable if all its members are equal and
2614 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2615 Value *Elt = CA->getElementAsConstant(0);
2616 Value *Val = isBytewiseValue(Elt);
2620 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2621 if (CA->getElementAsConstant(I) != Elt)
2627 // Conceptually, we could handle things like:
2628 // %a = zext i8 %X to i16
2629 // %b = shl i16 %a, 8
2630 // %c = or i16 %a, %b
2631 // but until there is an example that actually needs this, it doesn't seem
2632 // worth worrying about.
2637 // This is the recursive version of BuildSubAggregate. It takes a few different
2638 // arguments. Idxs is the index within the nested struct From that we are
2639 // looking at now (which is of type IndexedType). IdxSkip is the number of
2640 // indices from Idxs that should be left out when inserting into the resulting
2641 // struct. To is the result struct built so far, new insertvalue instructions
2643 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2644 SmallVectorImpl<unsigned> &Idxs,
2646 Instruction *InsertBefore) {
2647 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2649 // Save the original To argument so we can modify it
2651 // General case, the type indexed by Idxs is a struct
2652 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2653 // Process each struct element recursively
2656 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2660 // Couldn't find any inserted value for this index? Cleanup
2661 while (PrevTo != OrigTo) {
2662 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2663 PrevTo = Del->getAggregateOperand();
2664 Del->eraseFromParent();
2666 // Stop processing elements
2670 // If we successfully found a value for each of our subaggregates
2674 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2675 // the struct's elements had a value that was inserted directly. In the latter
2676 // case, perhaps we can't determine each of the subelements individually, but
2677 // we might be able to find the complete struct somewhere.
2679 // Find the value that is at that particular spot
2680 Value *V = FindInsertedValue(From, Idxs);
2685 // Insert the value in the new (sub) aggregrate
2686 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2687 "tmp", InsertBefore);
2690 // This helper takes a nested struct and extracts a part of it (which is again a
2691 // struct) into a new value. For example, given the struct:
2692 // { a, { b, { c, d }, e } }
2693 // and the indices "1, 1" this returns
2696 // It does this by inserting an insertvalue for each element in the resulting
2697 // struct, as opposed to just inserting a single struct. This will only work if
2698 // each of the elements of the substruct are known (ie, inserted into From by an
2699 // insertvalue instruction somewhere).
2701 // All inserted insertvalue instructions are inserted before InsertBefore
2702 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2703 Instruction *InsertBefore) {
2704 assert(InsertBefore && "Must have someplace to insert!");
2705 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2707 Value *To = UndefValue::get(IndexedType);
2708 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2709 unsigned IdxSkip = Idxs.size();
2711 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2714 /// Given an aggregrate and an sequence of indices, see if
2715 /// the scalar value indexed is already around as a register, for example if it
2716 /// were inserted directly into the aggregrate.
2718 /// If InsertBefore is not null, this function will duplicate (modified)
2719 /// insertvalues when a part of a nested struct is extracted.
2720 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2721 Instruction *InsertBefore) {
2722 // Nothing to index? Just return V then (this is useful at the end of our
2724 if (idx_range.empty())
2726 // We have indices, so V should have an indexable type.
2727 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2728 "Not looking at a struct or array?");
2729 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2730 "Invalid indices for type?");
2732 if (Constant *C = dyn_cast<Constant>(V)) {
2733 C = C->getAggregateElement(idx_range[0]);
2734 if (!C) return nullptr;
2735 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2738 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2739 // Loop the indices for the insertvalue instruction in parallel with the
2740 // requested indices
2741 const unsigned *req_idx = idx_range.begin();
2742 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2743 i != e; ++i, ++req_idx) {
2744 if (req_idx == idx_range.end()) {
2745 // We can't handle this without inserting insertvalues
2749 // The requested index identifies a part of a nested aggregate. Handle
2750 // this specially. For example,
2751 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2752 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2753 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2754 // This can be changed into
2755 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2756 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2757 // which allows the unused 0,0 element from the nested struct to be
2759 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2763 // This insert value inserts something else than what we are looking for.
2764 // See if the (aggregate) value inserted into has the value we are
2765 // looking for, then.
2767 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2770 // If we end up here, the indices of the insertvalue match with those
2771 // requested (though possibly only partially). Now we recursively look at
2772 // the inserted value, passing any remaining indices.
2773 return FindInsertedValue(I->getInsertedValueOperand(),
2774 makeArrayRef(req_idx, idx_range.end()),
2778 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2779 // If we're extracting a value from an aggregate that was extracted from
2780 // something else, we can extract from that something else directly instead.
2781 // However, we will need to chain I's indices with the requested indices.
2783 // Calculate the number of indices required
2784 unsigned size = I->getNumIndices() + idx_range.size();
2785 // Allocate some space to put the new indices in
2786 SmallVector<unsigned, 5> Idxs;
2788 // Add indices from the extract value instruction
2789 Idxs.append(I->idx_begin(), I->idx_end());
2791 // Add requested indices
2792 Idxs.append(idx_range.begin(), idx_range.end());
2794 assert(Idxs.size() == size
2795 && "Number of indices added not correct?");
2797 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2799 // Otherwise, we don't know (such as, extracting from a function return value
2800 // or load instruction)
2804 /// Analyze the specified pointer to see if it can be expressed as a base
2805 /// pointer plus a constant offset. Return the base and offset to the caller.
2806 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2807 const DataLayout &DL) {
2808 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2809 APInt ByteOffset(BitWidth, 0);
2811 if (Ptr->getType()->isVectorTy())
2814 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2815 APInt GEPOffset(BitWidth, 0);
2816 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2819 ByteOffset += GEPOffset;
2821 Ptr = GEP->getPointerOperand();
2822 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2823 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2824 Ptr = cast<Operator>(Ptr)->getOperand(0);
2825 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2826 if (GA->mayBeOverridden())
2828 Ptr = GA->getAliasee();
2833 Offset = ByteOffset.getSExtValue();
2838 /// This function computes the length of a null-terminated C string pointed to
2839 /// by V. If successful, it returns true and returns the string in Str.
2840 /// If unsuccessful, it returns false.
2841 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2842 uint64_t Offset, bool TrimAtNul) {
2845 // Look through bitcast instructions and geps.
2846 V = V->stripPointerCasts();
2848 // If the value is a GEP instruction or constant expression, treat it as an
2850 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2851 // Make sure the GEP has exactly three arguments.
2852 if (GEP->getNumOperands() != 3)
2855 // Make sure the index-ee is a pointer to array of i8.
2856 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2857 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2858 if (!AT || !AT->getElementType()->isIntegerTy(8))
2861 // Check to make sure that the first operand of the GEP is an integer and
2862 // has value 0 so that we are sure we're indexing into the initializer.
2863 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2864 if (!FirstIdx || !FirstIdx->isZero())
2867 // If the second index isn't a ConstantInt, then this is a variable index
2868 // into the array. If this occurs, we can't say anything meaningful about
2870 uint64_t StartIdx = 0;
2871 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2872 StartIdx = CI->getZExtValue();
2875 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2879 // The GEP instruction, constant or instruction, must reference a global
2880 // variable that is a constant and is initialized. The referenced constant
2881 // initializer is the array that we'll use for optimization.
2882 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2883 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2886 // Handle the all-zeros case
2887 if (GV->getInitializer()->isNullValue()) {
2888 // This is a degenerate case. The initializer is constant zero so the
2889 // length of the string must be zero.
2894 // Must be a Constant Array
2895 const ConstantDataArray *Array =
2896 dyn_cast<ConstantDataArray>(GV->getInitializer());
2897 if (!Array || !Array->isString())
2900 // Get the number of elements in the array
2901 uint64_t NumElts = Array->getType()->getArrayNumElements();
2903 // Start out with the entire array in the StringRef.
2904 Str = Array->getAsString();
2906 if (Offset > NumElts)
2909 // Skip over 'offset' bytes.
2910 Str = Str.substr(Offset);
2913 // Trim off the \0 and anything after it. If the array is not nul
2914 // terminated, we just return the whole end of string. The client may know
2915 // some other way that the string is length-bound.
2916 Str = Str.substr(0, Str.find('\0'));
2921 // These next two are very similar to the above, but also look through PHI
2923 // TODO: See if we can integrate these two together.
2925 /// If we can compute the length of the string pointed to by
2926 /// the specified pointer, return 'len+1'. If we can't, return 0.
2927 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2928 // Look through noop bitcast instructions.
2929 V = V->stripPointerCasts();
2931 // If this is a PHI node, there are two cases: either we have already seen it
2933 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2934 if (!PHIs.insert(PN).second)
2935 return ~0ULL; // already in the set.
2937 // If it was new, see if all the input strings are the same length.
2938 uint64_t LenSoFar = ~0ULL;
2939 for (Value *IncValue : PN->incoming_values()) {
2940 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2941 if (Len == 0) return 0; // Unknown length -> unknown.
2943 if (Len == ~0ULL) continue;
2945 if (Len != LenSoFar && LenSoFar != ~0ULL)
2946 return 0; // Disagree -> unknown.
2950 // Success, all agree.
2954 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2955 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2956 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2957 if (Len1 == 0) return 0;
2958 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2959 if (Len2 == 0) return 0;
2960 if (Len1 == ~0ULL) return Len2;
2961 if (Len2 == ~0ULL) return Len1;
2962 if (Len1 != Len2) return 0;
2966 // Otherwise, see if we can read the string.
2968 if (!getConstantStringInfo(V, StrData))
2971 return StrData.size()+1;
2974 /// If we can compute the length of the string pointed to by
2975 /// the specified pointer, return 'len+1'. If we can't, return 0.
2976 uint64_t llvm::GetStringLength(Value *V) {
2977 if (!V->getType()->isPointerTy()) return 0;
2979 SmallPtrSet<PHINode*, 32> PHIs;
2980 uint64_t Len = GetStringLengthH(V, PHIs);
2981 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2982 // an empty string as a length.
2983 return Len == ~0ULL ? 1 : Len;
2986 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
2987 /// previous iteration of the loop was referring to the same object as \p PN.
2988 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
2989 // Find the loop-defined value.
2990 Loop *L = LI->getLoopFor(PN->getParent());
2991 if (PN->getNumIncomingValues() != 2)
2994 // Find the value from previous iteration.
2995 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
2996 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
2997 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
2998 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3001 // If a new pointer is loaded in the loop, the pointer references a different
3002 // object in every iteration. E.g.:
3006 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3007 if (!L->isLoopInvariant(Load->getPointerOperand()))
3012 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3013 unsigned MaxLookup) {
3014 if (!V->getType()->isPointerTy())
3016 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3017 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3018 V = GEP->getPointerOperand();
3019 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3020 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3021 V = cast<Operator>(V)->getOperand(0);
3022 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3023 if (GA->mayBeOverridden())
3025 V = GA->getAliasee();
3027 // See if InstructionSimplify knows any relevant tricks.
3028 if (Instruction *I = dyn_cast<Instruction>(V))
3029 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3030 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
3037 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3042 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3043 const DataLayout &DL, LoopInfo *LI,
3044 unsigned MaxLookup) {
3045 SmallPtrSet<Value *, 4> Visited;
3046 SmallVector<Value *, 4> Worklist;
3047 Worklist.push_back(V);
3049 Value *P = Worklist.pop_back_val();
3050 P = GetUnderlyingObject(P, DL, MaxLookup);
3052 if (!Visited.insert(P).second)
3055 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3056 Worklist.push_back(SI->getTrueValue());
3057 Worklist.push_back(SI->getFalseValue());
3061 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3062 // If this PHI changes the underlying object in every iteration of the
3063 // loop, don't look through it. Consider:
3066 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3070 // Prev is tracking Curr one iteration behind so they refer to different
3071 // underlying objects.
3072 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3073 isSameUnderlyingObjectInLoop(PN, LI))
3074 for (Value *IncValue : PN->incoming_values())
3075 Worklist.push_back(IncValue);
3079 Objects.push_back(P);
3080 } while (!Worklist.empty());
3083 /// Return true if the only users of this pointer are lifetime markers.
3084 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3085 for (const User *U : V->users()) {
3086 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3087 if (!II) return false;
3089 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3090 II->getIntrinsicID() != Intrinsic::lifetime_end)
3096 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
3097 Type *Ty, const DataLayout &DL,
3098 const Instruction *CtxI,
3099 const DominatorTree *DT,
3100 const TargetLibraryInfo *TLI) {
3101 assert(Offset.isNonNegative() && "offset can't be negative");
3102 assert(Ty->isSized() && "must be sized");
3104 APInt DerefBytes(Offset.getBitWidth(), 0);
3105 bool CheckForNonNull = false;
3106 if (const Argument *A = dyn_cast<Argument>(BV)) {
3107 DerefBytes = A->getDereferenceableBytes();
3108 if (!DerefBytes.getBoolValue()) {
3109 DerefBytes = A->getDereferenceableOrNullBytes();
3110 CheckForNonNull = true;
3112 } else if (auto CS = ImmutableCallSite(BV)) {
3113 DerefBytes = CS.getDereferenceableBytes(0);
3114 if (!DerefBytes.getBoolValue()) {
3115 DerefBytes = CS.getDereferenceableOrNullBytes(0);
3116 CheckForNonNull = true;
3118 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
3119 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
3120 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3121 DerefBytes = CI->getLimitedValue();
3123 if (!DerefBytes.getBoolValue()) {
3125 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
3126 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3127 DerefBytes = CI->getLimitedValue();
3129 CheckForNonNull = true;
3133 if (DerefBytes.getBoolValue())
3134 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
3135 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
3141 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
3142 const Instruction *CtxI,
3143 const DominatorTree *DT,
3144 const TargetLibraryInfo *TLI) {
3145 Type *VTy = V->getType();
3146 Type *Ty = VTy->getPointerElementType();
3150 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3151 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
3154 static bool isAligned(const Value *Base, APInt Offset, unsigned Align,
3155 const DataLayout &DL) {
3156 APInt BaseAlign(Offset.getBitWidth(), getAlignment(Base, DL));
3159 Type *Ty = Base->getType()->getPointerElementType();
3160 BaseAlign = DL.getABITypeAlignment(Ty);
3163 APInt Alignment(Offset.getBitWidth(), Align);
3165 assert(Alignment.isPowerOf2() && "must be a power of 2!");
3166 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1));
3169 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) {
3170 APInt Offset(DL.getTypeStoreSizeInBits(Base->getType()), 0);
3171 return isAligned(Base, Offset, Align, DL);
3174 /// Test if V is always a pointer to allocated and suitably aligned memory for
3175 /// a simple load or store.
3176 static bool isDereferenceableAndAlignedPointer(
3177 const Value *V, unsigned Align, const DataLayout &DL,
3178 const Instruction *CtxI, const DominatorTree *DT,
3179 const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited) {
3180 // Note that it is not safe to speculate into a malloc'd region because
3181 // malloc may return null.
3183 // These are obviously ok if aligned.
3184 if (isa<AllocaInst>(V))
3185 return isAligned(V, Align, DL);
3187 // It's not always safe to follow a bitcast, for example:
3188 // bitcast i8* (alloca i8) to i32*
3189 // would result in a 4-byte load from a 1-byte alloca. However,
3190 // if we're casting from a pointer from a type of larger size
3191 // to a type of smaller size (or the same size), and the alignment
3192 // is at least as large as for the resulting pointer type, then
3193 // we can look through the bitcast.
3194 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
3195 Type *STy = BC->getSrcTy()->getPointerElementType(),
3196 *DTy = BC->getDestTy()->getPointerElementType();
3197 if (STy->isSized() && DTy->isSized() &&
3198 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
3199 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
3200 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, DL,
3201 CtxI, DT, TLI, Visited);
3204 // Global variables which can't collapse to null are ok.
3205 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
3206 if (!GV->hasExternalWeakLinkage())
3207 return isAligned(V, Align, DL);
3209 // byval arguments are okay.
3210 if (const Argument *A = dyn_cast<Argument>(V))
3211 if (A->hasByValAttr())
3212 return isAligned(V, Align, DL);
3214 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
3215 return isAligned(V, Align, DL);
3217 // For GEPs, determine if the indexing lands within the allocated object.
3218 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3219 Type *VTy = GEP->getType();
3220 Type *Ty = VTy->getPointerElementType();
3221 const Value *Base = GEP->getPointerOperand();
3223 // Conservatively require that the base pointer be fully dereferenceable
3225 if (!Visited.insert(Base).second)
3227 if (!isDereferenceableAndAlignedPointer(Base, Align, DL, CtxI, DT, TLI,
3231 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
3232 if (!GEP->accumulateConstantOffset(DL, Offset))
3235 // Check if the load is within the bounds of the underlying object
3236 // and offset is aligned.
3237 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3238 Type *BaseType = Base->getType()->getPointerElementType();
3239 assert(isPowerOf2_32(Align) && "must be a power of 2!");
3240 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType)) &&
3241 !(Offset & APInt(Offset.getBitWidth(), Align-1));
3244 // For gc.relocate, look through relocations
3245 if (const IntrinsicInst *I = dyn_cast<IntrinsicInst>(V))
3246 if (I->getIntrinsicID() == Intrinsic::experimental_gc_relocate) {
3247 GCRelocateOperands RelocateInst(I);
3248 return isDereferenceableAndAlignedPointer(
3249 RelocateInst.getDerivedPtr(), Align, DL, CtxI, DT, TLI, Visited);
3252 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3253 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, DL,
3254 CtxI, DT, TLI, Visited);
3256 // If we don't know, assume the worst.
3260 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align,
3261 const DataLayout &DL,
3262 const Instruction *CtxI,
3263 const DominatorTree *DT,
3264 const TargetLibraryInfo *TLI) {
3265 // When dereferenceability information is provided by a dereferenceable
3266 // attribute, we know exactly how many bytes are dereferenceable. If we can
3267 // determine the exact offset to the attributed variable, we can use that
3268 // information here.
3269 Type *VTy = V->getType();
3270 Type *Ty = VTy->getPointerElementType();
3272 // Require ABI alignment for loads without alignment specification
3274 Align = DL.getABITypeAlignment(Ty);
3276 if (Ty->isSized()) {
3277 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3278 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3280 if (Offset.isNonNegative())
3281 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL, CtxI, DT, TLI) &&
3282 isAligned(BV, Offset, Align, DL))
3286 SmallPtrSet<const Value *, 32> Visited;
3287 return ::isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT, TLI,
3291 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3292 const Instruction *CtxI,
3293 const DominatorTree *DT,
3294 const TargetLibraryInfo *TLI) {
3295 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT, TLI);
3298 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3299 const Instruction *CtxI,
3300 const DominatorTree *DT,
3301 const TargetLibraryInfo *TLI) {
3302 const Operator *Inst = dyn_cast<Operator>(V);
3306 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3307 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3311 switch (Inst->getOpcode()) {
3314 case Instruction::UDiv:
3315 case Instruction::URem: {
3316 // x / y is undefined if y == 0.
3318 if (match(Inst->getOperand(1), m_APInt(V)))
3322 case Instruction::SDiv:
3323 case Instruction::SRem: {
3324 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3325 const APInt *Numerator, *Denominator;
3326 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3328 // We cannot hoist this division if the denominator is 0.
3329 if (*Denominator == 0)
3331 // It's safe to hoist if the denominator is not 0 or -1.
3332 if (*Denominator != -1)
3334 // At this point we know that the denominator is -1. It is safe to hoist as
3335 // long we know that the numerator is not INT_MIN.
3336 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3337 return !Numerator->isMinSignedValue();
3338 // The numerator *might* be MinSignedValue.
3341 case Instruction::Load: {
3342 const LoadInst *LI = cast<LoadInst>(Inst);
3343 if (!LI->isUnordered() ||
3344 // Speculative load may create a race that did not exist in the source.
3345 LI->getParent()->getParent()->hasFnAttribute(
3346 Attribute::SanitizeThread) ||
3347 // Speculative load may load data from dirty regions.
3348 LI->getParent()->getParent()->hasFnAttribute(
3349 Attribute::SanitizeAddress))
3351 const DataLayout &DL = LI->getModule()->getDataLayout();
3352 return isDereferenceableAndAlignedPointer(
3353 LI->getPointerOperand(), LI->getAlignment(), DL, CtxI, DT, TLI);
3355 case Instruction::Call: {
3356 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3357 switch (II->getIntrinsicID()) {
3358 // These synthetic intrinsics have no side-effects and just mark
3359 // information about their operands.
3360 // FIXME: There are other no-op synthetic instructions that potentially
3361 // should be considered at least *safe* to speculate...
3362 case Intrinsic::dbg_declare:
3363 case Intrinsic::dbg_value:
3366 case Intrinsic::bswap:
3367 case Intrinsic::ctlz:
3368 case Intrinsic::ctpop:
3369 case Intrinsic::cttz:
3370 case Intrinsic::objectsize:
3371 case Intrinsic::sadd_with_overflow:
3372 case Intrinsic::smul_with_overflow:
3373 case Intrinsic::ssub_with_overflow:
3374 case Intrinsic::uadd_with_overflow:
3375 case Intrinsic::umul_with_overflow:
3376 case Intrinsic::usub_with_overflow:
3378 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3379 // errno like libm sqrt would.
3380 case Intrinsic::sqrt:
3381 case Intrinsic::fma:
3382 case Intrinsic::fmuladd:
3383 case Intrinsic::fabs:
3384 case Intrinsic::minnum:
3385 case Intrinsic::maxnum:
3387 // TODO: some fp intrinsics are marked as having the same error handling
3388 // as libm. They're safe to speculate when they won't error.
3389 // TODO: are convert_{from,to}_fp16 safe?
3390 // TODO: can we list target-specific intrinsics here?
3394 return false; // The called function could have undefined behavior or
3395 // side-effects, even if marked readnone nounwind.
3397 case Instruction::VAArg:
3398 case Instruction::Alloca:
3399 case Instruction::Invoke:
3400 case Instruction::PHI:
3401 case Instruction::Store:
3402 case Instruction::Ret:
3403 case Instruction::Br:
3404 case Instruction::IndirectBr:
3405 case Instruction::Switch:
3406 case Instruction::Unreachable:
3407 case Instruction::Fence:
3408 case Instruction::AtomicRMW:
3409 case Instruction::AtomicCmpXchg:
3410 case Instruction::LandingPad:
3411 case Instruction::Resume:
3412 case Instruction::CatchPad:
3413 case Instruction::CatchEndPad:
3414 case Instruction::CatchRet:
3415 case Instruction::CleanupPad:
3416 case Instruction::CleanupEndPad:
3417 case Instruction::CleanupRet:
3418 case Instruction::TerminatePad:
3419 return false; // Misc instructions which have effects
3423 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3424 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3427 /// Return true if we know that the specified value is never null.
3428 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3429 assert(V->getType()->isPointerTy() && "V must be pointer type");
3431 // Alloca never returns null, malloc might.
3432 if (isa<AllocaInst>(V)) return true;
3434 // A byval, inalloca, or nonnull argument is never null.
3435 if (const Argument *A = dyn_cast<Argument>(V))
3436 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3438 // A global variable in address space 0 is non null unless extern weak.
3439 // Other address spaces may have null as a valid address for a global,
3440 // so we can't assume anything.
3441 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3442 return !GV->hasExternalWeakLinkage() &&
3443 GV->getType()->getAddressSpace() == 0;
3445 // A Load tagged w/nonnull metadata is never null.
3446 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3447 return LI->getMetadata(LLVMContext::MD_nonnull);
3449 if (auto CS = ImmutableCallSite(V))
3450 if (CS.isReturnNonNull())
3453 // operator new never returns null.
3454 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
3460 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3461 const Instruction *CtxI,
3462 const DominatorTree *DT) {
3463 assert(V->getType()->isPointerTy() && "V must be pointer type");
3465 unsigned NumUsesExplored = 0;
3466 for (auto U : V->users()) {
3467 // Avoid massive lists
3468 if (NumUsesExplored >= DomConditionsMaxUses)
3471 // Consider only compare instructions uniquely controlling a branch
3472 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3476 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3479 for (auto *CmpU : Cmp->users()) {
3480 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3484 assert(BI->isConditional() && "uses a comparison!");
3486 BasicBlock *NonNullSuccessor = nullptr;
3487 CmpInst::Predicate Pred;
3489 if (match(const_cast<ICmpInst*>(Cmp),
3490 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3491 if (Pred == ICmpInst::ICMP_EQ)
3492 NonNullSuccessor = BI->getSuccessor(1);
3493 else if (Pred == ICmpInst::ICMP_NE)
3494 NonNullSuccessor = BI->getSuccessor(0);
3497 if (NonNullSuccessor) {
3498 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3499 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3508 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3509 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3510 if (isKnownNonNull(V, TLI))
3513 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3516 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3517 const DataLayout &DL,
3518 AssumptionCache *AC,
3519 const Instruction *CxtI,
3520 const DominatorTree *DT) {
3521 // Multiplying n * m significant bits yields a result of n + m significant
3522 // bits. If the total number of significant bits does not exceed the
3523 // result bit width (minus 1), there is no overflow.
3524 // This means if we have enough leading zero bits in the operands
3525 // we can guarantee that the result does not overflow.
3526 // Ref: "Hacker's Delight" by Henry Warren
3527 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3528 APInt LHSKnownZero(BitWidth, 0);
3529 APInt LHSKnownOne(BitWidth, 0);
3530 APInt RHSKnownZero(BitWidth, 0);
3531 APInt RHSKnownOne(BitWidth, 0);
3532 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3534 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3536 // Note that underestimating the number of zero bits gives a more
3537 // conservative answer.
3538 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3539 RHSKnownZero.countLeadingOnes();
3540 // First handle the easy case: if we have enough zero bits there's
3541 // definitely no overflow.
3542 if (ZeroBits >= BitWidth)
3543 return OverflowResult::NeverOverflows;
3545 // Get the largest possible values for each operand.
3546 APInt LHSMax = ~LHSKnownZero;
3547 APInt RHSMax = ~RHSKnownZero;
3549 // We know the multiply operation doesn't overflow if the maximum values for
3550 // each operand will not overflow after we multiply them together.
3552 LHSMax.umul_ov(RHSMax, MaxOverflow);
3554 return OverflowResult::NeverOverflows;
3556 // We know it always overflows if multiplying the smallest possible values for
3557 // the operands also results in overflow.
3559 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3561 return OverflowResult::AlwaysOverflows;
3563 return OverflowResult::MayOverflow;
3566 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3567 const DataLayout &DL,
3568 AssumptionCache *AC,
3569 const Instruction *CxtI,
3570 const DominatorTree *DT) {
3571 bool LHSKnownNonNegative, LHSKnownNegative;
3572 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3574 if (LHSKnownNonNegative || LHSKnownNegative) {
3575 bool RHSKnownNonNegative, RHSKnownNegative;
3576 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3579 if (LHSKnownNegative && RHSKnownNegative) {
3580 // The sign bit is set in both cases: this MUST overflow.
3581 // Create a simple add instruction, and insert it into the struct.
3582 return OverflowResult::AlwaysOverflows;
3585 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3586 // The sign bit is clear in both cases: this CANNOT overflow.
3587 // Create a simple add instruction, and insert it into the struct.
3588 return OverflowResult::NeverOverflows;
3592 return OverflowResult::MayOverflow;
3595 static OverflowResult computeOverflowForSignedAdd(
3596 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3597 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3598 if (Add && Add->hasNoSignedWrap()) {
3599 return OverflowResult::NeverOverflows;
3602 bool LHSKnownNonNegative, LHSKnownNegative;
3603 bool RHSKnownNonNegative, RHSKnownNegative;
3604 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3606 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3609 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3610 (LHSKnownNegative && RHSKnownNonNegative)) {
3611 // The sign bits are opposite: this CANNOT overflow.
3612 return OverflowResult::NeverOverflows;
3615 // The remaining code needs Add to be available. Early returns if not so.
3617 return OverflowResult::MayOverflow;
3619 // If the sign of Add is the same as at least one of the operands, this add
3620 // CANNOT overflow. This is particularly useful when the sum is
3621 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3623 bool LHSOrRHSKnownNonNegative =
3624 (LHSKnownNonNegative || RHSKnownNonNegative);
3625 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3626 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3627 bool AddKnownNonNegative, AddKnownNegative;
3628 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3629 /*Depth=*/0, AC, CxtI, DT);
3630 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3631 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3632 return OverflowResult::NeverOverflows;
3636 return OverflowResult::MayOverflow;
3639 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3640 const DataLayout &DL,
3641 AssumptionCache *AC,
3642 const Instruction *CxtI,
3643 const DominatorTree *DT) {
3644 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3645 Add, DL, AC, CxtI, DT);
3648 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3649 const DataLayout &DL,
3650 AssumptionCache *AC,
3651 const Instruction *CxtI,
3652 const DominatorTree *DT) {
3653 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3656 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3657 // FIXME: This conservative implementation can be relaxed. E.g. most
3658 // atomic operations are guaranteed to terminate on most platforms
3659 // and most functions terminate.
3661 return !I->isAtomic() && // atomics may never succeed on some platforms
3662 !isa<CallInst>(I) && // could throw and might not terminate
3663 !isa<InvokeInst>(I) && // might not terminate and could throw to
3664 // non-successor (see bug 24185 for details).
3665 !isa<ResumeInst>(I) && // has no successors
3666 !isa<ReturnInst>(I); // has no successors
3669 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3671 // The loop header is guaranteed to be executed for every iteration.
3673 // FIXME: Relax this constraint to cover all basic blocks that are
3674 // guaranteed to be executed at every iteration.
3675 if (I->getParent() != L->getHeader()) return false;
3677 for (const Instruction &LI : *L->getHeader()) {
3678 if (&LI == I) return true;
3679 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3681 llvm_unreachable("Instruction not contained in its own parent basic block.");
3684 bool llvm::propagatesFullPoison(const Instruction *I) {
3685 switch (I->getOpcode()) {
3686 case Instruction::Add:
3687 case Instruction::Sub:
3688 case Instruction::Xor:
3689 case Instruction::Trunc:
3690 case Instruction::BitCast:
3691 case Instruction::AddrSpaceCast:
3692 // These operations all propagate poison unconditionally. Note that poison
3693 // is not any particular value, so xor or subtraction of poison with
3694 // itself still yields poison, not zero.
3697 case Instruction::AShr:
3698 case Instruction::SExt:
3699 // For these operations, one bit of the input is replicated across
3700 // multiple output bits. A replicated poison bit is still poison.
3703 case Instruction::Shl: {
3704 // Left shift *by* a poison value is poison. The number of
3705 // positions to shift is unsigned, so no negative values are
3706 // possible there. Left shift by zero places preserves poison. So
3707 // it only remains to consider left shift of poison by a positive
3708 // number of places.
3710 // A left shift by a positive number of places leaves the lowest order bit
3711 // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3712 // make the poison operand violate that flag, yielding a fresh full-poison
3714 auto *OBO = cast<OverflowingBinaryOperator>(I);
3715 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3718 case Instruction::Mul: {
3719 // A multiplication by zero yields a non-poison zero result, so we need to
3720 // rule out zero as an operand. Conservatively, multiplication by a
3721 // non-zero constant is not multiplication by zero.
3723 // Multiplication by a non-zero constant can leave some bits
3724 // non-poisoned. For example, a multiplication by 2 leaves the lowest
3725 // order bit unpoisoned. So we need to consider that.
3727 // Multiplication by 1 preserves poison. If the multiplication has a
3728 // no-wrap flag, then we can make the poison operand violate that flag
3729 // when multiplied by any integer other than 0 and 1.
3730 auto *OBO = cast<OverflowingBinaryOperator>(I);
3731 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3732 for (Value *V : OBO->operands()) {
3733 if (auto *CI = dyn_cast<ConstantInt>(V)) {
3734 // A ConstantInt cannot yield poison, so we can assume that it is
3735 // the other operand that is poison.
3736 return !CI->isZero();
3743 case Instruction::GetElementPtr:
3744 // A GEP implicitly represents a sequence of additions, subtractions,
3745 // truncations, sign extensions and multiplications. The multiplications
3746 // are by the non-zero sizes of some set of types, so we do not have to be
3747 // concerned with multiplication by zero. If the GEP is in-bounds, then
3748 // these operations are implicitly no-signed-wrap so poison is propagated
3749 // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3750 return cast<GEPOperator>(I)->isInBounds();
3757 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3758 switch (I->getOpcode()) {
3759 case Instruction::Store:
3760 return cast<StoreInst>(I)->getPointerOperand();
3762 case Instruction::Load:
3763 return cast<LoadInst>(I)->getPointerOperand();
3765 case Instruction::AtomicCmpXchg:
3766 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3768 case Instruction::AtomicRMW:
3769 return cast<AtomicRMWInst>(I)->getPointerOperand();
3771 case Instruction::UDiv:
3772 case Instruction::SDiv:
3773 case Instruction::URem:
3774 case Instruction::SRem:
3775 return I->getOperand(1);
3782 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3783 // We currently only look for uses of poison values within the same basic
3784 // block, as that makes it easier to guarantee that the uses will be
3785 // executed given that PoisonI is executed.
3787 // FIXME: Expand this to consider uses beyond the same basic block. To do
3788 // this, look out for the distinction between post-dominance and strong
3790 const BasicBlock *BB = PoisonI->getParent();
3792 // Set of instructions that we have proved will yield poison if PoisonI
3794 SmallSet<const Value *, 16> YieldsPoison;
3795 YieldsPoison.insert(PoisonI);
3797 for (BasicBlock::const_iterator I = PoisonI->getIterator(), E = BB->end();
3799 if (&*I != PoisonI) {
3800 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&*I);
3801 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) return true;
3802 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
3806 // Mark poison that propagates from I through uses of I.
3807 if (YieldsPoison.count(&*I)) {
3808 for (const User *User : I->users()) {
3809 const Instruction *UserI = cast<Instruction>(User);
3810 if (UserI->getParent() == BB && propagatesFullPoison(UserI))
3811 YieldsPoison.insert(User);
3818 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3822 if (auto *C = dyn_cast<ConstantFP>(V))
3827 static bool isKnownNonZero(Value *V) {
3828 if (auto *C = dyn_cast<ConstantFP>(V))
3829 return !C->isZero();
3833 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3835 Value *CmpLHS, Value *CmpRHS,
3836 Value *TrueVal, Value *FalseVal,
3837 Value *&LHS, Value *&RHS) {
3841 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3842 // return inconsistent results between implementations.
3843 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3844 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3845 // Therefore we behave conservatively and only proceed if at least one of the
3846 // operands is known to not be zero, or if we don't care about signed zeroes.
3849 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3850 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3851 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3852 !isKnownNonZero(CmpRHS))
3853 return {SPF_UNKNOWN, SPNB_NA, false};
3856 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3857 bool Ordered = false;
3859 // When given one NaN and one non-NaN input:
3860 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3861 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3862 // ordered comparison fails), which could be NaN or non-NaN.
3863 // so here we discover exactly what NaN behavior is required/accepted.
3864 if (CmpInst::isFPPredicate(Pred)) {
3865 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3866 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3868 if (LHSSafe && RHSSafe) {
3869 // Both operands are known non-NaN.
3870 NaNBehavior = SPNB_RETURNS_ANY;
3871 } else if (CmpInst::isOrdered(Pred)) {
3872 // An ordered comparison will return false when given a NaN, so it
3876 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3877 NaNBehavior = SPNB_RETURNS_NAN;
3879 NaNBehavior = SPNB_RETURNS_OTHER;
3881 // Completely unsafe.
3882 return {SPF_UNKNOWN, SPNB_NA, false};
3885 // An unordered comparison will return true when given a NaN, so it
3888 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3889 NaNBehavior = SPNB_RETURNS_OTHER;
3891 NaNBehavior = SPNB_RETURNS_NAN;
3893 // Completely unsafe.
3894 return {SPF_UNKNOWN, SPNB_NA, false};
3898 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3899 std::swap(CmpLHS, CmpRHS);
3900 Pred = CmpInst::getSwappedPredicate(Pred);
3901 if (NaNBehavior == SPNB_RETURNS_NAN)
3902 NaNBehavior = SPNB_RETURNS_OTHER;
3903 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3904 NaNBehavior = SPNB_RETURNS_NAN;
3908 // ([if]cmp X, Y) ? X : Y
3909 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3911 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3912 case ICmpInst::ICMP_UGT:
3913 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3914 case ICmpInst::ICMP_SGT:
3915 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3916 case ICmpInst::ICMP_ULT:
3917 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3918 case ICmpInst::ICMP_SLT:
3919 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3920 case FCmpInst::FCMP_UGT:
3921 case FCmpInst::FCMP_UGE:
3922 case FCmpInst::FCMP_OGT:
3923 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3924 case FCmpInst::FCMP_ULT:
3925 case FCmpInst::FCMP_ULE:
3926 case FCmpInst::FCMP_OLT:
3927 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3931 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3932 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3933 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3935 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3936 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3937 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3938 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3941 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3942 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3943 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3944 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3948 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3949 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3950 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3951 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3952 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3955 return {SPF_SMIN, SPNB_NA, false};
3960 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3962 return {SPF_UNKNOWN, SPNB_NA, false};
3965 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
3966 Instruction::CastOps *CastOp) {
3967 CastInst *CI = dyn_cast<CastInst>(V1);
3968 Constant *C = dyn_cast<Constant>(V2);
3969 CastInst *CI2 = dyn_cast<CastInst>(V2);
3972 *CastOp = CI->getOpcode();
3975 // If V1 and V2 are both the same cast from the same type, we can look
3977 if (CI2->getOpcode() == CI->getOpcode() &&
3978 CI2->getSrcTy() == CI->getSrcTy())
3979 return CI2->getOperand(0);
3985 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
3986 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
3987 // This is only valid if the truncated value can be sign-extended
3988 // back to the original value.
3989 if (ConstantExpr::getSExt(T, C->getType()) == C)
3993 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
3994 return ConstantExpr::getTrunc(C, CI->getSrcTy());
3996 if (isa<TruncInst>(CI))
3997 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
3999 if (isa<FPToUIInst>(CI))
4000 return ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
4002 if (isa<FPToSIInst>(CI))
4003 return ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
4005 if (isa<UIToFPInst>(CI))
4006 return ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
4008 if (isa<SIToFPInst>(CI))
4009 return ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
4011 if (isa<FPTruncInst>(CI))
4012 return ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
4014 if (isa<FPExtInst>(CI))
4015 return ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
4020 SelectPatternResult llvm::matchSelectPattern(Value *V,
4021 Value *&LHS, Value *&RHS,
4022 Instruction::CastOps *CastOp) {
4023 SelectInst *SI = dyn_cast<SelectInst>(V);
4024 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4026 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4027 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4029 CmpInst::Predicate Pred = CmpI->getPredicate();
4030 Value *CmpLHS = CmpI->getOperand(0);
4031 Value *CmpRHS = CmpI->getOperand(1);
4032 Value *TrueVal = SI->getTrueValue();
4033 Value *FalseVal = SI->getFalseValue();
4035 if (isa<FPMathOperator>(CmpI))
4036 FMF = CmpI->getFastMathFlags();
4039 if (CmpI->isEquality())
4040 return {SPF_UNKNOWN, SPNB_NA, false};
4042 // Deal with type mismatches.
4043 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4044 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
4045 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4046 cast<CastInst>(TrueVal)->getOperand(0), C,
4048 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
4049 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4050 C, cast<CastInst>(FalseVal)->getOperand(0),
4053 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4057 ConstantRange llvm::getConstantRangeFromMetadata(MDNode &Ranges) {
4058 const unsigned NumRanges = Ranges.getNumOperands() / 2;
4059 assert(NumRanges >= 1 && "Must have at least one range!");
4060 assert(Ranges.getNumOperands() % 2 == 0 && "Must be a sequence of pairs");
4062 auto *FirstLow = mdconst::extract<ConstantInt>(Ranges.getOperand(0));
4063 auto *FirstHigh = mdconst::extract<ConstantInt>(Ranges.getOperand(1));
4065 ConstantRange CR(FirstLow->getValue(), FirstHigh->getValue());
4067 for (unsigned i = 1; i < NumRanges; ++i) {
4068 auto *Low = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
4069 auto *High = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
4071 // Note: unionWith will potentially create a range that contains values not
4072 // contained in any of the original N ranges.
4073 CR = CR.unionWith(ConstantRange(Low->getValue(), High->getValue()));