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,
372 unsigned BitWidth = KnownZero.getBitWidth();
373 unsigned NumRanges = Ranges.getNumOperands() / 2;
374 assert(NumRanges >= 1);
376 KnownZero.setAllBits();
377 KnownOne.setAllBits();
379 for (unsigned i = 0; i < NumRanges; ++i) {
381 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
383 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
384 ConstantRange Range(Lower->getValue(), Upper->getValue());
386 // The first CommonPrefixBits of all values in Range are equal.
387 unsigned CommonPrefixBits =
388 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
390 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
391 KnownOne &= Range.getUnsignedMax() & Mask;
392 KnownZero &= ~Range.getUnsignedMax() & Mask;
396 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
397 SmallVector<const Value *, 16> WorkSet(1, I);
398 SmallPtrSet<const Value *, 32> Visited;
399 SmallPtrSet<const Value *, 16> EphValues;
401 // The instruction defining an assumption's condition itself is always
402 // considered ephemeral to that assumption (even if it has other
403 // non-ephemeral users). See r246696's test case for an example.
404 if (std::find(I->op_begin(), I->op_end(), E) != I->op_end())
407 while (!WorkSet.empty()) {
408 const Value *V = WorkSet.pop_back_val();
409 if (!Visited.insert(V).second)
412 // If all uses of this value are ephemeral, then so is this value.
413 if (std::all_of(V->user_begin(), V->user_end(),
414 [&](const User *U) { return EphValues.count(U); })) {
419 if (const User *U = dyn_cast<User>(V))
420 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
422 if (isSafeToSpeculativelyExecute(*J))
423 WorkSet.push_back(*J);
431 // Is this an intrinsic that cannot be speculated but also cannot trap?
432 static bool isAssumeLikeIntrinsic(const Instruction *I) {
433 if (const CallInst *CI = dyn_cast<CallInst>(I))
434 if (Function *F = CI->getCalledFunction())
435 switch (F->getIntrinsicID()) {
437 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
438 case Intrinsic::assume:
439 case Intrinsic::dbg_declare:
440 case Intrinsic::dbg_value:
441 case Intrinsic::invariant_start:
442 case Intrinsic::invariant_end:
443 case Intrinsic::lifetime_start:
444 case Intrinsic::lifetime_end:
445 case Intrinsic::objectsize:
446 case Intrinsic::ptr_annotation:
447 case Intrinsic::var_annotation:
454 static bool isValidAssumeForContext(Value *V, const Query &Q) {
455 Instruction *Inv = cast<Instruction>(V);
457 // There are two restrictions on the use of an assume:
458 // 1. The assume must dominate the context (or the control flow must
459 // reach the assume whenever it reaches the context).
460 // 2. The context must not be in the assume's set of ephemeral values
461 // (otherwise we will use the assume to prove that the condition
462 // feeding the assume is trivially true, thus causing the removal of
466 if (Q.DT->dominates(Inv, Q.CxtI)) {
468 } else if (Inv->getParent() == Q.CxtI->getParent()) {
469 // The context comes first, but they're both in the same block. Make sure
470 // there is nothing in between that might interrupt the control flow.
471 for (BasicBlock::const_iterator I =
472 std::next(BasicBlock::const_iterator(Q.CxtI)),
473 IE(Inv); I != IE; ++I)
474 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
477 return !isEphemeralValueOf(Inv, Q.CxtI);
483 // When we don't have a DT, we do a limited search...
484 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
486 } else if (Inv->getParent() == Q.CxtI->getParent()) {
487 // Search forward from the assume until we reach the context (or the end
488 // of the block); the common case is that the assume will come first.
489 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
490 IE = Inv->getParent()->end(); I != IE; ++I)
494 // The context must come first...
495 for (BasicBlock::const_iterator I =
496 std::next(BasicBlock::const_iterator(Q.CxtI)),
497 IE(Inv); I != IE; ++I)
498 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
501 return !isEphemeralValueOf(Inv, Q.CxtI);
507 bool llvm::isValidAssumeForContext(const Instruction *I,
508 const Instruction *CxtI,
509 const DominatorTree *DT) {
510 return ::isValidAssumeForContext(const_cast<Instruction *>(I),
511 Query(nullptr, CxtI, DT));
514 template<typename LHS, typename RHS>
515 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
516 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
517 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
518 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
521 template<typename LHS, typename RHS>
522 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
523 BinaryOp_match<RHS, LHS, Instruction::And>>
524 m_c_And(const LHS &L, const RHS &R) {
525 return m_CombineOr(m_And(L, R), m_And(R, L));
528 template<typename LHS, typename RHS>
529 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
530 BinaryOp_match<RHS, LHS, Instruction::Or>>
531 m_c_Or(const LHS &L, const RHS &R) {
532 return m_CombineOr(m_Or(L, R), m_Or(R, L));
535 template<typename LHS, typename RHS>
536 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
537 BinaryOp_match<RHS, LHS, Instruction::Xor>>
538 m_c_Xor(const LHS &L, const RHS &R) {
539 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
542 /// Compute known bits in 'V' under the assumption that the condition 'Cmp' is
543 /// true (at the context instruction.) This is mostly a utility function for
544 /// the prototype dominating conditions reasoning below.
545 static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp,
548 const DataLayout &DL,
549 unsigned Depth, const Query &Q) {
550 Value *LHS = Cmp->getOperand(0);
551 Value *RHS = Cmp->getOperand(1);
552 // TODO: We could potentially be more aggressive here. This would be worth
553 // evaluating. If we can, explore commoning this code with the assume
555 if (LHS != V && RHS != V)
558 const unsigned BitWidth = KnownZero.getBitWidth();
560 switch (Cmp->getPredicate()) {
562 // We know nothing from this condition
564 // TODO: implement unsigned bound from below (known one bits)
565 // TODO: common condition check implementations with assumes
566 // TODO: implement other patterns from assume (e.g. V & B == A)
567 case ICmpInst::ICMP_SGT:
569 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
570 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
571 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) {
572 // We know that the sign bit is zero.
573 KnownZero |= APInt::getSignBit(BitWidth);
577 case ICmpInst::ICMP_EQ:
579 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
581 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
583 computeKnownBits(LHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
585 llvm_unreachable("missing use?");
586 KnownZero |= KnownZeroTemp;
587 KnownOne |= KnownOneTemp;
590 case ICmpInst::ICMP_ULE:
592 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
593 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
594 // The known zero bits carry over
595 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
596 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
599 case ICmpInst::ICMP_ULT:
601 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0);
602 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q);
603 // Whatever high bits in rhs are zero are known to be zero (if rhs is a
604 // power of 2, then one more).
605 unsigned SignBits = KnownZeroTemp.countLeadingOnes();
606 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL))
608 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits);
614 /// Compute known bits in 'V' from conditions which are known to be true along
615 /// all paths leading to the context instruction. In particular, look for
616 /// cases where one branch of an interesting condition dominates the context
617 /// instruction. This does not do general dataflow.
618 /// NOTE: This code is EXPERIMENTAL and currently off by default.
619 static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero,
621 const DataLayout &DL,
624 // Need both the dominator tree and the query location to do anything useful
625 if (!Q.DT || !Q.CxtI)
627 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI);
628 // The context instruction might be in a statically unreachable block. If
629 // so, asking dominator queries may yield suprising results. (e.g. the block
630 // may not have a dom tree node)
631 if (!Q.DT->isReachableFromEntry(Cxt->getParent()))
634 // Avoid useless work
635 if (auto VI = dyn_cast<Instruction>(V))
636 if (VI->getParent() == Cxt->getParent())
639 // Note: We currently implement two options. It's not clear which of these
640 // will survive long term, we need data for that.
641 // Option 1 - Try walking the dominator tree looking for conditions which
642 // might apply. This works well for local conditions (loop guards, etc..),
643 // but not as well for things far from the context instruction (presuming a
644 // low max blocks explored). If we can set an high enough limit, this would
646 // Option 2 - We restrict out search to those conditions which are uses of
647 // the value we're interested in. This is independent of dom structure,
648 // but is slightly less powerful without looking through lots of use chains.
649 // It does handle conditions far from the context instruction (e.g. early
650 // function exits on entry) really well though.
652 // Option 1 - Search the dom tree
653 unsigned NumBlocksExplored = 0;
654 BasicBlock *Current = Cxt->getParent();
656 // Stop searching if we've gone too far up the chain
657 if (NumBlocksExplored >= DomConditionsMaxDomBlocks)
661 if (!Q.DT->getNode(Current)->getIDom())
663 Current = Q.DT->getNode(Current)->getIDom()->getBlock();
665 // found function entry
668 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator());
669 if (!BI || BI->isUnconditional())
671 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition());
675 // We're looking for conditions that are guaranteed to hold at the context
676 // instruction. Finding a condition where one path dominates the context
677 // isn't enough because both the true and false cases could merge before
678 // the context instruction we're actually interested in. Instead, we need
679 // to ensure that the taken *edge* dominates the context instruction. We
680 // know that the edge must be reachable since we started from a reachable
682 BasicBlock *BB0 = BI->getSuccessor(0);
683 BasicBlockEdge Edge(BI->getParent(), BB0);
684 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
687 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
691 // Option 2 - Search the other uses of V
692 unsigned NumUsesExplored = 0;
693 for (auto U : V->users()) {
694 // Avoid massive lists
695 if (NumUsesExplored >= DomConditionsMaxUses)
698 // Consider only compare instructions uniquely controlling a branch
699 ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
703 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
706 for (auto *CmpU : Cmp->users()) {
707 BranchInst *BI = dyn_cast<BranchInst>(CmpU);
708 if (!BI || BI->isUnconditional())
710 // We're looking for conditions that are guaranteed to hold at the
711 // context instruction. Finding a condition where one path dominates
712 // the context isn't enough because both the true and false cases could
713 // merge before the context instruction we're actually interested in.
714 // Instead, we need to ensure that the taken *edge* dominates the context
716 BasicBlock *BB0 = BI->getSuccessor(0);
717 BasicBlockEdge Edge(BI->getParent(), BB0);
718 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent()))
721 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth,
727 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
728 APInt &KnownOne, const DataLayout &DL,
729 unsigned Depth, const Query &Q) {
730 // Use of assumptions is context-sensitive. If we don't have a context, we
732 if (!Q.AC || !Q.CxtI)
735 unsigned BitWidth = KnownZero.getBitWidth();
737 for (auto &AssumeVH : Q.AC->assumptions()) {
740 CallInst *I = cast<CallInst>(AssumeVH);
741 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
742 "Got assumption for the wrong function!");
743 if (Q.ExclInvs.count(I))
746 // Warning: This loop can end up being somewhat performance sensetive.
747 // We're running this loop for once for each value queried resulting in a
748 // runtime of ~O(#assumes * #values).
750 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
751 "must be an assume intrinsic");
753 Value *Arg = I->getArgOperand(0);
755 if (Arg == V && isValidAssumeForContext(I, Q)) {
756 assert(BitWidth == 1 && "assume operand is not i1?");
757 KnownZero.clearAllBits();
758 KnownOne.setAllBits();
762 // The remaining tests are all recursive, so bail out if we hit the limit.
763 if (Depth == MaxDepth)
767 auto m_V = m_CombineOr(m_Specific(V),
768 m_CombineOr(m_PtrToInt(m_Specific(V)),
769 m_BitCast(m_Specific(V))));
771 CmpInst::Predicate Pred;
774 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
775 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
776 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
777 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
778 KnownZero |= RHSKnownZero;
779 KnownOne |= RHSKnownOne;
781 } else if (match(Arg,
782 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
783 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
784 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
785 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
786 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
787 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
789 // For those bits in the mask that are known to be one, we can propagate
790 // known bits from the RHS to V.
791 KnownZero |= RHSKnownZero & MaskKnownOne;
792 KnownOne |= RHSKnownOne & MaskKnownOne;
793 // assume(~(v & b) = a)
794 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
796 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
797 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
798 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
799 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
800 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
802 // For those bits in the mask that are known to be one, we can propagate
803 // inverted known bits from the RHS to V.
804 KnownZero |= RHSKnownOne & MaskKnownOne;
805 KnownOne |= RHSKnownZero & MaskKnownOne;
807 } else if (match(Arg,
808 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
809 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
810 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
811 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
812 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
813 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
815 // For those bits in B that are known to be zero, we can propagate known
816 // bits from the RHS to V.
817 KnownZero |= RHSKnownZero & BKnownZero;
818 KnownOne |= RHSKnownOne & BKnownZero;
819 // assume(~(v | b) = a)
820 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
822 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
823 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
824 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
825 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
826 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
828 // For those bits in B that are known to be zero, we can propagate
829 // inverted known bits from the RHS to V.
830 KnownZero |= RHSKnownOne & BKnownZero;
831 KnownOne |= RHSKnownZero & BKnownZero;
833 } else if (match(Arg,
834 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
835 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
836 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
837 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
838 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
839 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
841 // For those bits in B that are known to be zero, we can propagate known
842 // bits from the RHS to V. For those bits in B that are known to be one,
843 // we can propagate inverted known bits from the RHS to V.
844 KnownZero |= RHSKnownZero & BKnownZero;
845 KnownOne |= RHSKnownOne & BKnownZero;
846 KnownZero |= RHSKnownOne & BKnownOne;
847 KnownOne |= RHSKnownZero & BKnownOne;
848 // assume(~(v ^ b) = a)
849 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
851 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
852 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
853 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
854 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
855 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
857 // For those bits in B that are known to be zero, we can propagate
858 // inverted known bits from the RHS to V. For those bits in B that are
859 // known to be one, we can propagate known bits from the RHS to V.
860 KnownZero |= RHSKnownOne & BKnownZero;
861 KnownOne |= RHSKnownZero & BKnownZero;
862 KnownZero |= RHSKnownZero & BKnownOne;
863 KnownOne |= RHSKnownOne & BKnownOne;
864 // assume(v << c = a)
865 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
867 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
868 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
869 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
870 // For those bits in RHS that are known, we can propagate them to known
871 // bits in V shifted to the right by C.
872 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
873 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
874 // assume(~(v << c) = a)
875 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
877 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
878 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
879 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
880 // For those bits in RHS that are known, we can propagate them inverted
881 // to known bits in V shifted to the right by C.
882 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
883 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
884 // assume(v >> c = a)
885 } else if (match(Arg,
886 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
887 m_AShr(m_V, m_ConstantInt(C))),
889 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
890 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
891 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
892 // For those bits in RHS that are known, we can propagate them to known
893 // bits in V shifted to the right by C.
894 KnownZero |= RHSKnownZero << C->getZExtValue();
895 KnownOne |= RHSKnownOne << C->getZExtValue();
896 // assume(~(v >> c) = a)
897 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
898 m_LShr(m_V, m_ConstantInt(C)),
899 m_AShr(m_V, m_ConstantInt(C)))),
901 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) {
902 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
903 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
904 // For those bits in RHS that are known, we can propagate them inverted
905 // to known bits in V shifted to the right by C.
906 KnownZero |= RHSKnownOne << C->getZExtValue();
907 KnownOne |= RHSKnownZero << C->getZExtValue();
908 // assume(v >=_s c) where c is non-negative
909 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
910 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) {
911 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
912 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
914 if (RHSKnownZero.isNegative()) {
915 // We know that the sign bit is zero.
916 KnownZero |= APInt::getSignBit(BitWidth);
918 // assume(v >_s c) where c is at least -1.
919 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
920 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) {
921 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
922 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
924 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
925 // We know that the sign bit is zero.
926 KnownZero |= APInt::getSignBit(BitWidth);
928 // assume(v <=_s c) where c is negative
929 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
930 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) {
931 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
932 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
934 if (RHSKnownOne.isNegative()) {
935 // We know that the sign bit is one.
936 KnownOne |= APInt::getSignBit(BitWidth);
938 // assume(v <_s c) where c is non-positive
939 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
940 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) {
941 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
942 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
944 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
945 // We know that the sign bit is one.
946 KnownOne |= APInt::getSignBit(BitWidth);
949 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
950 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) {
951 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
952 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
954 // Whatever high bits in c are zero are known to be zero.
956 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
958 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
959 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) {
960 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
961 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
963 // Whatever high bits in c are zero are known to be zero (if c is a power
964 // of 2, then one more).
965 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL))
967 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
970 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
975 // Compute known bits from a shift operator, including those with a
976 // non-constant shift amount. KnownZero and KnownOne are the outputs of this
977 // function. KnownZero2 and KnownOne2 are pre-allocated temporaries with the
978 // same bit width as KnownZero and KnownOne. KZF and KOF are operator-specific
979 // functors that, given the known-zero or known-one bits respectively, and a
980 // shift amount, compute the implied known-zero or known-one bits of the shift
981 // operator's result respectively for that shift amount. The results from calling
982 // KZF and KOF are conservatively combined for all permitted shift amounts.
983 template <typename KZFunctor, typename KOFunctor>
984 static void computeKnownBitsFromShiftOperator(Operator *I,
985 APInt &KnownZero, APInt &KnownOne,
986 APInt &KnownZero2, APInt &KnownOne2,
987 const DataLayout &DL, unsigned Depth, const Query &Q,
988 KZFunctor KZF, KOFunctor KOF) {
989 unsigned BitWidth = KnownZero.getBitWidth();
991 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
992 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
994 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
995 KnownZero = KZF(KnownZero, ShiftAmt);
996 KnownOne = KOF(KnownOne, ShiftAmt);
1000 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1002 // Note: We cannot use KnownZero.getLimitedValue() here, because if
1003 // BitWidth > 64 and any upper bits are known, we'll end up returning the
1004 // limit value (which implies all bits are known).
1005 uint64_t ShiftAmtKZ = KnownZero.zextOrTrunc(64).getZExtValue();
1006 uint64_t ShiftAmtKO = KnownOne.zextOrTrunc(64).getZExtValue();
1008 // It would be more-clearly correct to use the two temporaries for this
1009 // calculation. Reusing the APInts here to prevent unnecessary allocations.
1010 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1012 // If we know the shifter operand is nonzero, we can sometimes infer more
1013 // known bits. However this is expensive to compute, so be lazy about it and
1014 // only compute it when absolutely necessary.
1015 Optional<bool> ShifterOperandIsNonZero;
1017 // Early exit if we can't constrain any well-defined shift amount.
1018 if (!(ShiftAmtKZ & (BitWidth - 1)) && !(ShiftAmtKO & (BitWidth - 1))) {
1019 ShifterOperandIsNonZero =
1020 isKnownNonZero(I->getOperand(1), DL, Depth + 1, Q);
1021 if (!*ShifterOperandIsNonZero)
1025 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1027 KnownZero = KnownOne = APInt::getAllOnesValue(BitWidth);
1028 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
1029 // Combine the shifted known input bits only for those shift amounts
1030 // compatible with its known constraints.
1031 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
1033 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
1035 // If we know the shifter is nonzero, we may be able to infer more known
1036 // bits. This check is sunk down as far as possible to avoid the expensive
1037 // call to isKnownNonZero if the cheaper checks above fail.
1038 if (ShiftAmt == 0) {
1039 if (!ShifterOperandIsNonZero.hasValue())
1040 ShifterOperandIsNonZero =
1041 isKnownNonZero(I->getOperand(1), DL, Depth + 1, Q);
1042 if (*ShifterOperandIsNonZero)
1046 KnownZero &= KZF(KnownZero2, ShiftAmt);
1047 KnownOne &= KOF(KnownOne2, ShiftAmt);
1050 // If there are no compatible shift amounts, then we've proven that the shift
1051 // amount must be >= the BitWidth, and the result is undefined. We could
1052 // return anything we'd like, but we need to make sure the sets of known bits
1053 // stay disjoint (it should be better for some other code to actually
1054 // propagate the undef than to pick a value here using known bits).
1055 if ((KnownZero & KnownOne) != 0)
1056 KnownZero.clearAllBits(), KnownOne.clearAllBits();
1059 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero,
1060 APInt &KnownOne, const DataLayout &DL,
1061 unsigned Depth, const Query &Q) {
1062 unsigned BitWidth = KnownZero.getBitWidth();
1064 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
1065 switch (I->getOpcode()) {
1067 case Instruction::Load:
1068 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
1069 computeKnownBitsFromRangeMetadata(*MD, KnownZero, KnownOne);
1071 case Instruction::And: {
1072 // If either the LHS or the RHS are Zero, the result is zero.
1073 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1074 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1076 // Output known-1 bits are only known if set in both the LHS & RHS.
1077 KnownOne &= KnownOne2;
1078 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1079 KnownZero |= KnownZero2;
1081 // and(x, add (x, -1)) is a common idiom that always clears the low bit;
1082 // here we handle the more general case of adding any odd number by
1083 // matching the form add(x, add(x, y)) where y is odd.
1084 // TODO: This could be generalized to clearing any bit set in y where the
1085 // following bit is known to be unset in y.
1087 if (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)),
1089 match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)),
1091 APInt KnownZero3(BitWidth, 0), KnownOne3(BitWidth, 0);
1092 computeKnownBits(Y, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1093 if (KnownOne3.countTrailingOnes() > 0)
1094 KnownZero |= APInt::getLowBitsSet(BitWidth, 1);
1098 case Instruction::Or: {
1099 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1100 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1102 // Output known-0 bits are only known if clear in both the LHS & RHS.
1103 KnownZero &= KnownZero2;
1104 // Output known-1 are known to be set if set in either the LHS | RHS.
1105 KnownOne |= KnownOne2;
1108 case Instruction::Xor: {
1109 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
1110 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1112 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1113 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1114 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1115 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1116 KnownZero = KnownZeroOut;
1119 case Instruction::Mul: {
1120 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1121 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero,
1122 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1125 case Instruction::UDiv: {
1126 // For the purposes of computing leading zeros we can conservatively
1127 // treat a udiv as a logical right shift by the power of 2 known to
1128 // be less than the denominator.
1129 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1130 unsigned LeadZ = KnownZero2.countLeadingOnes();
1132 KnownOne2.clearAllBits();
1133 KnownZero2.clearAllBits();
1134 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1135 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
1136 if (RHSUnknownLeadingOnes != BitWidth)
1137 LeadZ = std::min(BitWidth,
1138 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
1140 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
1143 case Instruction::Select:
1144 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q);
1145 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1147 // Only known if known in both the LHS and RHS.
1148 KnownOne &= KnownOne2;
1149 KnownZero &= KnownZero2;
1151 case Instruction::FPTrunc:
1152 case Instruction::FPExt:
1153 case Instruction::FPToUI:
1154 case Instruction::FPToSI:
1155 case Instruction::SIToFP:
1156 case Instruction::UIToFP:
1157 break; // Can't work with floating point.
1158 case Instruction::PtrToInt:
1159 case Instruction::IntToPtr:
1160 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
1161 // FALL THROUGH and handle them the same as zext/trunc.
1162 case Instruction::ZExt:
1163 case Instruction::Trunc: {
1164 Type *SrcTy = I->getOperand(0)->getType();
1166 unsigned SrcBitWidth;
1167 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1168 // which fall through here.
1169 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType());
1171 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1172 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
1173 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
1174 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1175 KnownZero = KnownZero.zextOrTrunc(BitWidth);
1176 KnownOne = KnownOne.zextOrTrunc(BitWidth);
1177 // Any top bits are known to be zero.
1178 if (BitWidth > SrcBitWidth)
1179 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1182 case Instruction::BitCast: {
1183 Type *SrcTy = I->getOperand(0)->getType();
1184 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy() ||
1185 SrcTy->isFloatingPointTy()) &&
1186 // TODO: For now, not handling conversions like:
1187 // (bitcast i64 %x to <2 x i32>)
1188 !I->getType()->isVectorTy()) {
1189 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1194 case Instruction::SExt: {
1195 // Compute the bits in the result that are not present in the input.
1196 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1198 KnownZero = KnownZero.trunc(SrcBitWidth);
1199 KnownOne = KnownOne.trunc(SrcBitWidth);
1200 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1201 KnownZero = KnownZero.zext(BitWidth);
1202 KnownOne = KnownOne.zext(BitWidth);
1204 // If the sign bit of the input is known set or clear, then we know the
1205 // top bits of the result.
1206 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1207 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1208 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1209 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1212 case Instruction::Shl: {
1213 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1214 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1215 return (KnownZero << ShiftAmt) |
1216 APInt::getLowBitsSet(BitWidth, ShiftAmt); // Low bits known 0.
1219 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1220 return KnownOne << ShiftAmt;
1223 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1224 KnownZero2, KnownOne2, DL, Depth, Q,
1228 case Instruction::LShr: {
1229 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1230 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1231 return APIntOps::lshr(KnownZero, ShiftAmt) |
1232 // High bits known zero.
1233 APInt::getHighBitsSet(BitWidth, ShiftAmt);
1236 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1237 return APIntOps::lshr(KnownOne, ShiftAmt);
1240 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1241 KnownZero2, KnownOne2, DL, Depth, Q,
1245 case Instruction::AShr: {
1246 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1247 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) {
1248 return APIntOps::ashr(KnownZero, ShiftAmt);
1251 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) {
1252 return APIntOps::ashr(KnownOne, ShiftAmt);
1255 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne,
1256 KnownZero2, KnownOne2, DL, Depth, Q,
1260 case Instruction::Sub: {
1261 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1262 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1263 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1267 case Instruction::Add: {
1268 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1269 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1270 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1274 case Instruction::SRem:
1275 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1276 APInt RA = Rem->getValue().abs();
1277 if (RA.isPowerOf2()) {
1278 APInt LowBits = RA - 1;
1279 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1,
1282 // The low bits of the first operand are unchanged by the srem.
1283 KnownZero = KnownZero2 & LowBits;
1284 KnownOne = KnownOne2 & LowBits;
1286 // If the first operand is non-negative or has all low bits zero, then
1287 // the upper bits are all zero.
1288 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1289 KnownZero |= ~LowBits;
1291 // If the first operand is negative and not all low bits are zero, then
1292 // the upper bits are all one.
1293 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1294 KnownOne |= ~LowBits;
1296 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1300 // The sign bit is the LHS's sign bit, except when the result of the
1301 // remainder is zero.
1302 if (KnownZero.isNonNegative()) {
1303 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1304 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL,
1306 // If it's known zero, our sign bit is also zero.
1307 if (LHSKnownZero.isNegative())
1308 KnownZero.setBit(BitWidth - 1);
1312 case Instruction::URem: {
1313 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1314 APInt RA = Rem->getValue();
1315 if (RA.isPowerOf2()) {
1316 APInt LowBits = (RA - 1);
1317 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
1319 KnownZero |= ~LowBits;
1320 KnownOne &= LowBits;
1325 // Since the result is less than or equal to either operand, any leading
1326 // zero bits in either operand must also exist in the result.
1327 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q);
1328 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q);
1330 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1331 KnownZero2.countLeadingOnes());
1332 KnownOne.clearAllBits();
1333 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1337 case Instruction::Alloca: {
1338 AllocaInst *AI = cast<AllocaInst>(I);
1339 unsigned Align = AI->getAlignment();
1341 Align = DL.getABITypeAlignment(AI->getType()->getElementType());
1344 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1347 case Instruction::GetElementPtr: {
1348 // Analyze all of the subscripts of this getelementptr instruction
1349 // to determine if we can prove known low zero bits.
1350 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1351 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL,
1353 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1355 gep_type_iterator GTI = gep_type_begin(I);
1356 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1357 Value *Index = I->getOperand(i);
1358 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1359 // Handle struct member offset arithmetic.
1361 // Handle case when index is vector zeroinitializer
1362 Constant *CIndex = cast<Constant>(Index);
1363 if (CIndex->isZeroValue())
1366 if (CIndex->getType()->isVectorTy())
1367 Index = CIndex->getSplatValue();
1369 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1370 const StructLayout *SL = DL.getStructLayout(STy);
1371 uint64_t Offset = SL->getElementOffset(Idx);
1372 TrailZ = std::min<unsigned>(TrailZ,
1373 countTrailingZeros(Offset));
1375 // Handle array index arithmetic.
1376 Type *IndexedTy = GTI.getIndexedType();
1377 if (!IndexedTy->isSized()) {
1381 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1382 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy);
1383 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1384 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1,
1386 TrailZ = std::min(TrailZ,
1387 unsigned(countTrailingZeros(TypeSize) +
1388 LocalKnownZero.countTrailingOnes()));
1392 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1395 case Instruction::PHI: {
1396 PHINode *P = cast<PHINode>(I);
1397 // Handle the case of a simple two-predecessor recurrence PHI.
1398 // There's a lot more that could theoretically be done here, but
1399 // this is sufficient to catch some interesting cases.
1400 if (P->getNumIncomingValues() == 2) {
1401 for (unsigned i = 0; i != 2; ++i) {
1402 Value *L = P->getIncomingValue(i);
1403 Value *R = P->getIncomingValue(!i);
1404 Operator *LU = dyn_cast<Operator>(L);
1407 unsigned Opcode = LU->getOpcode();
1408 // Check for operations that have the property that if
1409 // both their operands have low zero bits, the result
1410 // will have low zero bits.
1411 if (Opcode == Instruction::Add ||
1412 Opcode == Instruction::Sub ||
1413 Opcode == Instruction::And ||
1414 Opcode == Instruction::Or ||
1415 Opcode == Instruction::Mul) {
1416 Value *LL = LU->getOperand(0);
1417 Value *LR = LU->getOperand(1);
1418 // Find a recurrence.
1425 // Ok, we have a PHI of the form L op= R. Check for low
1427 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q);
1429 // We need to take the minimum number of known bits
1430 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1431 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q);
1433 KnownZero = APInt::getLowBitsSet(BitWidth,
1434 std::min(KnownZero2.countTrailingOnes(),
1435 KnownZero3.countTrailingOnes()));
1441 // Unreachable blocks may have zero-operand PHI nodes.
1442 if (P->getNumIncomingValues() == 0)
1445 // Otherwise take the unions of the known bit sets of the operands,
1446 // taking conservative care to avoid excessive recursion.
1447 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1448 // Skip if every incoming value references to ourself.
1449 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1452 KnownZero = APInt::getAllOnesValue(BitWidth);
1453 KnownOne = APInt::getAllOnesValue(BitWidth);
1454 for (Value *IncValue : P->incoming_values()) {
1455 // Skip direct self references.
1456 if (IncValue == P) continue;
1458 KnownZero2 = APInt(BitWidth, 0);
1459 KnownOne2 = APInt(BitWidth, 0);
1460 // Recurse, but cap the recursion to one level, because we don't
1461 // want to waste time spinning around in loops.
1462 computeKnownBits(IncValue, KnownZero2, KnownOne2, DL,
1464 KnownZero &= KnownZero2;
1465 KnownOne &= KnownOne2;
1466 // If all bits have been ruled out, there's no need to check
1468 if (!KnownZero && !KnownOne)
1474 case Instruction::Call:
1475 case Instruction::Invoke:
1476 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1477 computeKnownBitsFromRangeMetadata(*MD, KnownZero, KnownOne);
1478 // If a range metadata is attached to this IntrinsicInst, intersect the
1479 // explicit range specified by the metadata and the implicit range of
1481 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1482 switch (II->getIntrinsicID()) {
1484 case Intrinsic::bswap:
1485 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1487 KnownZero |= KnownZero2.byteSwap();
1488 KnownOne |= KnownOne2.byteSwap();
1490 case Intrinsic::ctlz:
1491 case Intrinsic::cttz: {
1492 unsigned LowBits = Log2_32(BitWidth)+1;
1493 // If this call is undefined for 0, the result will be less than 2^n.
1494 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1496 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1499 case Intrinsic::ctpop: {
1500 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL,
1502 // We can bound the space the count needs. Also, bits known to be zero
1503 // can't contribute to the population.
1504 unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation();
1505 unsigned LeadingZeros =
1506 APInt(BitWidth, BitsPossiblySet).countLeadingZeros();
1507 assert(LeadingZeros <= BitWidth);
1508 KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros);
1509 KnownOne &= ~KnownZero;
1510 // TODO: we could bound KnownOne using the lower bound on the number
1511 // of bits which might be set provided by popcnt KnownOne2.
1514 case Intrinsic::fabs: {
1515 Type *Ty = II->getType();
1516 APInt SignBit = APInt::getSignBit(Ty->getScalarSizeInBits());
1517 KnownZero |= APInt::getSplat(Ty->getPrimitiveSizeInBits(), SignBit);
1520 case Intrinsic::x86_sse42_crc32_64_64:
1521 KnownZero |= APInt::getHighBitsSet(64, 32);
1526 case Instruction::ExtractValue:
1527 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1528 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1529 if (EVI->getNumIndices() != 1) break;
1530 if (EVI->getIndices()[0] == 0) {
1531 switch (II->getIntrinsicID()) {
1533 case Intrinsic::uadd_with_overflow:
1534 case Intrinsic::sadd_with_overflow:
1535 computeKnownBitsAddSub(true, II->getArgOperand(0),
1536 II->getArgOperand(1), false, KnownZero,
1537 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1539 case Intrinsic::usub_with_overflow:
1540 case Intrinsic::ssub_with_overflow:
1541 computeKnownBitsAddSub(false, II->getArgOperand(0),
1542 II->getArgOperand(1), false, KnownZero,
1543 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q);
1545 case Intrinsic::umul_with_overflow:
1546 case Intrinsic::smul_with_overflow:
1547 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1548 KnownZero, KnownOne, KnownZero2, KnownOne2, DL,
1557 static unsigned getAlignment(const Value *V, const DataLayout &DL) {
1559 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1560 Align = GO->getAlignment();
1562 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
1563 Type *ObjectType = GVar->getType()->getElementType();
1564 if (ObjectType->isSized()) {
1565 // If the object is defined in the current Module, we'll be giving
1566 // it the preferred alignment. Otherwise, we have to assume that it
1567 // may only have the minimum ABI alignment.
1568 if (GVar->isStrongDefinitionForLinker())
1569 Align = DL.getPreferredAlignment(GVar);
1571 Align = DL.getABITypeAlignment(ObjectType);
1575 } else if (const Argument *A = dyn_cast<Argument>(V)) {
1576 Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
1578 if (!Align && A->hasStructRetAttr()) {
1579 // An sret parameter has at least the ABI alignment of the return type.
1580 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
1581 if (EltTy->isSized())
1582 Align = DL.getABITypeAlignment(EltTy);
1584 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
1585 Align = AI->getAlignment();
1586 else if (auto CS = ImmutableCallSite(V))
1587 Align = CS.getAttributes().getParamAlignment(AttributeSet::ReturnIndex);
1588 else if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1589 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_align)) {
1590 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
1591 Align = CI->getLimitedValue();
1597 /// Determine which bits of V are known to be either zero or one and return
1598 /// them in the KnownZero/KnownOne bit sets.
1600 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1601 /// we cannot optimize based on the assumption that it is zero without changing
1602 /// it to be an explicit zero. If we don't change it to zero, other code could
1603 /// optimized based on the contradictory assumption that it is non-zero.
1604 /// Because instcombine aggressively folds operations with undef args anyway,
1605 /// this won't lose us code quality.
1607 /// This function is defined on values with integer type, values with pointer
1608 /// type, and vectors of integers. In the case
1609 /// where V is a vector, known zero, and known one values are the
1610 /// same width as the vector element, and the bit is set only if it is true
1611 /// for all of the elements in the vector.
1612 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
1613 const DataLayout &DL, unsigned Depth, const Query &Q) {
1614 assert(V && "No Value?");
1615 assert(Depth <= MaxDepth && "Limit Search Depth");
1616 unsigned BitWidth = KnownZero.getBitWidth();
1618 assert((V->getType()->isIntOrIntVectorTy() ||
1619 V->getType()->isFPOrFPVectorTy() ||
1620 V->getType()->getScalarType()->isPointerTy()) &&
1621 "Not integer, floating point, or pointer type!");
1622 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1623 (!V->getType()->isIntOrIntVectorTy() ||
1624 V->getType()->getScalarSizeInBits() == BitWidth) &&
1625 KnownZero.getBitWidth() == BitWidth &&
1626 KnownOne.getBitWidth() == BitWidth &&
1627 "V, KnownOne and KnownZero should have same BitWidth");
1629 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1630 // We know all of the bits for a constant!
1631 KnownOne = CI->getValue();
1632 KnownZero = ~KnownOne;
1635 // Null and aggregate-zero are all-zeros.
1636 if (isa<ConstantPointerNull>(V) ||
1637 isa<ConstantAggregateZero>(V)) {
1638 KnownOne.clearAllBits();
1639 KnownZero = APInt::getAllOnesValue(BitWidth);
1642 // Handle a constant vector by taking the intersection of the known bits of
1643 // each element. There is no real need to handle ConstantVector here, because
1644 // we don't handle undef in any particularly useful way.
1645 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1646 // We know that CDS must be a vector of integers. Take the intersection of
1648 KnownZero.setAllBits(); KnownOne.setAllBits();
1649 APInt Elt(KnownZero.getBitWidth(), 0);
1650 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1651 Elt = CDS->getElementAsInteger(i);
1658 // Start out not knowing anything.
1659 KnownZero.clearAllBits(); KnownOne.clearAllBits();
1661 // Limit search depth.
1662 // All recursive calls that increase depth must come after this.
1663 if (Depth == MaxDepth)
1666 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1667 // the bits of its aliasee.
1668 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1669 if (!GA->mayBeOverridden())
1670 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q);
1674 if (Operator *I = dyn_cast<Operator>(V))
1675 computeKnownBitsFromOperator(I, KnownZero, KnownOne, DL, Depth, Q);
1677 // Aligned pointers have trailing zeros - refine KnownZero set
1678 if (V->getType()->isPointerTy()) {
1679 unsigned Align = getAlignment(V, DL);
1681 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1684 // computeKnownBitsFromAssume and computeKnownBitsFromDominatingCondition
1685 // strictly refines KnownZero and KnownOne. Therefore, we run them after
1686 // computeKnownBitsFromOperator.
1688 // Check whether a nearby assume intrinsic can determine some known bits.
1689 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q);
1691 // Check whether there's a dominating condition which implies something about
1692 // this value at the given context.
1693 if (EnableDomConditions && Depth <= DomConditionsMaxDepth)
1694 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth,
1697 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1700 /// Determine whether the sign bit is known to be zero or one.
1701 /// Convenience wrapper around computeKnownBits.
1702 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1703 const DataLayout &DL, unsigned Depth, const Query &Q) {
1704 unsigned BitWidth = getBitWidth(V->getType(), DL);
1710 APInt ZeroBits(BitWidth, 0);
1711 APInt OneBits(BitWidth, 0);
1712 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q);
1713 KnownOne = OneBits[BitWidth - 1];
1714 KnownZero = ZeroBits[BitWidth - 1];
1717 /// Return true if the given value is known to have exactly one
1718 /// bit set when defined. For vectors return true if every element is known to
1719 /// be a power of two when defined. Supports values with integer or pointer
1720 /// types and vectors of integers.
1721 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1722 const Query &Q, const DataLayout &DL) {
1723 if (Constant *C = dyn_cast<Constant>(V)) {
1724 if (C->isNullValue())
1726 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1727 return CI->getValue().isPowerOf2();
1728 // TODO: Handle vector constants.
1731 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1732 // it is shifted off the end then the result is undefined.
1733 if (match(V, m_Shl(m_One(), m_Value())))
1736 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1737 // bottom. If it is shifted off the bottom then the result is undefined.
1738 if (match(V, m_LShr(m_SignBit(), m_Value())))
1741 // The remaining tests are all recursive, so bail out if we hit the limit.
1742 if (Depth++ == MaxDepth)
1745 Value *X = nullptr, *Y = nullptr;
1746 // A shift left or a logical shift right of a power of two is a power of two
1748 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1749 match(V, m_LShr(m_Value(X), m_Value()))))
1750 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL);
1752 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1753 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL);
1755 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1756 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) &&
1757 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL);
1759 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1760 // A power of two and'd with anything is a power of two or zero.
1761 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) ||
1762 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL))
1764 // X & (-X) is always a power of two or zero.
1765 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1770 // Adding a power-of-two or zero to the same power-of-two or zero yields
1771 // either the original power-of-two, a larger power-of-two or zero.
1772 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1773 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1774 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1775 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1776 match(X, m_And(m_Value(), m_Specific(Y))))
1777 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL))
1779 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1780 match(Y, m_And(m_Value(), m_Specific(X))))
1781 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL))
1784 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1785 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1786 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q);
1788 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1789 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q);
1790 // If i8 V is a power of two or zero:
1791 // ZeroBits: 1 1 1 0 1 1 1 1
1792 // ~ZeroBits: 0 0 0 1 0 0 0 0
1793 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1794 // If OrZero isn't set, we cannot give back a zero result.
1795 // Make sure either the LHS or RHS has a bit set.
1796 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1801 // An exact divide or right shift can only shift off zero bits, so the result
1802 // is a power of two only if the first operand is a power of two and not
1803 // copying a sign bit (sdiv int_min, 2).
1804 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1805 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1806 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1813 /// \brief Test whether a GEP's result is known to be non-null.
1815 /// Uses properties inherent in a GEP to try to determine whether it is known
1818 /// Currently this routine does not support vector GEPs.
1819 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL,
1820 unsigned Depth, const Query &Q) {
1821 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1824 // FIXME: Support vector-GEPs.
1825 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1827 // If the base pointer is non-null, we cannot walk to a null address with an
1828 // inbounds GEP in address space zero.
1829 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1832 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1833 // If so, then the GEP cannot produce a null pointer, as doing so would
1834 // inherently violate the inbounds contract within address space zero.
1835 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1836 GTI != GTE; ++GTI) {
1837 // Struct types are easy -- they must always be indexed by a constant.
1838 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1839 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1840 unsigned ElementIdx = OpC->getZExtValue();
1841 const StructLayout *SL = DL.getStructLayout(STy);
1842 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1843 if (ElementOffset > 0)
1848 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1849 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1852 // Fast path the constant operand case both for efficiency and so we don't
1853 // increment Depth when just zipping down an all-constant GEP.
1854 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1860 // We post-increment Depth here because while isKnownNonZero increments it
1861 // as well, when we pop back up that increment won't persist. We don't want
1862 // to recurse 10k times just because we have 10k GEP operands. We don't
1863 // bail completely out because we want to handle constant GEPs regardless
1865 if (Depth++ >= MaxDepth)
1868 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1875 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1876 /// ensure that the value it's attached to is never Value? 'RangeType' is
1877 /// is the type of the value described by the range.
1878 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1879 const APInt& Value) {
1880 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1881 assert(NumRanges >= 1);
1882 for (unsigned i = 0; i < NumRanges; ++i) {
1883 ConstantInt *Lower =
1884 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1885 ConstantInt *Upper =
1886 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1887 ConstantRange Range(Lower->getValue(), Upper->getValue());
1888 if (Range.contains(Value))
1894 /// Return true if the given value is known to be non-zero when defined.
1895 /// For vectors return true if every element is known to be non-zero when
1896 /// defined. Supports values with integer or pointer type and vectors of
1898 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth,
1900 if (Constant *C = dyn_cast<Constant>(V)) {
1901 if (C->isNullValue())
1903 if (isa<ConstantInt>(C))
1904 // Must be non-zero due to null test above.
1906 // TODO: Handle vectors
1910 if (Instruction* I = dyn_cast<Instruction>(V)) {
1911 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1912 // If the possible ranges don't contain zero, then the value is
1913 // definitely non-zero.
1914 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1915 const APInt ZeroValue(Ty->getBitWidth(), 0);
1916 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1922 // The remaining tests are all recursive, so bail out if we hit the limit.
1923 if (Depth++ >= MaxDepth)
1926 // Check for pointer simplifications.
1927 if (V->getType()->isPointerTy()) {
1928 if (isKnownNonNull(V))
1930 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1931 if (isGEPKnownNonNull(GEP, DL, Depth, Q))
1935 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL);
1937 // X | Y != 0 if X != 0 or Y != 0.
1938 Value *X = nullptr, *Y = nullptr;
1939 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1940 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q);
1942 // ext X != 0 if X != 0.
1943 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1944 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q);
1946 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1947 // if the lowest bit is shifted off the end.
1948 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1949 // shl nuw can't remove any non-zero bits.
1950 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1951 if (BO->hasNoUnsignedWrap())
1952 return isKnownNonZero(X, DL, Depth, Q);
1954 APInt KnownZero(BitWidth, 0);
1955 APInt KnownOne(BitWidth, 0);
1956 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1960 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1961 // defined if the sign bit is shifted off the end.
1962 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1963 // shr exact can only shift out zero bits.
1964 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1966 return isKnownNonZero(X, DL, Depth, Q);
1968 bool XKnownNonNegative, XKnownNegative;
1969 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1973 // If the shifter operand is a constant, and all of the bits shifted
1974 // out are known to be zero, and X is known non-zero then at least one
1975 // non-zero bit must remain.
1976 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1977 APInt KnownZero(BitWidth, 0);
1978 APInt KnownOne(BitWidth, 0);
1979 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
1981 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1982 // Is there a known one in the portion not shifted out?
1983 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal)
1985 // Are all the bits to be shifted out known zero?
1986 if (KnownZero.countTrailingOnes() >= ShiftVal)
1987 return isKnownNonZero(X, DL, Depth, Q);
1990 // div exact can only produce a zero if the dividend is zero.
1991 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1992 return isKnownNonZero(X, DL, Depth, Q);
1995 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1996 bool XKnownNonNegative, XKnownNegative;
1997 bool YKnownNonNegative, YKnownNegative;
1998 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q);
1999 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q);
2001 // If X and Y are both non-negative (as signed values) then their sum is not
2002 // zero unless both X and Y are zero.
2003 if (XKnownNonNegative && YKnownNonNegative)
2004 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q))
2007 // If X and Y are both negative (as signed values) then their sum is not
2008 // zero unless both X and Y equal INT_MIN.
2009 if (BitWidth && XKnownNegative && YKnownNegative) {
2010 APInt KnownZero(BitWidth, 0);
2011 APInt KnownOne(BitWidth, 0);
2012 APInt Mask = APInt::getSignedMaxValue(BitWidth);
2013 // The sign bit of X is set. If some other bit is set then X is not equal
2015 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q);
2016 if ((KnownOne & Mask) != 0)
2018 // The sign bit of Y is set. If some other bit is set then Y is not equal
2020 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q);
2021 if ((KnownOne & Mask) != 0)
2025 // The sum of a non-negative number and a power of two is not zero.
2026 if (XKnownNonNegative &&
2027 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL))
2029 if (YKnownNonNegative &&
2030 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL))
2034 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2035 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2036 // If X and Y are non-zero then so is X * Y as long as the multiplication
2037 // does not overflow.
2038 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
2039 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q))
2042 // (C ? X : Y) != 0 if X != 0 and Y != 0.
2043 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2044 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) &&
2045 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q))
2049 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
2050 // Try and detect a recurrence that monotonically increases from a
2051 // starting value, as these are common as induction variables.
2052 if (PN->getNumIncomingValues() == 2) {
2053 Value *Start = PN->getIncomingValue(0);
2054 Value *Induction = PN->getIncomingValue(1);
2055 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2056 std::swap(Start, Induction);
2057 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2058 if (!C->isZero() && !C->isNegative()) {
2060 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2061 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2069 if (!BitWidth) return false;
2070 APInt KnownZero(BitWidth, 0);
2071 APInt KnownOne(BitWidth, 0);
2072 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2073 return KnownOne != 0;
2076 /// Return true if V2 == V1 + X, where X is known non-zero.
2077 static bool isAddOfNonZero(Value *V1, Value *V2, const DataLayout &DL,
2079 BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2080 if (!BO || BO->getOpcode() != Instruction::Add)
2082 Value *Op = nullptr;
2083 if (V2 == BO->getOperand(0))
2084 Op = BO->getOperand(1);
2085 else if (V2 == BO->getOperand(1))
2086 Op = BO->getOperand(0);
2089 return isKnownNonZero(Op, DL, 0, Q);
2092 /// Return true if it is known that V1 != V2.
2093 static bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL,
2095 if (V1->getType()->isVectorTy() || V1 == V2)
2097 if (V1->getType() != V2->getType())
2098 // We can't look through casts yet.
2100 if (isAddOfNonZero(V1, V2, DL, Q) || isAddOfNonZero(V2, V1, DL, Q))
2103 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
2104 // Are any known bits in V1 contradictory to known bits in V2? If V1
2105 // has a known zero where V2 has a known one, they must not be equal.
2106 auto BitWidth = Ty->getBitWidth();
2107 APInt KnownZero1(BitWidth, 0);
2108 APInt KnownOne1(BitWidth, 0);
2109 computeKnownBits(V1, KnownZero1, KnownOne1, DL, 0, Q);
2110 APInt KnownZero2(BitWidth, 0);
2111 APInt KnownOne2(BitWidth, 0);
2112 computeKnownBits(V2, KnownZero2, KnownOne2, DL, 0, Q);
2114 auto OppositeBits = (KnownZero1 & KnownOne2) | (KnownZero2 & KnownOne1);
2115 if (OppositeBits.getBoolValue())
2121 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2122 /// simplify operations downstream. Mask is known to be zero for bits that V
2125 /// This function is defined on values with integer type, values with pointer
2126 /// type, and vectors of integers. In the case
2127 /// where V is a vector, the mask, known zero, and known one values are the
2128 /// same width as the vector element, and the bit is set only if it is true
2129 /// for all of the elements in the vector.
2130 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL,
2131 unsigned Depth, const Query &Q) {
2132 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
2133 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2134 return (KnownZero & Mask) == Mask;
2139 /// Return the number of times the sign bit of the register is replicated into
2140 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2141 /// (itself), but other cases can give us information. For example, immediately
2142 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2143 /// other, so we return 3.
2145 /// 'Op' must have a scalar integer type.
2147 unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth,
2149 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType());
2151 unsigned FirstAnswer = 1;
2153 // Note that ConstantInt is handled by the general computeKnownBits case
2157 return 1; // Limit search depth.
2159 Operator *U = dyn_cast<Operator>(V);
2160 switch (Operator::getOpcode(V)) {
2162 case Instruction::SExt:
2163 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2164 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp;
2166 case Instruction::SDiv: {
2167 const APInt *Denominator;
2168 // sdiv X, C -> adds log(C) sign bits.
2169 if (match(U->getOperand(1), m_APInt(Denominator))) {
2171 // Ignore non-positive denominator.
2172 if (!Denominator->isStrictlyPositive())
2175 // Calculate the incoming numerator bits.
2176 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2178 // Add floor(log(C)) bits to the numerator bits.
2179 return std::min(TyBits, NumBits + Denominator->logBase2());
2184 case Instruction::SRem: {
2185 const APInt *Denominator;
2186 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2187 // positive constant. This let us put a lower bound on the number of sign
2189 if (match(U->getOperand(1), m_APInt(Denominator))) {
2191 // Ignore non-positive denominator.
2192 if (!Denominator->isStrictlyPositive())
2195 // Calculate the incoming numerator bits. SRem by a positive constant
2196 // can't lower the number of sign bits.
2198 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2200 // Calculate the leading sign bit constraints by examining the
2201 // denominator. Given that the denominator is positive, there are two
2204 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2205 // (1 << ceilLogBase2(C)).
2207 // 2. the numerator is negative. Then the result range is (-C,0] and
2208 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2210 // Thus a lower bound on the number of sign bits is `TyBits -
2211 // ceilLogBase2(C)`.
2213 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2214 return std::max(NumrBits, ResBits);
2219 case Instruction::AShr: {
2220 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2221 // ashr X, C -> adds C sign bits. Vectors too.
2223 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2224 Tmp += ShAmt->getZExtValue();
2225 if (Tmp > TyBits) Tmp = TyBits;
2229 case Instruction::Shl: {
2231 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2232 // shl destroys sign bits.
2233 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2234 Tmp2 = ShAmt->getZExtValue();
2235 if (Tmp2 >= TyBits || // Bad shift.
2236 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2241 case Instruction::And:
2242 case Instruction::Or:
2243 case Instruction::Xor: // NOT is handled here.
2244 // Logical binary ops preserve the number of sign bits at the worst.
2245 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2247 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2248 FirstAnswer = std::min(Tmp, Tmp2);
2249 // We computed what we know about the sign bits as our first
2250 // answer. Now proceed to the generic code that uses
2251 // computeKnownBits, and pick whichever answer is better.
2255 case Instruction::Select:
2256 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2257 if (Tmp == 1) return 1; // Early out.
2258 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q);
2259 return std::min(Tmp, Tmp2);
2261 case Instruction::Add:
2262 // Add can have at most one carry bit. Thus we know that the output
2263 // is, at worst, one more bit than the inputs.
2264 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2265 if (Tmp == 1) return 1; // Early out.
2267 // Special case decrementing a value (ADD X, -1):
2268 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2269 if (CRHS->isAllOnesValue()) {
2270 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2271 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1,
2274 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2276 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2279 // If we are subtracting one from a positive number, there is no carry
2280 // out of the result.
2281 if (KnownZero.isNegative())
2285 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2286 if (Tmp2 == 1) return 1;
2287 return std::min(Tmp, Tmp2)-1;
2289 case Instruction::Sub:
2290 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q);
2291 if (Tmp2 == 1) return 1;
2294 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2295 if (CLHS->isNullValue()) {
2296 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2297 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1,
2299 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2301 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
2304 // If the input is known to be positive (the sign bit is known clear),
2305 // the output of the NEG has the same number of sign bits as the input.
2306 if (KnownZero.isNegative())
2309 // Otherwise, we treat this like a SUB.
2312 // Sub can have at most one carry bit. Thus we know that the output
2313 // is, at worst, one more bit than the inputs.
2314 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q);
2315 if (Tmp == 1) return 1; // Early out.
2316 return std::min(Tmp, Tmp2)-1;
2318 case Instruction::PHI: {
2319 PHINode *PN = cast<PHINode>(U);
2320 unsigned NumIncomingValues = PN->getNumIncomingValues();
2321 // Don't analyze large in-degree PHIs.
2322 if (NumIncomingValues > 4) break;
2323 // Unreachable blocks may have zero-operand PHI nodes.
2324 if (NumIncomingValues == 0) break;
2326 // Take the minimum of all incoming values. This can't infinitely loop
2327 // because of our depth threshold.
2328 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q);
2329 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2330 if (Tmp == 1) return Tmp;
2332 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q));
2337 case Instruction::Trunc:
2338 // FIXME: it's tricky to do anything useful for this, but it is an important
2339 // case for targets like X86.
2343 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2344 // use this information.
2345 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2347 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q);
2349 if (KnownZero.isNegative()) { // sign bit is 0
2351 } else if (KnownOne.isNegative()) { // sign bit is 1;
2358 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2359 // the number of identical bits in the top of the input value.
2361 Mask <<= Mask.getBitWidth()-TyBits;
2362 // Return # leading zeros. We use 'min' here in case Val was zero before
2363 // shifting. We don't want to return '64' as for an i32 "0".
2364 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2367 /// This function computes the integer multiple of Base that equals V.
2368 /// If successful, it returns true and returns the multiple in
2369 /// Multiple. If unsuccessful, it returns false. It looks
2370 /// through SExt instructions only if LookThroughSExt is true.
2371 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2372 bool LookThroughSExt, unsigned Depth) {
2373 const unsigned MaxDepth = 6;
2375 assert(V && "No Value?");
2376 assert(Depth <= MaxDepth && "Limit Search Depth");
2377 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2379 Type *T = V->getType();
2381 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2391 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2392 Constant *BaseVal = ConstantInt::get(T, Base);
2393 if (CO && CO == BaseVal) {
2395 Multiple = ConstantInt::get(T, 1);
2399 if (CI && CI->getZExtValue() % Base == 0) {
2400 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2404 if (Depth == MaxDepth) return false; // Limit search depth.
2406 Operator *I = dyn_cast<Operator>(V);
2407 if (!I) return false;
2409 switch (I->getOpcode()) {
2411 case Instruction::SExt:
2412 if (!LookThroughSExt) return false;
2413 // otherwise fall through to ZExt
2414 case Instruction::ZExt:
2415 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2416 LookThroughSExt, Depth+1);
2417 case Instruction::Shl:
2418 case Instruction::Mul: {
2419 Value *Op0 = I->getOperand(0);
2420 Value *Op1 = I->getOperand(1);
2422 if (I->getOpcode() == Instruction::Shl) {
2423 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2424 if (!Op1CI) return false;
2425 // Turn Op0 << Op1 into Op0 * 2^Op1
2426 APInt Op1Int = Op1CI->getValue();
2427 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2428 APInt API(Op1Int.getBitWidth(), 0);
2429 API.setBit(BitToSet);
2430 Op1 = ConstantInt::get(V->getContext(), API);
2433 Value *Mul0 = nullptr;
2434 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2435 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2436 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2437 if (Op1C->getType()->getPrimitiveSizeInBits() <
2438 MulC->getType()->getPrimitiveSizeInBits())
2439 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2440 if (Op1C->getType()->getPrimitiveSizeInBits() >
2441 MulC->getType()->getPrimitiveSizeInBits())
2442 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2444 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2445 Multiple = ConstantExpr::getMul(MulC, Op1C);
2449 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2450 if (Mul0CI->getValue() == 1) {
2451 // V == Base * Op1, so return Op1
2457 Value *Mul1 = nullptr;
2458 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2459 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2460 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2461 if (Op0C->getType()->getPrimitiveSizeInBits() <
2462 MulC->getType()->getPrimitiveSizeInBits())
2463 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2464 if (Op0C->getType()->getPrimitiveSizeInBits() >
2465 MulC->getType()->getPrimitiveSizeInBits())
2466 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2468 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2469 Multiple = ConstantExpr::getMul(MulC, Op0C);
2473 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2474 if (Mul1CI->getValue() == 1) {
2475 // V == Base * Op0, so return Op0
2483 // We could not determine if V is a multiple of Base.
2487 /// Return true if we can prove that the specified FP value is never equal to
2490 /// NOTE: this function will need to be revisited when we support non-default
2493 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2494 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2495 return !CFP->getValueAPF().isNegZero();
2497 // FIXME: Magic number! At the least, this should be given a name because it's
2498 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2499 // expose it as a parameter, so it can be used for testing / experimenting.
2501 return false; // Limit search depth.
2503 const Operator *I = dyn_cast<Operator>(V);
2504 if (!I) return false;
2506 // Check if the nsz fast-math flag is set
2507 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2508 if (FPO->hasNoSignedZeros())
2511 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2512 if (I->getOpcode() == Instruction::FAdd)
2513 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2514 if (CFP->isNullValue())
2517 // sitofp and uitofp turn into +0.0 for zero.
2518 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2521 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2522 // sqrt(-0.0) = -0.0, no other negative results are possible.
2523 if (II->getIntrinsicID() == Intrinsic::sqrt)
2524 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2526 if (const CallInst *CI = dyn_cast<CallInst>(I))
2527 if (const Function *F = CI->getCalledFunction()) {
2528 if (F->isDeclaration()) {
2530 if (F->getName() == "abs") return true;
2531 // fabs[lf](x) != -0.0
2532 if (F->getName() == "fabs") return true;
2533 if (F->getName() == "fabsf") return true;
2534 if (F->getName() == "fabsl") return true;
2535 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2536 F->getName() == "sqrtl")
2537 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2544 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2545 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2546 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2548 // FIXME: Magic number! At the least, this should be given a name because it's
2549 // used similarly in CannotBeNegativeZero(). A better fix may be to
2550 // expose it as a parameter, so it can be used for testing / experimenting.
2552 return false; // Limit search depth.
2554 const Operator *I = dyn_cast<Operator>(V);
2555 if (!I) return false;
2557 switch (I->getOpcode()) {
2559 // Unsigned integers are always nonnegative.
2560 case Instruction::UIToFP:
2562 case Instruction::FMul:
2563 // x*x is always non-negative or a NaN.
2564 if (I->getOperand(0) == I->getOperand(1))
2567 case Instruction::FAdd:
2568 case Instruction::FDiv:
2569 case Instruction::FRem:
2570 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2571 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2572 case Instruction::Select:
2573 return CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1) &&
2574 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2575 case Instruction::FPExt:
2576 case Instruction::FPTrunc:
2577 // Widening/narrowing never change sign.
2578 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2579 case Instruction::Call:
2580 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2581 switch (II->getIntrinsicID()) {
2583 case Intrinsic::maxnum:
2584 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) ||
2585 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2586 case Intrinsic::minnum:
2587 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2588 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2589 case Intrinsic::exp:
2590 case Intrinsic::exp2:
2591 case Intrinsic::fabs:
2592 case Intrinsic::sqrt:
2594 case Intrinsic::powi:
2595 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2596 // powi(x,n) is non-negative if n is even.
2597 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2600 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2601 case Intrinsic::fma:
2602 case Intrinsic::fmuladd:
2603 // x*x+y is non-negative if y is non-negative.
2604 return I->getOperand(0) == I->getOperand(1) &&
2605 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2612 /// If the specified value can be set by repeating the same byte in memory,
2613 /// return the i8 value that it is represented with. This is
2614 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2615 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2616 /// byte store (e.g. i16 0x1234), return null.
2617 Value *llvm::isBytewiseValue(Value *V) {
2618 // All byte-wide stores are splatable, even of arbitrary variables.
2619 if (V->getType()->isIntegerTy(8)) return V;
2621 // Handle 'null' ConstantArrayZero etc.
2622 if (Constant *C = dyn_cast<Constant>(V))
2623 if (C->isNullValue())
2624 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2626 // Constant float and double values can be handled as integer values if the
2627 // corresponding integer value is "byteable". An important case is 0.0.
2628 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2629 if (CFP->getType()->isFloatTy())
2630 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2631 if (CFP->getType()->isDoubleTy())
2632 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2633 // Don't handle long double formats, which have strange constraints.
2636 // We can handle constant integers that are multiple of 8 bits.
2637 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2638 if (CI->getBitWidth() % 8 == 0) {
2639 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2641 if (!CI->getValue().isSplat(8))
2643 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2647 // A ConstantDataArray/Vector is splatable if all its members are equal and
2649 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2650 Value *Elt = CA->getElementAsConstant(0);
2651 Value *Val = isBytewiseValue(Elt);
2655 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2656 if (CA->getElementAsConstant(I) != Elt)
2662 // Conceptually, we could handle things like:
2663 // %a = zext i8 %X to i16
2664 // %b = shl i16 %a, 8
2665 // %c = or i16 %a, %b
2666 // but until there is an example that actually needs this, it doesn't seem
2667 // worth worrying about.
2672 // This is the recursive version of BuildSubAggregate. It takes a few different
2673 // arguments. Idxs is the index within the nested struct From that we are
2674 // looking at now (which is of type IndexedType). IdxSkip is the number of
2675 // indices from Idxs that should be left out when inserting into the resulting
2676 // struct. To is the result struct built so far, new insertvalue instructions
2678 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2679 SmallVectorImpl<unsigned> &Idxs,
2681 Instruction *InsertBefore) {
2682 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2684 // Save the original To argument so we can modify it
2686 // General case, the type indexed by Idxs is a struct
2687 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2688 // Process each struct element recursively
2691 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2695 // Couldn't find any inserted value for this index? Cleanup
2696 while (PrevTo != OrigTo) {
2697 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2698 PrevTo = Del->getAggregateOperand();
2699 Del->eraseFromParent();
2701 // Stop processing elements
2705 // If we successfully found a value for each of our subaggregates
2709 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2710 // the struct's elements had a value that was inserted directly. In the latter
2711 // case, perhaps we can't determine each of the subelements individually, but
2712 // we might be able to find the complete struct somewhere.
2714 // Find the value that is at that particular spot
2715 Value *V = FindInsertedValue(From, Idxs);
2720 // Insert the value in the new (sub) aggregrate
2721 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2722 "tmp", InsertBefore);
2725 // This helper takes a nested struct and extracts a part of it (which is again a
2726 // struct) into a new value. For example, given the struct:
2727 // { a, { b, { c, d }, e } }
2728 // and the indices "1, 1" this returns
2731 // It does this by inserting an insertvalue for each element in the resulting
2732 // struct, as opposed to just inserting a single struct. This will only work if
2733 // each of the elements of the substruct are known (ie, inserted into From by an
2734 // insertvalue instruction somewhere).
2736 // All inserted insertvalue instructions are inserted before InsertBefore
2737 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2738 Instruction *InsertBefore) {
2739 assert(InsertBefore && "Must have someplace to insert!");
2740 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2742 Value *To = UndefValue::get(IndexedType);
2743 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2744 unsigned IdxSkip = Idxs.size();
2746 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2749 /// Given an aggregrate and an sequence of indices, see if
2750 /// the scalar value indexed is already around as a register, for example if it
2751 /// were inserted directly into the aggregrate.
2753 /// If InsertBefore is not null, this function will duplicate (modified)
2754 /// insertvalues when a part of a nested struct is extracted.
2755 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2756 Instruction *InsertBefore) {
2757 // Nothing to index? Just return V then (this is useful at the end of our
2759 if (idx_range.empty())
2761 // We have indices, so V should have an indexable type.
2762 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2763 "Not looking at a struct or array?");
2764 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2765 "Invalid indices for type?");
2767 if (Constant *C = dyn_cast<Constant>(V)) {
2768 C = C->getAggregateElement(idx_range[0]);
2769 if (!C) return nullptr;
2770 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2773 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2774 // Loop the indices for the insertvalue instruction in parallel with the
2775 // requested indices
2776 const unsigned *req_idx = idx_range.begin();
2777 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2778 i != e; ++i, ++req_idx) {
2779 if (req_idx == idx_range.end()) {
2780 // We can't handle this without inserting insertvalues
2784 // The requested index identifies a part of a nested aggregate. Handle
2785 // this specially. For example,
2786 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2787 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2788 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2789 // This can be changed into
2790 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2791 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2792 // which allows the unused 0,0 element from the nested struct to be
2794 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2798 // This insert value inserts something else than what we are looking for.
2799 // See if the (aggregate) value inserted into has the value we are
2800 // looking for, then.
2802 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2805 // If we end up here, the indices of the insertvalue match with those
2806 // requested (though possibly only partially). Now we recursively look at
2807 // the inserted value, passing any remaining indices.
2808 return FindInsertedValue(I->getInsertedValueOperand(),
2809 makeArrayRef(req_idx, idx_range.end()),
2813 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2814 // If we're extracting a value from an aggregate that was extracted from
2815 // something else, we can extract from that something else directly instead.
2816 // However, we will need to chain I's indices with the requested indices.
2818 // Calculate the number of indices required
2819 unsigned size = I->getNumIndices() + idx_range.size();
2820 // Allocate some space to put the new indices in
2821 SmallVector<unsigned, 5> Idxs;
2823 // Add indices from the extract value instruction
2824 Idxs.append(I->idx_begin(), I->idx_end());
2826 // Add requested indices
2827 Idxs.append(idx_range.begin(), idx_range.end());
2829 assert(Idxs.size() == size
2830 && "Number of indices added not correct?");
2832 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2834 // Otherwise, we don't know (such as, extracting from a function return value
2835 // or load instruction)
2839 /// Analyze the specified pointer to see if it can be expressed as a base
2840 /// pointer plus a constant offset. Return the base and offset to the caller.
2841 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2842 const DataLayout &DL) {
2843 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2844 APInt ByteOffset(BitWidth, 0);
2846 // We walk up the defs but use a visited set to handle unreachable code. In
2847 // that case, we stop after accumulating the cycle once (not that it
2849 SmallPtrSet<Value *, 16> Visited;
2850 while (Visited.insert(Ptr).second) {
2851 if (Ptr->getType()->isVectorTy())
2854 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2855 APInt GEPOffset(BitWidth, 0);
2856 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2859 ByteOffset += GEPOffset;
2861 Ptr = GEP->getPointerOperand();
2862 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2863 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2864 Ptr = cast<Operator>(Ptr)->getOperand(0);
2865 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2866 if (GA->mayBeOverridden())
2868 Ptr = GA->getAliasee();
2873 Offset = ByteOffset.getSExtValue();
2878 /// This function computes the length of a null-terminated C string pointed to
2879 /// by V. If successful, it returns true and returns the string in Str.
2880 /// If unsuccessful, it returns false.
2881 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2882 uint64_t Offset, bool TrimAtNul) {
2885 // Look through bitcast instructions and geps.
2886 V = V->stripPointerCasts();
2888 // If the value is a GEP instruction or constant expression, treat it as an
2890 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2891 // Make sure the GEP has exactly three arguments.
2892 if (GEP->getNumOperands() != 3)
2895 // Make sure the index-ee is a pointer to array of i8.
2896 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2897 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2898 if (!AT || !AT->getElementType()->isIntegerTy(8))
2901 // Check to make sure that the first operand of the GEP is an integer and
2902 // has value 0 so that we are sure we're indexing into the initializer.
2903 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2904 if (!FirstIdx || !FirstIdx->isZero())
2907 // If the second index isn't a ConstantInt, then this is a variable index
2908 // into the array. If this occurs, we can't say anything meaningful about
2910 uint64_t StartIdx = 0;
2911 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2912 StartIdx = CI->getZExtValue();
2915 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
2919 // The GEP instruction, constant or instruction, must reference a global
2920 // variable that is a constant and is initialized. The referenced constant
2921 // initializer is the array that we'll use for optimization.
2922 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2923 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2926 // Handle the all-zeros case
2927 if (GV->getInitializer()->isNullValue()) {
2928 // This is a degenerate case. The initializer is constant zero so the
2929 // length of the string must be zero.
2934 // Must be a Constant Array
2935 const ConstantDataArray *Array =
2936 dyn_cast<ConstantDataArray>(GV->getInitializer());
2937 if (!Array || !Array->isString())
2940 // Get the number of elements in the array
2941 uint64_t NumElts = Array->getType()->getArrayNumElements();
2943 // Start out with the entire array in the StringRef.
2944 Str = Array->getAsString();
2946 if (Offset > NumElts)
2949 // Skip over 'offset' bytes.
2950 Str = Str.substr(Offset);
2953 // Trim off the \0 and anything after it. If the array is not nul
2954 // terminated, we just return the whole end of string. The client may know
2955 // some other way that the string is length-bound.
2956 Str = Str.substr(0, Str.find('\0'));
2961 // These next two are very similar to the above, but also look through PHI
2963 // TODO: See if we can integrate these two together.
2965 /// If we can compute the length of the string pointed to by
2966 /// the specified pointer, return 'len+1'. If we can't, return 0.
2967 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2968 // Look through noop bitcast instructions.
2969 V = V->stripPointerCasts();
2971 // If this is a PHI node, there are two cases: either we have already seen it
2973 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2974 if (!PHIs.insert(PN).second)
2975 return ~0ULL; // already in the set.
2977 // If it was new, see if all the input strings are the same length.
2978 uint64_t LenSoFar = ~0ULL;
2979 for (Value *IncValue : PN->incoming_values()) {
2980 uint64_t Len = GetStringLengthH(IncValue, PHIs);
2981 if (Len == 0) return 0; // Unknown length -> unknown.
2983 if (Len == ~0ULL) continue;
2985 if (Len != LenSoFar && LenSoFar != ~0ULL)
2986 return 0; // Disagree -> unknown.
2990 // Success, all agree.
2994 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2995 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2996 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2997 if (Len1 == 0) return 0;
2998 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2999 if (Len2 == 0) return 0;
3000 if (Len1 == ~0ULL) return Len2;
3001 if (Len2 == ~0ULL) return Len1;
3002 if (Len1 != Len2) return 0;
3006 // Otherwise, see if we can read the string.
3008 if (!getConstantStringInfo(V, StrData))
3011 return StrData.size()+1;
3014 /// If we can compute the length of the string pointed to by
3015 /// the specified pointer, return 'len+1'. If we can't, return 0.
3016 uint64_t llvm::GetStringLength(Value *V) {
3017 if (!V->getType()->isPointerTy()) return 0;
3019 SmallPtrSet<PHINode*, 32> PHIs;
3020 uint64_t Len = GetStringLengthH(V, PHIs);
3021 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3022 // an empty string as a length.
3023 return Len == ~0ULL ? 1 : Len;
3026 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
3027 /// previous iteration of the loop was referring to the same object as \p PN.
3028 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) {
3029 // Find the loop-defined value.
3030 Loop *L = LI->getLoopFor(PN->getParent());
3031 if (PN->getNumIncomingValues() != 2)
3034 // Find the value from previous iteration.
3035 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3036 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3037 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3038 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3041 // If a new pointer is loaded in the loop, the pointer references a different
3042 // object in every iteration. E.g.:
3046 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3047 if (!L->isLoopInvariant(Load->getPointerOperand()))
3052 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3053 unsigned MaxLookup) {
3054 if (!V->getType()->isPointerTy())
3056 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3057 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3058 V = GEP->getPointerOperand();
3059 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3060 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3061 V = cast<Operator>(V)->getOperand(0);
3062 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3063 if (GA->mayBeOverridden())
3065 V = GA->getAliasee();
3067 // See if InstructionSimplify knows any relevant tricks.
3068 if (Instruction *I = dyn_cast<Instruction>(V))
3069 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3070 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
3077 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3082 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3083 const DataLayout &DL, LoopInfo *LI,
3084 unsigned MaxLookup) {
3085 SmallPtrSet<Value *, 4> Visited;
3086 SmallVector<Value *, 4> Worklist;
3087 Worklist.push_back(V);
3089 Value *P = Worklist.pop_back_val();
3090 P = GetUnderlyingObject(P, DL, MaxLookup);
3092 if (!Visited.insert(P).second)
3095 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3096 Worklist.push_back(SI->getTrueValue());
3097 Worklist.push_back(SI->getFalseValue());
3101 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3102 // If this PHI changes the underlying object in every iteration of the
3103 // loop, don't look through it. Consider:
3106 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3110 // Prev is tracking Curr one iteration behind so they refer to different
3111 // underlying objects.
3112 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3113 isSameUnderlyingObjectInLoop(PN, LI))
3114 for (Value *IncValue : PN->incoming_values())
3115 Worklist.push_back(IncValue);
3119 Objects.push_back(P);
3120 } while (!Worklist.empty());
3123 /// Return true if the only users of this pointer are lifetime markers.
3124 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3125 for (const User *U : V->users()) {
3126 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3127 if (!II) return false;
3129 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3130 II->getIntrinsicID() != Intrinsic::lifetime_end)
3136 static bool isDereferenceableFromAttribute(const Value *BV, APInt Offset,
3137 Type *Ty, const DataLayout &DL,
3138 const Instruction *CtxI,
3139 const DominatorTree *DT,
3140 const TargetLibraryInfo *TLI) {
3141 assert(Offset.isNonNegative() && "offset can't be negative");
3142 assert(Ty->isSized() && "must be sized");
3144 APInt DerefBytes(Offset.getBitWidth(), 0);
3145 bool CheckForNonNull = false;
3146 if (const Argument *A = dyn_cast<Argument>(BV)) {
3147 DerefBytes = A->getDereferenceableBytes();
3148 if (!DerefBytes.getBoolValue()) {
3149 DerefBytes = A->getDereferenceableOrNullBytes();
3150 CheckForNonNull = true;
3152 } else if (auto CS = ImmutableCallSite(BV)) {
3153 DerefBytes = CS.getDereferenceableBytes(0);
3154 if (!DerefBytes.getBoolValue()) {
3155 DerefBytes = CS.getDereferenceableOrNullBytes(0);
3156 CheckForNonNull = true;
3158 } else if (const LoadInst *LI = dyn_cast<LoadInst>(BV)) {
3159 if (MDNode *MD = LI->getMetadata(LLVMContext::MD_dereferenceable)) {
3160 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3161 DerefBytes = CI->getLimitedValue();
3163 if (!DerefBytes.getBoolValue()) {
3165 LI->getMetadata(LLVMContext::MD_dereferenceable_or_null)) {
3166 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(0));
3167 DerefBytes = CI->getLimitedValue();
3169 CheckForNonNull = true;
3173 if (DerefBytes.getBoolValue())
3174 if (DerefBytes.uge(Offset + DL.getTypeStoreSize(Ty)))
3175 if (!CheckForNonNull || isKnownNonNullAt(BV, CtxI, DT, TLI))
3181 static bool isDereferenceableFromAttribute(const Value *V, const DataLayout &DL,
3182 const Instruction *CtxI,
3183 const DominatorTree *DT,
3184 const TargetLibraryInfo *TLI) {
3185 Type *VTy = V->getType();
3186 Type *Ty = VTy->getPointerElementType();
3190 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3191 return isDereferenceableFromAttribute(V, Offset, Ty, DL, CtxI, DT, TLI);
3194 static bool isAligned(const Value *Base, APInt Offset, unsigned Align,
3195 const DataLayout &DL) {
3196 APInt BaseAlign(Offset.getBitWidth(), getAlignment(Base, DL));
3199 Type *Ty = Base->getType()->getPointerElementType();
3202 BaseAlign = DL.getABITypeAlignment(Ty);
3205 APInt Alignment(Offset.getBitWidth(), Align);
3207 assert(Alignment.isPowerOf2() && "must be a power of 2!");
3208 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1));
3211 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) {
3212 Type *Ty = Base->getType();
3213 assert(Ty->isSized() && "must be sized");
3214 APInt Offset(DL.getTypeStoreSizeInBits(Ty), 0);
3215 return isAligned(Base, Offset, Align, DL);
3218 /// Test if V is always a pointer to allocated and suitably aligned memory for
3219 /// a simple load or store.
3220 static bool isDereferenceableAndAlignedPointer(
3221 const Value *V, unsigned Align, const DataLayout &DL,
3222 const Instruction *CtxI, const DominatorTree *DT,
3223 const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited) {
3224 // Note that it is not safe to speculate into a malloc'd region because
3225 // malloc may return null.
3227 // These are obviously ok if aligned.
3228 if (isa<AllocaInst>(V))
3229 return isAligned(V, Align, DL);
3231 // It's not always safe to follow a bitcast, for example:
3232 // bitcast i8* (alloca i8) to i32*
3233 // would result in a 4-byte load from a 1-byte alloca. However,
3234 // if we're casting from a pointer from a type of larger size
3235 // to a type of smaller size (or the same size), and the alignment
3236 // is at least as large as for the resulting pointer type, then
3237 // we can look through the bitcast.
3238 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
3239 Type *STy = BC->getSrcTy()->getPointerElementType(),
3240 *DTy = BC->getDestTy()->getPointerElementType();
3241 if (STy->isSized() && DTy->isSized() &&
3242 (DL.getTypeStoreSize(STy) >= DL.getTypeStoreSize(DTy)) &&
3243 (DL.getABITypeAlignment(STy) >= DL.getABITypeAlignment(DTy)))
3244 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, DL,
3245 CtxI, DT, TLI, Visited);
3248 // Global variables which can't collapse to null are ok.
3249 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
3250 if (!GV->hasExternalWeakLinkage())
3251 return isAligned(V, Align, DL);
3253 // byval arguments are okay.
3254 if (const Argument *A = dyn_cast<Argument>(V))
3255 if (A->hasByValAttr())
3256 return isAligned(V, Align, DL);
3258 if (isDereferenceableFromAttribute(V, DL, CtxI, DT, TLI))
3259 return isAligned(V, Align, DL);
3261 // For GEPs, determine if the indexing lands within the allocated object.
3262 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3263 Type *VTy = GEP->getType();
3264 Type *Ty = VTy->getPointerElementType();
3265 const Value *Base = GEP->getPointerOperand();
3267 // Conservatively require that the base pointer be fully dereferenceable
3269 if (!Visited.insert(Base).second)
3271 if (!isDereferenceableAndAlignedPointer(Base, Align, DL, CtxI, DT, TLI,
3275 APInt Offset(DL.getPointerTypeSizeInBits(VTy), 0);
3276 if (!GEP->accumulateConstantOffset(DL, Offset))
3279 // Check if the load is within the bounds of the underlying object
3280 // and offset is aligned.
3281 uint64_t LoadSize = DL.getTypeStoreSize(Ty);
3282 Type *BaseType = Base->getType()->getPointerElementType();
3283 assert(isPowerOf2_32(Align) && "must be a power of 2!");
3284 return (Offset + LoadSize).ule(DL.getTypeAllocSize(BaseType)) &&
3285 !(Offset & APInt(Offset.getBitWidth(), Align-1));
3288 // For gc.relocate, look through relocations
3289 if (const GCRelocateInst *RelocateInst = dyn_cast<GCRelocateInst>(V))
3290 return isDereferenceableAndAlignedPointer(
3291 RelocateInst->getDerivedPtr(), Align, DL, CtxI, DT, TLI, Visited);
3293 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V))
3294 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, DL,
3295 CtxI, DT, TLI, Visited);
3297 // If we don't know, assume the worst.
3301 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align,
3302 const DataLayout &DL,
3303 const Instruction *CtxI,
3304 const DominatorTree *DT,
3305 const TargetLibraryInfo *TLI) {
3306 // When dereferenceability information is provided by a dereferenceable
3307 // attribute, we know exactly how many bytes are dereferenceable. If we can
3308 // determine the exact offset to the attributed variable, we can use that
3309 // information here.
3310 Type *VTy = V->getType();
3311 Type *Ty = VTy->getPointerElementType();
3313 // Require ABI alignment for loads without alignment specification
3315 Align = DL.getABITypeAlignment(Ty);
3317 if (Ty->isSized()) {
3318 APInt Offset(DL.getTypeStoreSizeInBits(VTy), 0);
3319 const Value *BV = V->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
3321 if (Offset.isNonNegative())
3322 if (isDereferenceableFromAttribute(BV, Offset, Ty, DL, CtxI, DT, TLI) &&
3323 isAligned(BV, Offset, Align, DL))
3327 SmallPtrSet<const Value *, 32> Visited;
3328 return ::isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT, TLI,
3332 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL,
3333 const Instruction *CtxI,
3334 const DominatorTree *DT,
3335 const TargetLibraryInfo *TLI) {
3336 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT, TLI);
3339 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3340 const Instruction *CtxI,
3341 const DominatorTree *DT,
3342 const TargetLibraryInfo *TLI) {
3343 const Operator *Inst = dyn_cast<Operator>(V);
3347 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3348 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3352 switch (Inst->getOpcode()) {
3355 case Instruction::UDiv:
3356 case Instruction::URem: {
3357 // x / y is undefined if y == 0.
3359 if (match(Inst->getOperand(1), m_APInt(V)))
3363 case Instruction::SDiv:
3364 case Instruction::SRem: {
3365 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3366 const APInt *Numerator, *Denominator;
3367 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3369 // We cannot hoist this division if the denominator is 0.
3370 if (*Denominator == 0)
3372 // It's safe to hoist if the denominator is not 0 or -1.
3373 if (*Denominator != -1)
3375 // At this point we know that the denominator is -1. It is safe to hoist as
3376 // long we know that the numerator is not INT_MIN.
3377 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3378 return !Numerator->isMinSignedValue();
3379 // The numerator *might* be MinSignedValue.
3382 case Instruction::Load: {
3383 const LoadInst *LI = cast<LoadInst>(Inst);
3384 if (!LI->isUnordered() ||
3385 // Speculative load may create a race that did not exist in the source.
3386 LI->getParent()->getParent()->hasFnAttribute(
3387 Attribute::SanitizeThread) ||
3388 // Speculative load may load data from dirty regions.
3389 LI->getParent()->getParent()->hasFnAttribute(
3390 Attribute::SanitizeAddress))
3392 const DataLayout &DL = LI->getModule()->getDataLayout();
3393 return isDereferenceableAndAlignedPointer(
3394 LI->getPointerOperand(), LI->getAlignment(), DL, CtxI, DT, TLI);
3396 case Instruction::Call: {
3397 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3398 switch (II->getIntrinsicID()) {
3399 // These synthetic intrinsics have no side-effects and just mark
3400 // information about their operands.
3401 // FIXME: There are other no-op synthetic instructions that potentially
3402 // should be considered at least *safe* to speculate...
3403 case Intrinsic::dbg_declare:
3404 case Intrinsic::dbg_value:
3407 case Intrinsic::bswap:
3408 case Intrinsic::ctlz:
3409 case Intrinsic::ctpop:
3410 case Intrinsic::cttz:
3411 case Intrinsic::objectsize:
3412 case Intrinsic::sadd_with_overflow:
3413 case Intrinsic::smul_with_overflow:
3414 case Intrinsic::ssub_with_overflow:
3415 case Intrinsic::uadd_with_overflow:
3416 case Intrinsic::umul_with_overflow:
3417 case Intrinsic::usub_with_overflow:
3419 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
3420 // errno like libm sqrt would.
3421 case Intrinsic::sqrt:
3422 case Intrinsic::fma:
3423 case Intrinsic::fmuladd:
3424 case Intrinsic::fabs:
3425 case Intrinsic::minnum:
3426 case Intrinsic::maxnum:
3428 // TODO: some fp intrinsics are marked as having the same error handling
3429 // as libm. They're safe to speculate when they won't error.
3430 // TODO: are convert_{from,to}_fp16 safe?
3431 // TODO: can we list target-specific intrinsics here?
3435 return false; // The called function could have undefined behavior or
3436 // side-effects, even if marked readnone nounwind.
3438 case Instruction::VAArg:
3439 case Instruction::Alloca:
3440 case Instruction::Invoke:
3441 case Instruction::PHI:
3442 case Instruction::Store:
3443 case Instruction::Ret:
3444 case Instruction::Br:
3445 case Instruction::IndirectBr:
3446 case Instruction::Switch:
3447 case Instruction::Unreachable:
3448 case Instruction::Fence:
3449 case Instruction::AtomicRMW:
3450 case Instruction::AtomicCmpXchg:
3451 case Instruction::LandingPad:
3452 case Instruction::Resume:
3453 case Instruction::CatchSwitch:
3454 case Instruction::CatchPad:
3455 case Instruction::CatchRet:
3456 case Instruction::CleanupPad:
3457 case Instruction::CleanupRet:
3458 return false; // Misc instructions which have effects
3462 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3463 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3466 /// Return true if we know that the specified value is never null.
3467 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
3468 assert(V->getType()->isPointerTy() && "V must be pointer type");
3470 // Alloca never returns null, malloc might.
3471 if (isa<AllocaInst>(V)) return true;
3473 // A byval, inalloca, or nonnull argument is never null.
3474 if (const Argument *A = dyn_cast<Argument>(V))
3475 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3477 // A global variable in address space 0 is non null unless extern weak.
3478 // Other address spaces may have null as a valid address for a global,
3479 // so we can't assume anything.
3480 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3481 return !GV->hasExternalWeakLinkage() &&
3482 GV->getType()->getAddressSpace() == 0;
3484 // A Load tagged w/nonnull metadata is never null.
3485 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3486 return LI->getMetadata(LLVMContext::MD_nonnull);
3488 if (auto CS = ImmutableCallSite(V))
3489 if (CS.isReturnNonNull())
3495 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3496 const Instruction *CtxI,
3497 const DominatorTree *DT) {
3498 assert(V->getType()->isPointerTy() && "V must be pointer type");
3500 unsigned NumUsesExplored = 0;
3501 for (auto U : V->users()) {
3502 // Avoid massive lists
3503 if (NumUsesExplored >= DomConditionsMaxUses)
3506 // Consider only compare instructions uniquely controlling a branch
3507 const ICmpInst *Cmp = dyn_cast<ICmpInst>(U);
3511 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse())
3514 for (auto *CmpU : Cmp->users()) {
3515 const BranchInst *BI = dyn_cast<BranchInst>(CmpU);
3519 assert(BI->isConditional() && "uses a comparison!");
3521 BasicBlock *NonNullSuccessor = nullptr;
3522 CmpInst::Predicate Pred;
3524 if (match(const_cast<ICmpInst*>(Cmp),
3525 m_c_ICmp(Pred, m_Specific(V), m_Zero()))) {
3526 if (Pred == ICmpInst::ICMP_EQ)
3527 NonNullSuccessor = BI->getSuccessor(1);
3528 else if (Pred == ICmpInst::ICMP_NE)
3529 NonNullSuccessor = BI->getSuccessor(0);
3532 if (NonNullSuccessor) {
3533 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3534 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3543 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3544 const DominatorTree *DT, const TargetLibraryInfo *TLI) {
3545 if (isKnownNonNull(V, TLI))
3548 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false;
3551 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
3552 const DataLayout &DL,
3553 AssumptionCache *AC,
3554 const Instruction *CxtI,
3555 const DominatorTree *DT) {
3556 // Multiplying n * m significant bits yields a result of n + m significant
3557 // bits. If the total number of significant bits does not exceed the
3558 // result bit width (minus 1), there is no overflow.
3559 // This means if we have enough leading zero bits in the operands
3560 // we can guarantee that the result does not overflow.
3561 // Ref: "Hacker's Delight" by Henry Warren
3562 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3563 APInt LHSKnownZero(BitWidth, 0);
3564 APInt LHSKnownOne(BitWidth, 0);
3565 APInt RHSKnownZero(BitWidth, 0);
3566 APInt RHSKnownOne(BitWidth, 0);
3567 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3569 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
3571 // Note that underestimating the number of zero bits gives a more
3572 // conservative answer.
3573 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
3574 RHSKnownZero.countLeadingOnes();
3575 // First handle the easy case: if we have enough zero bits there's
3576 // definitely no overflow.
3577 if (ZeroBits >= BitWidth)
3578 return OverflowResult::NeverOverflows;
3580 // Get the largest possible values for each operand.
3581 APInt LHSMax = ~LHSKnownZero;
3582 APInt RHSMax = ~RHSKnownZero;
3584 // We know the multiply operation doesn't overflow if the maximum values for
3585 // each operand will not overflow after we multiply them together.
3587 LHSMax.umul_ov(RHSMax, MaxOverflow);
3589 return OverflowResult::NeverOverflows;
3591 // We know it always overflows if multiplying the smallest possible values for
3592 // the operands also results in overflow.
3594 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
3596 return OverflowResult::AlwaysOverflows;
3598 return OverflowResult::MayOverflow;
3601 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
3602 const DataLayout &DL,
3603 AssumptionCache *AC,
3604 const Instruction *CxtI,
3605 const DominatorTree *DT) {
3606 bool LHSKnownNonNegative, LHSKnownNegative;
3607 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3609 if (LHSKnownNonNegative || LHSKnownNegative) {
3610 bool RHSKnownNonNegative, RHSKnownNegative;
3611 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3614 if (LHSKnownNegative && RHSKnownNegative) {
3615 // The sign bit is set in both cases: this MUST overflow.
3616 // Create a simple add instruction, and insert it into the struct.
3617 return OverflowResult::AlwaysOverflows;
3620 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3621 // The sign bit is clear in both cases: this CANNOT overflow.
3622 // Create a simple add instruction, and insert it into the struct.
3623 return OverflowResult::NeverOverflows;
3627 return OverflowResult::MayOverflow;
3630 static OverflowResult computeOverflowForSignedAdd(
3631 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL,
3632 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) {
3633 if (Add && Add->hasNoSignedWrap()) {
3634 return OverflowResult::NeverOverflows;
3637 bool LHSKnownNonNegative, LHSKnownNegative;
3638 bool RHSKnownNonNegative, RHSKnownNegative;
3639 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3641 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3644 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3645 (LHSKnownNegative && RHSKnownNonNegative)) {
3646 // The sign bits are opposite: this CANNOT overflow.
3647 return OverflowResult::NeverOverflows;
3650 // The remaining code needs Add to be available. Early returns if not so.
3652 return OverflowResult::MayOverflow;
3654 // If the sign of Add is the same as at least one of the operands, this add
3655 // CANNOT overflow. This is particularly useful when the sum is
3656 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3658 bool LHSOrRHSKnownNonNegative =
3659 (LHSKnownNonNegative || RHSKnownNonNegative);
3660 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3661 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3662 bool AddKnownNonNegative, AddKnownNegative;
3663 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3664 /*Depth=*/0, AC, CxtI, DT);
3665 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3666 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3667 return OverflowResult::NeverOverflows;
3671 return OverflowResult::MayOverflow;
3674 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add,
3675 const DataLayout &DL,
3676 AssumptionCache *AC,
3677 const Instruction *CxtI,
3678 const DominatorTree *DT) {
3679 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3680 Add, DL, AC, CxtI, DT);
3683 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS,
3684 const DataLayout &DL,
3685 AssumptionCache *AC,
3686 const Instruction *CxtI,
3687 const DominatorTree *DT) {
3688 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3691 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3692 // FIXME: This conservative implementation can be relaxed. E.g. most
3693 // atomic operations are guaranteed to terminate on most platforms
3694 // and most functions terminate.
3696 return !I->isAtomic() && // atomics may never succeed on some platforms
3697 !isa<CallInst>(I) && // could throw and might not terminate
3698 !isa<InvokeInst>(I) && // might not terminate and could throw to
3699 // non-successor (see bug 24185 for details).
3700 !isa<ResumeInst>(I) && // has no successors
3701 !isa<ReturnInst>(I); // has no successors
3704 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3706 // The loop header is guaranteed to be executed for every iteration.
3708 // FIXME: Relax this constraint to cover all basic blocks that are
3709 // guaranteed to be executed at every iteration.
3710 if (I->getParent() != L->getHeader()) return false;
3712 for (const Instruction &LI : *L->getHeader()) {
3713 if (&LI == I) return true;
3714 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3716 llvm_unreachable("Instruction not contained in its own parent basic block.");
3719 bool llvm::propagatesFullPoison(const Instruction *I) {
3720 switch (I->getOpcode()) {
3721 case Instruction::Add:
3722 case Instruction::Sub:
3723 case Instruction::Xor:
3724 case Instruction::Trunc:
3725 case Instruction::BitCast:
3726 case Instruction::AddrSpaceCast:
3727 // These operations all propagate poison unconditionally. Note that poison
3728 // is not any particular value, so xor or subtraction of poison with
3729 // itself still yields poison, not zero.
3732 case Instruction::AShr:
3733 case Instruction::SExt:
3734 // For these operations, one bit of the input is replicated across
3735 // multiple output bits. A replicated poison bit is still poison.
3738 case Instruction::Shl: {
3739 // Left shift *by* a poison value is poison. The number of
3740 // positions to shift is unsigned, so no negative values are
3741 // possible there. Left shift by zero places preserves poison. So
3742 // it only remains to consider left shift of poison by a positive
3743 // number of places.
3745 // A left shift by a positive number of places leaves the lowest order bit
3746 // non-poisoned. However, if such a shift has a no-wrap flag, then we can
3747 // make the poison operand violate that flag, yielding a fresh full-poison
3749 auto *OBO = cast<OverflowingBinaryOperator>(I);
3750 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
3753 case Instruction::Mul: {
3754 // A multiplication by zero yields a non-poison zero result, so we need to
3755 // rule out zero as an operand. Conservatively, multiplication by a
3756 // non-zero constant is not multiplication by zero.
3758 // Multiplication by a non-zero constant can leave some bits
3759 // non-poisoned. For example, a multiplication by 2 leaves the lowest
3760 // order bit unpoisoned. So we need to consider that.
3762 // Multiplication by 1 preserves poison. If the multiplication has a
3763 // no-wrap flag, then we can make the poison operand violate that flag
3764 // when multiplied by any integer other than 0 and 1.
3765 auto *OBO = cast<OverflowingBinaryOperator>(I);
3766 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) {
3767 for (Value *V : OBO->operands()) {
3768 if (auto *CI = dyn_cast<ConstantInt>(V)) {
3769 // A ConstantInt cannot yield poison, so we can assume that it is
3770 // the other operand that is poison.
3771 return !CI->isZero();
3778 case Instruction::GetElementPtr:
3779 // A GEP implicitly represents a sequence of additions, subtractions,
3780 // truncations, sign extensions and multiplications. The multiplications
3781 // are by the non-zero sizes of some set of types, so we do not have to be
3782 // concerned with multiplication by zero. If the GEP is in-bounds, then
3783 // these operations are implicitly no-signed-wrap so poison is propagated
3784 // by the arguments above for Add, Sub, Trunc, SExt and Mul.
3785 return cast<GEPOperator>(I)->isInBounds();
3792 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3793 switch (I->getOpcode()) {
3794 case Instruction::Store:
3795 return cast<StoreInst>(I)->getPointerOperand();
3797 case Instruction::Load:
3798 return cast<LoadInst>(I)->getPointerOperand();
3800 case Instruction::AtomicCmpXchg:
3801 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3803 case Instruction::AtomicRMW:
3804 return cast<AtomicRMWInst>(I)->getPointerOperand();
3806 case Instruction::UDiv:
3807 case Instruction::SDiv:
3808 case Instruction::URem:
3809 case Instruction::SRem:
3810 return I->getOperand(1);
3817 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3818 // We currently only look for uses of poison values within the same basic
3819 // block, as that makes it easier to guarantee that the uses will be
3820 // executed given that PoisonI is executed.
3822 // FIXME: Expand this to consider uses beyond the same basic block. To do
3823 // this, look out for the distinction between post-dominance and strong
3825 const BasicBlock *BB = PoisonI->getParent();
3827 // Set of instructions that we have proved will yield poison if PoisonI
3829 SmallSet<const Value *, 16> YieldsPoison;
3830 YieldsPoison.insert(PoisonI);
3832 for (BasicBlock::const_iterator I = PoisonI->getIterator(), E = BB->end();
3834 if (&*I != PoisonI) {
3835 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&*I);
3836 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) return true;
3837 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
3841 // Mark poison that propagates from I through uses of I.
3842 if (YieldsPoison.count(&*I)) {
3843 for (const User *User : I->users()) {
3844 const Instruction *UserI = cast<Instruction>(User);
3845 if (UserI->getParent() == BB && propagatesFullPoison(UserI))
3846 YieldsPoison.insert(User);
3853 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) {
3857 if (auto *C = dyn_cast<ConstantFP>(V))
3862 static bool isKnownNonZero(Value *V) {
3863 if (auto *C = dyn_cast<ConstantFP>(V))
3864 return !C->isZero();
3868 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
3870 Value *CmpLHS, Value *CmpRHS,
3871 Value *TrueVal, Value *FalseVal,
3872 Value *&LHS, Value *&RHS) {
3876 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
3877 // return inconsistent results between implementations.
3878 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
3879 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
3880 // Therefore we behave conservatively and only proceed if at least one of the
3881 // operands is known to not be zero, or if we don't care about signed zeroes.
3884 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
3885 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
3886 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
3887 !isKnownNonZero(CmpRHS))
3888 return {SPF_UNKNOWN, SPNB_NA, false};
3891 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
3892 bool Ordered = false;
3894 // When given one NaN and one non-NaN input:
3895 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
3896 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
3897 // ordered comparison fails), which could be NaN or non-NaN.
3898 // so here we discover exactly what NaN behavior is required/accepted.
3899 if (CmpInst::isFPPredicate(Pred)) {
3900 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
3901 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
3903 if (LHSSafe && RHSSafe) {
3904 // Both operands are known non-NaN.
3905 NaNBehavior = SPNB_RETURNS_ANY;
3906 } else if (CmpInst::isOrdered(Pred)) {
3907 // An ordered comparison will return false when given a NaN, so it
3911 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
3912 NaNBehavior = SPNB_RETURNS_NAN;
3914 NaNBehavior = SPNB_RETURNS_OTHER;
3916 // Completely unsafe.
3917 return {SPF_UNKNOWN, SPNB_NA, false};
3920 // An unordered comparison will return true when given a NaN, so it
3923 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
3924 NaNBehavior = SPNB_RETURNS_OTHER;
3926 NaNBehavior = SPNB_RETURNS_NAN;
3928 // Completely unsafe.
3929 return {SPF_UNKNOWN, SPNB_NA, false};
3933 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
3934 std::swap(CmpLHS, CmpRHS);
3935 Pred = CmpInst::getSwappedPredicate(Pred);
3936 if (NaNBehavior == SPNB_RETURNS_NAN)
3937 NaNBehavior = SPNB_RETURNS_OTHER;
3938 else if (NaNBehavior == SPNB_RETURNS_OTHER)
3939 NaNBehavior = SPNB_RETURNS_NAN;
3943 // ([if]cmp X, Y) ? X : Y
3944 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
3946 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
3947 case ICmpInst::ICMP_UGT:
3948 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
3949 case ICmpInst::ICMP_SGT:
3950 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
3951 case ICmpInst::ICMP_ULT:
3952 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
3953 case ICmpInst::ICMP_SLT:
3954 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
3955 case FCmpInst::FCMP_UGT:
3956 case FCmpInst::FCMP_UGE:
3957 case FCmpInst::FCMP_OGT:
3958 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
3959 case FCmpInst::FCMP_ULT:
3960 case FCmpInst::FCMP_ULE:
3961 case FCmpInst::FCMP_OLT:
3962 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
3966 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
3967 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
3968 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
3970 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
3971 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
3972 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
3973 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3976 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
3977 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
3978 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
3979 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
3983 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C)
3984 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) {
3985 if (C1->getType() == C2->getType() && ~C1->getValue() == C2->getValue() &&
3986 (match(TrueVal, m_Not(m_Specific(CmpLHS))) ||
3987 match(CmpLHS, m_Not(m_Specific(TrueVal))))) {
3990 return {SPF_SMIN, SPNB_NA, false};
3995 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
3997 return {SPF_UNKNOWN, SPNB_NA, false};
4000 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4001 Instruction::CastOps *CastOp) {
4002 CastInst *CI = dyn_cast<CastInst>(V1);
4003 Constant *C = dyn_cast<Constant>(V2);
4004 CastInst *CI2 = dyn_cast<CastInst>(V2);
4007 *CastOp = CI->getOpcode();
4010 // If V1 and V2 are both the same cast from the same type, we can look
4012 if (CI2->getOpcode() == CI->getOpcode() &&
4013 CI2->getSrcTy() == CI->getSrcTy())
4014 return CI2->getOperand(0);
4020 if (isa<SExtInst>(CI) && CmpI->isSigned()) {
4021 Constant *T = ConstantExpr::getTrunc(C, CI->getSrcTy());
4022 // This is only valid if the truncated value can be sign-extended
4023 // back to the original value.
4024 if (ConstantExpr::getSExt(T, C->getType()) == C)
4028 if (isa<ZExtInst>(CI) && CmpI->isUnsigned())
4029 return ConstantExpr::getTrunc(C, CI->getSrcTy());
4031 if (isa<TruncInst>(CI))
4032 return ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned());
4034 if (isa<FPToUIInst>(CI))
4035 return ConstantExpr::getUIToFP(C, CI->getSrcTy(), true);
4037 if (isa<FPToSIInst>(CI))
4038 return ConstantExpr::getSIToFP(C, CI->getSrcTy(), true);
4040 if (isa<UIToFPInst>(CI))
4041 return ConstantExpr::getFPToUI(C, CI->getSrcTy(), true);
4043 if (isa<SIToFPInst>(CI))
4044 return ConstantExpr::getFPToSI(C, CI->getSrcTy(), true);
4046 if (isa<FPTruncInst>(CI))
4047 return ConstantExpr::getFPExtend(C, CI->getSrcTy(), true);
4049 if (isa<FPExtInst>(CI))
4050 return ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true);
4055 SelectPatternResult llvm::matchSelectPattern(Value *V,
4056 Value *&LHS, Value *&RHS,
4057 Instruction::CastOps *CastOp) {
4058 SelectInst *SI = dyn_cast<SelectInst>(V);
4059 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4061 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4062 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4064 CmpInst::Predicate Pred = CmpI->getPredicate();
4065 Value *CmpLHS = CmpI->getOperand(0);
4066 Value *CmpRHS = CmpI->getOperand(1);
4067 Value *TrueVal = SI->getTrueValue();
4068 Value *FalseVal = SI->getFalseValue();
4070 if (isa<FPMathOperator>(CmpI))
4071 FMF = CmpI->getFastMathFlags();
4074 if (CmpI->isEquality())
4075 return {SPF_UNKNOWN, SPNB_NA, false};
4077 // Deal with type mismatches.
4078 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4079 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
4080 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4081 cast<CastInst>(TrueVal)->getOperand(0), C,
4083 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
4084 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4085 C, cast<CastInst>(FalseVal)->getOperand(0),
4088 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4092 ConstantRange llvm::getConstantRangeFromMetadata(MDNode &Ranges) {
4093 const unsigned NumRanges = Ranges.getNumOperands() / 2;
4094 assert(NumRanges >= 1 && "Must have at least one range!");
4095 assert(Ranges.getNumOperands() % 2 == 0 && "Must be a sequence of pairs");
4097 auto *FirstLow = mdconst::extract<ConstantInt>(Ranges.getOperand(0));
4098 auto *FirstHigh = mdconst::extract<ConstantInt>(Ranges.getOperand(1));
4100 ConstantRange CR(FirstLow->getValue(), FirstHigh->getValue());
4102 for (unsigned i = 1; i < NumRanges; ++i) {
4103 auto *Low = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
4104 auto *High = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
4106 // Note: unionWith will potentially create a range that contains values not
4107 // contained in any of the original N ranges.
4108 CR = CR.unionWith(ConstantRange(Low->getValue(), High->getValue()));
4114 /// Return true if "icmp Pred LHS RHS" is always true.
4115 static bool isTruePredicate(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
4116 const DataLayout &DL, unsigned Depth,
4117 AssumptionCache *AC, const Instruction *CxtI,
4118 const DominatorTree *DT) {
4119 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
4120 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
4127 case CmpInst::ICMP_SLE: {
4130 // LHS s<= LHS +_{nsw} C if C >= 0
4131 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
4132 return !C->isNegative();
4136 case CmpInst::ICMP_ULE: {
4139 // LHS u<= LHS +_{nuw} C for any C
4140 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
4143 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
4144 auto MatchNUWAddsToSameValue = [&](Value *A, Value *B, Value *&X,
4145 const APInt *&CA, const APInt *&CB) {
4146 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
4147 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
4150 // If X & C == 0 then (X | C) == X +_{nuw} C
4151 if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
4152 match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
4153 unsigned BitWidth = CA->getBitWidth();
4154 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4155 computeKnownBits(X, KnownZero, KnownOne, DL, Depth + 1, AC, CxtI, DT);
4157 if ((KnownZero & *CA) == *CA && (KnownZero & *CB) == *CB)
4165 const APInt *CLHS, *CRHS;
4166 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
4167 return CLHS->ule(*CRHS);
4174 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
4175 /// ALHS ARHS" is true.
4176 static bool isImpliedCondOperands(CmpInst::Predicate Pred, Value *ALHS,
4177 Value *ARHS, Value *BLHS, Value *BRHS,
4178 const DataLayout &DL, unsigned Depth,
4179 AssumptionCache *AC, const Instruction *CxtI,
4180 const DominatorTree *DT) {
4185 case CmpInst::ICMP_SLT:
4186 case CmpInst::ICMP_SLE:
4187 return isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth, AC, CxtI,
4189 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth, AC, CxtI,
4192 case CmpInst::ICMP_ULT:
4193 case CmpInst::ICMP_ULE:
4194 return isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth, AC, CxtI,
4196 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth, AC, CxtI,
4201 bool llvm::isImpliedCondition(Value *LHS, Value *RHS, const DataLayout &DL,
4202 unsigned Depth, AssumptionCache *AC,
4203 const Instruction *CxtI,
4204 const DominatorTree *DT) {
4205 assert(LHS->getType() == RHS->getType() && "mismatched type");
4206 Type *OpTy = LHS->getType();
4207 assert(OpTy->getScalarType()->isIntegerTy(1));
4209 // LHS ==> RHS by definition
4210 if (LHS == RHS) return true;
4212 if (OpTy->isVectorTy())
4213 // TODO: extending the code below to handle vectors
4215 assert(OpTy->isIntegerTy(1) && "implied by above");
4217 ICmpInst::Predicate APred, BPred;
4221 if (!match(LHS, m_ICmp(APred, m_Value(ALHS), m_Value(ARHS))) ||
4222 !match(RHS, m_ICmp(BPred, m_Value(BLHS), m_Value(BRHS))))
4226 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth, AC,