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/Loads.h"
22 #include "llvm/Analysis/LoopInfo.h"
23 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
24 #include "llvm/Analysis/VectorUtils.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/GlobalAlias.h"
33 #include "llvm/IR/GlobalVariable.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/LLVMContext.h"
37 #include "llvm/IR/Metadata.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/IR/PatternMatch.h"
40 #include "llvm/IR/Statepoint.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/KnownBits.h"
43 #include "llvm/Support/MathExtras.h"
48 using namespace llvm::PatternMatch;
50 const unsigned MaxDepth = 6;
52 // Controls the number of uses of the value searched for possible
53 // dominating comparisons.
54 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
55 cl::Hidden, cl::init(20));
57 // This optimization is known to cause performance regressions is some cases,
58 // keep it under a temporary flag for now.
60 DontImproveNonNegativePhiBits("dont-improve-non-negative-phi-bits",
61 cl::Hidden, cl::init(true));
63 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
64 /// returns the element type's bitwidth.
65 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
66 if (unsigned BitWidth = Ty->getScalarSizeInBits())
69 return DL.getPointerTypeSizeInBits(Ty);
73 // Simplifying using an assume can only be done in a particular control-flow
74 // context (the context instruction provides that context). If an assume and
75 // the context instruction are not in the same block then the DT helps in
76 // figuring out if we can use it.
80 const Instruction *CxtI;
81 const DominatorTree *DT;
82 // Unlike the other analyses, this may be a nullptr because not all clients
83 // provide it currently.
84 OptimizationRemarkEmitter *ORE;
86 /// Set of assumptions that should be excluded from further queries.
87 /// This is because of the potential for mutual recursion to cause
88 /// computeKnownBits to repeatedly visit the same assume intrinsic. The
89 /// classic case of this is assume(x = y), which will attempt to determine
90 /// bits in x from bits in y, which will attempt to determine bits in y from
91 /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
92 /// isKnownNonZero, which calls computeKnownBits and isKnownToBeAPowerOfTwo
93 /// (all of which can call computeKnownBits), and so on.
94 std::array<const Value *, MaxDepth> Excluded;
97 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
98 const DominatorTree *DT, OptimizationRemarkEmitter *ORE = nullptr)
99 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), NumExcluded(0) {}
101 Query(const Query &Q, const Value *NewExcl)
102 : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE),
103 NumExcluded(Q.NumExcluded) {
104 Excluded = Q.Excluded;
105 Excluded[NumExcluded++] = NewExcl;
106 assert(NumExcluded <= Excluded.size());
109 bool isExcluded(const Value *Value) const {
110 if (NumExcluded == 0)
112 auto End = Excluded.begin() + NumExcluded;
113 return std::find(Excluded.begin(), End, Value) != End;
116 } // end anonymous namespace
118 // Given the provided Value and, potentially, a context instruction, return
119 // the preferred context instruction (if any).
120 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
121 // If we've been provided with a context instruction, then use that (provided
122 // it has been inserted).
123 if (CxtI && CxtI->getParent())
126 // If the value is really an already-inserted instruction, then use that.
127 CxtI = dyn_cast<Instruction>(V);
128 if (CxtI && CxtI->getParent())
134 static void computeKnownBits(const Value *V, KnownBits &Known,
135 unsigned Depth, const Query &Q);
137 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
138 const DataLayout &DL, unsigned Depth,
139 AssumptionCache *AC, const Instruction *CxtI,
140 const DominatorTree *DT,
141 OptimizationRemarkEmitter *ORE) {
142 ::computeKnownBits(V, Known, Depth,
143 Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
146 static KnownBits computeKnownBits(const Value *V, unsigned Depth,
149 KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL,
150 unsigned Depth, AssumptionCache *AC,
151 const Instruction *CxtI,
152 const DominatorTree *DT,
153 OptimizationRemarkEmitter *ORE) {
154 return ::computeKnownBits(V, Depth,
155 Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
158 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
159 const DataLayout &DL,
160 AssumptionCache *AC, const Instruction *CxtI,
161 const DominatorTree *DT) {
162 assert(LHS->getType() == RHS->getType() &&
163 "LHS and RHS should have the same type");
164 assert(LHS->getType()->isIntOrIntVectorTy() &&
165 "LHS and RHS should be integers");
166 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
167 KnownBits LHSKnown(IT->getBitWidth());
168 KnownBits RHSKnown(IT->getBitWidth());
169 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT);
170 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT);
171 return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
175 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
178 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
180 unsigned Depth, AssumptionCache *AC,
181 const Instruction *CxtI,
182 const DominatorTree *DT) {
183 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
184 Query(DL, AC, safeCxtI(V, CxtI), DT));
187 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
189 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
190 AssumptionCache *AC, const Instruction *CxtI,
191 const DominatorTree *DT) {
192 return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
195 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
197 AssumptionCache *AC, const Instruction *CxtI,
198 const DominatorTree *DT) {
199 KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
200 return Known.isNonNegative();
203 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
204 AssumptionCache *AC, const Instruction *CxtI,
205 const DominatorTree *DT) {
206 if (auto *CI = dyn_cast<ConstantInt>(V))
207 return CI->getValue().isStrictlyPositive();
209 // TODO: We'd doing two recursive queries here. We should factor this such
210 // that only a single query is needed.
211 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) &&
212 isKnownNonZero(V, DL, Depth, AC, CxtI, DT);
215 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
216 AssumptionCache *AC, const Instruction *CxtI,
217 const DominatorTree *DT) {
218 KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
219 return Known.isNegative();
222 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
224 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
225 const DataLayout &DL,
226 AssumptionCache *AC, const Instruction *CxtI,
227 const DominatorTree *DT) {
228 return ::isKnownNonEqual(V1, V2, Query(DL, AC,
229 safeCxtI(V1, safeCxtI(V2, CxtI)),
233 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
236 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
237 const DataLayout &DL,
238 unsigned Depth, AssumptionCache *AC,
239 const Instruction *CxtI, const DominatorTree *DT) {
240 return ::MaskedValueIsZero(V, Mask, Depth,
241 Query(DL, AC, safeCxtI(V, CxtI), DT));
244 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
247 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
248 unsigned Depth, AssumptionCache *AC,
249 const Instruction *CxtI,
250 const DominatorTree *DT) {
251 return ::ComputeNumSignBits(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
254 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
256 KnownBits &KnownOut, KnownBits &Known2,
257 unsigned Depth, const Query &Q) {
258 unsigned BitWidth = KnownOut.getBitWidth();
260 // If an initial sequence of bits in the result is not needed, the
261 // corresponding bits in the operands are not needed.
262 KnownBits LHSKnown(BitWidth);
263 computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
264 computeKnownBits(Op1, Known2, Depth + 1, Q);
266 // Carry in a 1 for a subtract, rather than a 0.
267 uint64_t CarryIn = 0;
269 // Sum = LHS + ~RHS + 1
270 std::swap(Known2.Zero, Known2.One);
274 APInt PossibleSumZero = ~LHSKnown.Zero + ~Known2.Zero + CarryIn;
275 APInt PossibleSumOne = LHSKnown.One + Known2.One + CarryIn;
277 // Compute known bits of the carry.
278 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnown.Zero ^ Known2.Zero);
279 APInt CarryKnownOne = PossibleSumOne ^ LHSKnown.One ^ Known2.One;
281 // Compute set of known bits (where all three relevant bits are known).
282 APInt LHSKnownUnion = LHSKnown.Zero | LHSKnown.One;
283 APInt RHSKnownUnion = Known2.Zero | Known2.One;
284 APInt CarryKnownUnion = CarryKnownZero | CarryKnownOne;
285 APInt Known = LHSKnownUnion & RHSKnownUnion & CarryKnownUnion;
287 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
288 "known bits of sum differ");
290 // Compute known bits of the result.
291 KnownOut.Zero = ~PossibleSumOne & Known;
292 KnownOut.One = PossibleSumOne & Known;
294 // Are we still trying to solve for the sign bit?
295 if (!Known.isSignBitSet()) {
297 // Adding two non-negative numbers, or subtracting a negative number from
298 // a non-negative one, can't wrap into negative.
299 if (LHSKnown.isNonNegative() && Known2.isNonNegative())
300 KnownOut.makeNonNegative();
301 // Adding two negative numbers, or subtracting a non-negative number from
302 // a negative one, can't wrap into non-negative.
303 else if (LHSKnown.isNegative() && Known2.isNegative())
304 KnownOut.makeNegative();
309 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
310 KnownBits &Known, KnownBits &Known2,
311 unsigned Depth, const Query &Q) {
312 unsigned BitWidth = Known.getBitWidth();
313 computeKnownBits(Op1, Known, Depth + 1, Q);
314 computeKnownBits(Op0, Known2, Depth + 1, Q);
316 bool isKnownNegative = false;
317 bool isKnownNonNegative = false;
318 // If the multiplication is known not to overflow, compute the sign bit.
321 // The product of a number with itself is non-negative.
322 isKnownNonNegative = true;
324 bool isKnownNonNegativeOp1 = Known.isNonNegative();
325 bool isKnownNonNegativeOp0 = Known2.isNonNegative();
326 bool isKnownNegativeOp1 = Known.isNegative();
327 bool isKnownNegativeOp0 = Known2.isNegative();
328 // The product of two numbers with the same sign is non-negative.
329 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
330 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
331 // The product of a negative number and a non-negative number is either
333 if (!isKnownNonNegative)
334 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
335 isKnownNonZero(Op0, Depth, Q)) ||
336 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
337 isKnownNonZero(Op1, Depth, Q));
341 // If low bits are zero in either operand, output low known-0 bits.
342 // Also compute a conservative estimate for high known-0 bits.
343 // More trickiness is possible, but this is sufficient for the
344 // interesting case of alignment computation.
345 unsigned TrailZ = Known.countMinTrailingZeros() +
346 Known2.countMinTrailingZeros();
347 unsigned LeadZ = std::max(Known.countMinLeadingZeros() +
348 Known2.countMinLeadingZeros(),
349 BitWidth) - BitWidth;
351 TrailZ = std::min(TrailZ, BitWidth);
352 LeadZ = std::min(LeadZ, BitWidth);
354 Known.Zero.setLowBits(TrailZ);
355 Known.Zero.setHighBits(LeadZ);
357 // Only make use of no-wrap flags if we failed to compute the sign bit
358 // directly. This matters if the multiplication always overflows, in
359 // which case we prefer to follow the result of the direct computation,
360 // though as the program is invoking undefined behaviour we can choose
361 // whatever we like here.
362 if (isKnownNonNegative && !Known.isNegative())
363 Known.makeNonNegative();
364 else if (isKnownNegative && !Known.isNonNegative())
365 Known.makeNegative();
368 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
370 unsigned BitWidth = Known.getBitWidth();
371 unsigned NumRanges = Ranges.getNumOperands() / 2;
372 assert(NumRanges >= 1);
374 Known.Zero.setAllBits();
375 Known.One.setAllBits();
377 for (unsigned i = 0; i < NumRanges; ++i) {
379 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
381 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
382 ConstantRange Range(Lower->getValue(), Upper->getValue());
384 // The first CommonPrefixBits of all values in Range are equal.
385 unsigned CommonPrefixBits =
386 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
388 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
389 Known.One &= Range.getUnsignedMax() & Mask;
390 Known.Zero &= ~Range.getUnsignedMax() & Mask;
394 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
395 SmallVector<const Value *, 16> WorkSet(1, I);
396 SmallPtrSet<const Value *, 32> Visited;
397 SmallPtrSet<const Value *, 16> EphValues;
399 // The instruction defining an assumption's condition itself is always
400 // considered ephemeral to that assumption (even if it has other
401 // non-ephemeral users). See r246696's test case for an example.
402 if (is_contained(I->operands(), E))
405 while (!WorkSet.empty()) {
406 const Value *V = WorkSet.pop_back_val();
407 if (!Visited.insert(V).second)
410 // If all uses of this value are ephemeral, then so is this value.
411 if (all_of(V->users(), [&](const User *U) { return EphValues.count(U); })) {
416 if (const User *U = dyn_cast<User>(V))
417 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
419 if (isSafeToSpeculativelyExecute(*J))
420 WorkSet.push_back(*J);
428 // Is this an intrinsic that cannot be speculated but also cannot trap?
429 static bool isAssumeLikeIntrinsic(const Instruction *I) {
430 if (const CallInst *CI = dyn_cast<CallInst>(I))
431 if (Function *F = CI->getCalledFunction())
432 switch (F->getIntrinsicID()) {
434 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
435 case Intrinsic::assume:
436 case Intrinsic::dbg_declare:
437 case Intrinsic::dbg_value:
438 case Intrinsic::invariant_start:
439 case Intrinsic::invariant_end:
440 case Intrinsic::lifetime_start:
441 case Intrinsic::lifetime_end:
442 case Intrinsic::objectsize:
443 case Intrinsic::ptr_annotation:
444 case Intrinsic::var_annotation:
451 bool llvm::isValidAssumeForContext(const Instruction *Inv,
452 const Instruction *CxtI,
453 const DominatorTree *DT) {
455 // There are two restrictions on the use of an assume:
456 // 1. The assume must dominate the context (or the control flow must
457 // reach the assume whenever it reaches the context).
458 // 2. The context must not be in the assume's set of ephemeral values
459 // (otherwise we will use the assume to prove that the condition
460 // feeding the assume is trivially true, thus causing the removal of
464 if (DT->dominates(Inv, CxtI))
466 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
467 // We don't have a DT, but this trivially dominates.
471 // With or without a DT, the only remaining case we will check is if the
472 // instructions are in the same BB. Give up if that is not the case.
473 if (Inv->getParent() != CxtI->getParent())
476 // If we have a dom tree, then we now know that the assume doens't dominate
477 // the other instruction. If we don't have a dom tree then we can check if
478 // the assume is first in the BB.
480 // Search forward from the assume until we reach the context (or the end
481 // of the block); the common case is that the assume will come first.
482 for (auto I = std::next(BasicBlock::const_iterator(Inv)),
483 IE = Inv->getParent()->end(); I != IE; ++I)
488 // The context comes first, but they're both in the same block. Make sure
489 // there is nothing in between that might interrupt the control flow.
490 for (BasicBlock::const_iterator I =
491 std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
493 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
496 return !isEphemeralValueOf(Inv, CxtI);
499 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
500 unsigned Depth, const Query &Q) {
501 // Use of assumptions is context-sensitive. If we don't have a context, we
503 if (!Q.AC || !Q.CxtI)
506 unsigned BitWidth = Known.getBitWidth();
508 // Note that the patterns below need to be kept in sync with the code
509 // in AssumptionCache::updateAffectedValues.
511 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
514 CallInst *I = cast<CallInst>(AssumeVH);
515 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
516 "Got assumption for the wrong function!");
520 // Warning: This loop can end up being somewhat performance sensetive.
521 // We're running this loop for once for each value queried resulting in a
522 // runtime of ~O(#assumes * #values).
524 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
525 "must be an assume intrinsic");
527 Value *Arg = I->getArgOperand(0);
529 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
530 assert(BitWidth == 1 && "assume operand is not i1?");
534 if (match(Arg, m_Not(m_Specific(V))) &&
535 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
536 assert(BitWidth == 1 && "assume operand is not i1?");
541 // The remaining tests are all recursive, so bail out if we hit the limit.
542 if (Depth == MaxDepth)
546 auto m_V = m_CombineOr(m_Specific(V),
547 m_CombineOr(m_PtrToInt(m_Specific(V)),
548 m_BitCast(m_Specific(V))));
550 CmpInst::Predicate Pred;
553 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
554 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
555 KnownBits RHSKnown(BitWidth);
556 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
557 Known.Zero |= RHSKnown.Zero;
558 Known.One |= RHSKnown.One;
560 } else if (match(Arg,
561 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
562 Pred == ICmpInst::ICMP_EQ &&
563 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
564 KnownBits RHSKnown(BitWidth);
565 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
566 KnownBits MaskKnown(BitWidth);
567 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
569 // For those bits in the mask that are known to be one, we can propagate
570 // known bits from the RHS to V.
571 Known.Zero |= RHSKnown.Zero & MaskKnown.One;
572 Known.One |= RHSKnown.One & MaskKnown.One;
573 // assume(~(v & b) = a)
574 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
576 Pred == ICmpInst::ICMP_EQ &&
577 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
578 KnownBits RHSKnown(BitWidth);
579 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
580 KnownBits MaskKnown(BitWidth);
581 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
583 // For those bits in the mask that are known to be one, we can propagate
584 // inverted known bits from the RHS to V.
585 Known.Zero |= RHSKnown.One & MaskKnown.One;
586 Known.One |= RHSKnown.Zero & MaskKnown.One;
588 } else if (match(Arg,
589 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
590 Pred == ICmpInst::ICMP_EQ &&
591 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
592 KnownBits RHSKnown(BitWidth);
593 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
594 KnownBits BKnown(BitWidth);
595 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
597 // For those bits in B that are known to be zero, we can propagate known
598 // bits from the RHS to V.
599 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
600 Known.One |= RHSKnown.One & BKnown.Zero;
601 // assume(~(v | b) = a)
602 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
604 Pred == ICmpInst::ICMP_EQ &&
605 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
606 KnownBits RHSKnown(BitWidth);
607 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
608 KnownBits BKnown(BitWidth);
609 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
611 // For those bits in B that are known to be zero, we can propagate
612 // inverted known bits from the RHS to V.
613 Known.Zero |= RHSKnown.One & BKnown.Zero;
614 Known.One |= RHSKnown.Zero & BKnown.Zero;
616 } else if (match(Arg,
617 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
618 Pred == ICmpInst::ICMP_EQ &&
619 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
620 KnownBits RHSKnown(BitWidth);
621 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
622 KnownBits BKnown(BitWidth);
623 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
625 // For those bits in B that are known to be zero, we can propagate known
626 // bits from the RHS to V. For those bits in B that are known to be one,
627 // we can propagate inverted known bits from the RHS to V.
628 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
629 Known.One |= RHSKnown.One & BKnown.Zero;
630 Known.Zero |= RHSKnown.One & BKnown.One;
631 Known.One |= RHSKnown.Zero & BKnown.One;
632 // assume(~(v ^ b) = a)
633 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
635 Pred == ICmpInst::ICMP_EQ &&
636 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
637 KnownBits RHSKnown(BitWidth);
638 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
639 KnownBits BKnown(BitWidth);
640 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
642 // For those bits in B that are known to be zero, we can propagate
643 // inverted known bits from the RHS to V. For those bits in B that are
644 // known to be one, we can propagate known bits from the RHS to V.
645 Known.Zero |= RHSKnown.One & BKnown.Zero;
646 Known.One |= RHSKnown.Zero & BKnown.Zero;
647 Known.Zero |= RHSKnown.Zero & BKnown.One;
648 Known.One |= RHSKnown.One & BKnown.One;
649 // assume(v << c = a)
650 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
652 Pred == ICmpInst::ICMP_EQ &&
653 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
654 KnownBits RHSKnown(BitWidth);
655 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
656 // For those bits in RHS that are known, we can propagate them to known
657 // bits in V shifted to the right by C.
658 RHSKnown.Zero.lshrInPlace(C->getZExtValue());
659 Known.Zero |= RHSKnown.Zero;
660 RHSKnown.One.lshrInPlace(C->getZExtValue());
661 Known.One |= RHSKnown.One;
662 // assume(~(v << c) = a)
663 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
665 Pred == ICmpInst::ICMP_EQ &&
666 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
667 KnownBits RHSKnown(BitWidth);
668 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
669 // For those bits in RHS that are known, we can propagate them inverted
670 // to known bits in V shifted to the right by C.
671 RHSKnown.One.lshrInPlace(C->getZExtValue());
672 Known.Zero |= RHSKnown.One;
673 RHSKnown.Zero.lshrInPlace(C->getZExtValue());
674 Known.One |= RHSKnown.Zero;
675 // assume(v >> c = a)
676 } else if (match(Arg,
677 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
678 m_AShr(m_V, m_ConstantInt(C))),
680 Pred == ICmpInst::ICMP_EQ &&
681 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
682 KnownBits RHSKnown(BitWidth);
683 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
684 // For those bits in RHS that are known, we can propagate them to known
685 // bits in V shifted to the right by C.
686 Known.Zero |= RHSKnown.Zero << C->getZExtValue();
687 Known.One |= RHSKnown.One << C->getZExtValue();
688 // assume(~(v >> c) = a)
689 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
690 m_LShr(m_V, m_ConstantInt(C)),
691 m_AShr(m_V, m_ConstantInt(C)))),
693 Pred == ICmpInst::ICMP_EQ &&
694 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
695 KnownBits RHSKnown(BitWidth);
696 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
697 // For those bits in RHS that are known, we can propagate them inverted
698 // to known bits in V shifted to the right by C.
699 Known.Zero |= RHSKnown.One << C->getZExtValue();
700 Known.One |= RHSKnown.Zero << C->getZExtValue();
701 // assume(v >=_s c) where c is non-negative
702 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
703 Pred == ICmpInst::ICMP_SGE &&
704 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
705 KnownBits RHSKnown(BitWidth);
706 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
708 if (RHSKnown.isNonNegative()) {
709 // We know that the sign bit is zero.
710 Known.makeNonNegative();
712 // assume(v >_s c) where c is at least -1.
713 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
714 Pred == ICmpInst::ICMP_SGT &&
715 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
716 KnownBits RHSKnown(BitWidth);
717 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
719 if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
720 // We know that the sign bit is zero.
721 Known.makeNonNegative();
723 // assume(v <=_s c) where c is negative
724 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
725 Pred == ICmpInst::ICMP_SLE &&
726 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
727 KnownBits RHSKnown(BitWidth);
728 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
730 if (RHSKnown.isNegative()) {
731 // We know that the sign bit is one.
732 Known.makeNegative();
734 // assume(v <_s c) where c is non-positive
735 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
736 Pred == ICmpInst::ICMP_SLT &&
737 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
738 KnownBits RHSKnown(BitWidth);
739 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
741 if (RHSKnown.isZero() || RHSKnown.isNegative()) {
742 // We know that the sign bit is one.
743 Known.makeNegative();
746 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
747 Pred == ICmpInst::ICMP_ULE &&
748 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
749 KnownBits RHSKnown(BitWidth);
750 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
752 // Whatever high bits in c are zero are known to be zero.
753 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
755 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
756 Pred == ICmpInst::ICMP_ULT &&
757 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
758 KnownBits RHSKnown(BitWidth);
759 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
761 // Whatever high bits in c are zero are known to be zero (if c is a power
762 // of 2, then one more).
763 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
764 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
766 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
770 // If assumptions conflict with each other or previous known bits, then we
771 // have a logical fallacy. It's possible that the assumption is not reachable,
772 // so this isn't a real bug. On the other hand, the program may have undefined
773 // behavior, or we might have a bug in the compiler. We can't assert/crash, so
774 // clear out the known bits, try to warn the user, and hope for the best.
775 if (Known.Zero.intersects(Known.One)) {
779 auto *CxtI = const_cast<Instruction *>(Q.CxtI);
780 OptimizationRemarkAnalysis ORA("value-tracking", "BadAssumption", CxtI);
781 Q.ORE->emit(ORA << "Detected conflicting code assumptions. Program may "
782 "have undefined behavior, or compiler may have "
788 // Compute known bits from a shift operator, including those with a
789 // non-constant shift amount. Known is the outputs of this function. Known2 is a
790 // pre-allocated temporary with the/ same bit width as Known. KZF and KOF are
791 // operator-specific functors that, given the known-zero or known-one bits
792 // respectively, and a shift amount, compute the implied known-zero or known-one
793 // bits of the shift operator's result respectively for that shift amount. The
794 // results from calling KZF and KOF are conservatively combined for all
795 // permitted shift amounts.
796 static void computeKnownBitsFromShiftOperator(
797 const Operator *I, KnownBits &Known, KnownBits &Known2,
798 unsigned Depth, const Query &Q,
799 function_ref<APInt(const APInt &, unsigned)> KZF,
800 function_ref<APInt(const APInt &, unsigned)> KOF) {
801 unsigned BitWidth = Known.getBitWidth();
803 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
804 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
806 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
807 Known.Zero = KZF(Known.Zero, ShiftAmt);
808 Known.One = KOF(Known.One, ShiftAmt);
809 // If there is conflict between Known.Zero and Known.One, this must be an
810 // overflowing left shift, so the shift result is undefined. Clear Known
811 // bits so that other code could propagate this undef.
812 if ((Known.Zero & Known.One) != 0)
818 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
820 // If the shift amount could be greater than or equal to the bit-width of the LHS, the
821 // value could be undef, so we don't know anything about it.
822 if ((~Known.Zero).uge(BitWidth)) {
827 // Note: We cannot use Known.Zero.getLimitedValue() here, because if
828 // BitWidth > 64 and any upper bits are known, we'll end up returning the
829 // limit value (which implies all bits are known).
830 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
831 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
833 // It would be more-clearly correct to use the two temporaries for this
834 // calculation. Reusing the APInts here to prevent unnecessary allocations.
837 // If we know the shifter operand is nonzero, we can sometimes infer more
838 // known bits. However this is expensive to compute, so be lazy about it and
839 // only compute it when absolutely necessary.
840 Optional<bool> ShifterOperandIsNonZero;
842 // Early exit if we can't constrain any well-defined shift amount.
843 if (!(ShiftAmtKZ & (BitWidth - 1)) && !(ShiftAmtKO & (BitWidth - 1))) {
844 ShifterOperandIsNonZero =
845 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
846 if (!*ShifterOperandIsNonZero)
850 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
852 Known.Zero.setAllBits();
853 Known.One.setAllBits();
854 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
855 // Combine the shifted known input bits only for those shift amounts
856 // compatible with its known constraints.
857 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
859 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
861 // If we know the shifter is nonzero, we may be able to infer more known
862 // bits. This check is sunk down as far as possible to avoid the expensive
863 // call to isKnownNonZero if the cheaper checks above fail.
865 if (!ShifterOperandIsNonZero.hasValue())
866 ShifterOperandIsNonZero =
867 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
868 if (*ShifterOperandIsNonZero)
872 Known.Zero &= KZF(Known2.Zero, ShiftAmt);
873 Known.One &= KOF(Known2.One, ShiftAmt);
876 // If there are no compatible shift amounts, then we've proven that the shift
877 // amount must be >= the BitWidth, and the result is undefined. We could
878 // return anything we'd like, but we need to make sure the sets of known bits
879 // stay disjoint (it should be better for some other code to actually
880 // propagate the undef than to pick a value here using known bits).
881 if (Known.Zero.intersects(Known.One))
885 static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
886 unsigned Depth, const Query &Q) {
887 unsigned BitWidth = Known.getBitWidth();
889 KnownBits Known2(Known);
890 switch (I->getOpcode()) {
892 case Instruction::Load:
893 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
894 computeKnownBitsFromRangeMetadata(*MD, Known);
896 case Instruction::And: {
897 // If either the LHS or the RHS are Zero, the result is zero.
898 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
899 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
901 // Output known-1 bits are only known if set in both the LHS & RHS.
902 Known.One &= Known2.One;
903 // Output known-0 are known to be clear if zero in either the LHS | RHS.
904 Known.Zero |= Known2.Zero;
906 // and(x, add (x, -1)) is a common idiom that always clears the low bit;
907 // here we handle the more general case of adding any odd number by
908 // matching the form add(x, add(x, y)) where y is odd.
909 // TODO: This could be generalized to clearing any bit set in y where the
910 // following bit is known to be unset in y.
912 if (!Known.Zero[0] && !Known.One[0] &&
913 (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)),
915 match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)),
918 computeKnownBits(Y, Known2, Depth + 1, Q);
919 if (Known2.countMinTrailingOnes() > 0)
920 Known.Zero.setBit(0);
924 case Instruction::Or: {
925 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
926 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
928 // Output known-0 bits are only known if clear in both the LHS & RHS.
929 Known.Zero &= Known2.Zero;
930 // Output known-1 are known to be set if set in either the LHS | RHS.
931 Known.One |= Known2.One;
934 case Instruction::Xor: {
935 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
936 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
938 // Output known-0 bits are known if clear or set in both the LHS & RHS.
939 APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
940 // Output known-1 are known to be set if set in only one of the LHS, RHS.
941 Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
942 Known.Zero = std::move(KnownZeroOut);
945 case Instruction::Mul: {
946 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
947 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
951 case Instruction::UDiv: {
952 // For the purposes of computing leading zeros we can conservatively
953 // treat a udiv as a logical right shift by the power of 2 known to
954 // be less than the denominator.
955 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
956 unsigned LeadZ = Known2.countMinLeadingZeros();
959 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
960 unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros();
961 if (RHSMaxLeadingZeros != BitWidth)
962 LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1);
964 Known.Zero.setHighBits(LeadZ);
967 case Instruction::Select: {
968 const Value *LHS, *RHS;
969 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
970 if (SelectPatternResult::isMinOrMax(SPF)) {
971 computeKnownBits(RHS, Known, Depth + 1, Q);
972 computeKnownBits(LHS, Known2, Depth + 1, Q);
974 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
975 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
978 unsigned MaxHighOnes = 0;
979 unsigned MaxHighZeros = 0;
980 if (SPF == SPF_SMAX) {
981 // If both sides are negative, the result is negative.
982 if (Known.isNegative() && Known2.isNegative())
983 // We can derive a lower bound on the result by taking the max of the
986 std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
987 // If either side is non-negative, the result is non-negative.
988 else if (Known.isNonNegative() || Known2.isNonNegative())
990 } else if (SPF == SPF_SMIN) {
991 // If both sides are non-negative, the result is non-negative.
992 if (Known.isNonNegative() && Known2.isNonNegative())
993 // We can derive an upper bound on the result by taking the max of the
994 // leading zero bits.
995 MaxHighZeros = std::max(Known.countMinLeadingZeros(),
996 Known2.countMinLeadingZeros());
997 // If either side is negative, the result is negative.
998 else if (Known.isNegative() || Known2.isNegative())
1000 } else if (SPF == SPF_UMAX) {
1001 // We can derive a lower bound on the result by taking the max of the
1002 // leading one bits.
1004 std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
1005 } else if (SPF == SPF_UMIN) {
1006 // We can derive an upper bound on the result by taking the max of the
1007 // leading zero bits.
1009 std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1012 // Only known if known in both the LHS and RHS.
1013 Known.One &= Known2.One;
1014 Known.Zero &= Known2.Zero;
1015 if (MaxHighOnes > 0)
1016 Known.One.setHighBits(MaxHighOnes);
1017 if (MaxHighZeros > 0)
1018 Known.Zero.setHighBits(MaxHighZeros);
1021 case Instruction::FPTrunc:
1022 case Instruction::FPExt:
1023 case Instruction::FPToUI:
1024 case Instruction::FPToSI:
1025 case Instruction::SIToFP:
1026 case Instruction::UIToFP:
1027 break; // Can't work with floating point.
1028 case Instruction::PtrToInt:
1029 case Instruction::IntToPtr:
1030 // Fall through and handle them the same as zext/trunc.
1032 case Instruction::ZExt:
1033 case Instruction::Trunc: {
1034 Type *SrcTy = I->getOperand(0)->getType();
1036 unsigned SrcBitWidth;
1037 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1038 // which fall through here.
1039 SrcBitWidth = Q.DL.getTypeSizeInBits(SrcTy->getScalarType());
1041 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1042 Known = Known.zextOrTrunc(SrcBitWidth);
1043 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1044 Known = Known.zextOrTrunc(BitWidth);
1045 // Any top bits are known to be zero.
1046 if (BitWidth > SrcBitWidth)
1047 Known.Zero.setBitsFrom(SrcBitWidth);
1050 case Instruction::BitCast: {
1051 Type *SrcTy = I->getOperand(0)->getType();
1052 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
1053 // TODO: For now, not handling conversions like:
1054 // (bitcast i64 %x to <2 x i32>)
1055 !I->getType()->isVectorTy()) {
1056 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1061 case Instruction::SExt: {
1062 // Compute the bits in the result that are not present in the input.
1063 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1065 Known = Known.trunc(SrcBitWidth);
1066 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1067 // If the sign bit of the input is known set or clear, then we know the
1068 // top bits of the result.
1069 Known = Known.sext(BitWidth);
1072 case Instruction::Shl: {
1073 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1074 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1075 auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1076 APInt KZResult = KnownZero << ShiftAmt;
1077 KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1078 // If this shift has "nsw" keyword, then the result is either a poison
1079 // value or has the same sign bit as the first operand.
1080 if (NSW && KnownZero.isSignBitSet())
1081 KZResult.setSignBit();
1085 auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1086 APInt KOResult = KnownOne << ShiftAmt;
1087 if (NSW && KnownOne.isSignBitSet())
1088 KOResult.setSignBit();
1092 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1095 case Instruction::LShr: {
1096 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1097 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1098 APInt KZResult = KnownZero.lshr(ShiftAmt);
1099 // High bits known zero.
1100 KZResult.setHighBits(ShiftAmt);
1104 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1105 return KnownOne.lshr(ShiftAmt);
1108 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1111 case Instruction::AShr: {
1112 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1113 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1114 return KnownZero.ashr(ShiftAmt);
1117 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1118 return KnownOne.ashr(ShiftAmt);
1121 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1124 case Instruction::Sub: {
1125 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1126 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1127 Known, Known2, Depth, Q);
1130 case Instruction::Add: {
1131 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1132 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1133 Known, Known2, Depth, Q);
1136 case Instruction::SRem:
1137 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1138 APInt RA = Rem->getValue().abs();
1139 if (RA.isPowerOf2()) {
1140 APInt LowBits = RA - 1;
1141 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1143 // The low bits of the first operand are unchanged by the srem.
1144 Known.Zero = Known2.Zero & LowBits;
1145 Known.One = Known2.One & LowBits;
1147 // If the first operand is non-negative or has all low bits zero, then
1148 // the upper bits are all zero.
1149 if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
1150 Known.Zero |= ~LowBits;
1152 // If the first operand is negative and not all low bits are zero, then
1153 // the upper bits are all one.
1154 if (Known2.isNegative() && LowBits.intersects(Known2.One))
1155 Known.One |= ~LowBits;
1157 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1162 // The sign bit is the LHS's sign bit, except when the result of the
1163 // remainder is zero.
1164 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1165 // If it's known zero, our sign bit is also zero.
1166 if (Known2.isNonNegative())
1167 Known.makeNonNegative();
1170 case Instruction::URem: {
1171 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1172 const APInt &RA = Rem->getValue();
1173 if (RA.isPowerOf2()) {
1174 APInt LowBits = (RA - 1);
1175 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1176 Known.Zero |= ~LowBits;
1177 Known.One &= LowBits;
1182 // Since the result is less than or equal to either operand, any leading
1183 // zero bits in either operand must also exist in the result.
1184 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1185 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1188 std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1190 Known.Zero.setHighBits(Leaders);
1194 case Instruction::Alloca: {
1195 const AllocaInst *AI = cast<AllocaInst>(I);
1196 unsigned Align = AI->getAlignment();
1198 Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1201 Known.Zero.setLowBits(countTrailingZeros(Align));
1204 case Instruction::GetElementPtr: {
1205 // Analyze all of the subscripts of this getelementptr instruction
1206 // to determine if we can prove known low zero bits.
1207 KnownBits LocalKnown(BitWidth);
1208 computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1209 unsigned TrailZ = LocalKnown.countMinTrailingZeros();
1211 gep_type_iterator GTI = gep_type_begin(I);
1212 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1213 Value *Index = I->getOperand(i);
1214 if (StructType *STy = GTI.getStructTypeOrNull()) {
1215 // Handle struct member offset arithmetic.
1217 // Handle case when index is vector zeroinitializer
1218 Constant *CIndex = cast<Constant>(Index);
1219 if (CIndex->isZeroValue())
1222 if (CIndex->getType()->isVectorTy())
1223 Index = CIndex->getSplatValue();
1225 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1226 const StructLayout *SL = Q.DL.getStructLayout(STy);
1227 uint64_t Offset = SL->getElementOffset(Idx);
1228 TrailZ = std::min<unsigned>(TrailZ,
1229 countTrailingZeros(Offset));
1231 // Handle array index arithmetic.
1232 Type *IndexedTy = GTI.getIndexedType();
1233 if (!IndexedTy->isSized()) {
1237 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1238 uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1239 LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1240 computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1241 TrailZ = std::min(TrailZ,
1242 unsigned(countTrailingZeros(TypeSize) +
1243 LocalKnown.countMinTrailingZeros()));
1247 Known.Zero.setLowBits(TrailZ);
1250 case Instruction::PHI: {
1251 const PHINode *P = cast<PHINode>(I);
1252 // Handle the case of a simple two-predecessor recurrence PHI.
1253 // There's a lot more that could theoretically be done here, but
1254 // this is sufficient to catch some interesting cases.
1255 if (P->getNumIncomingValues() == 2) {
1256 for (unsigned i = 0; i != 2; ++i) {
1257 Value *L = P->getIncomingValue(i);
1258 Value *R = P->getIncomingValue(!i);
1259 Operator *LU = dyn_cast<Operator>(L);
1262 unsigned Opcode = LU->getOpcode();
1263 // Check for operations that have the property that if
1264 // both their operands have low zero bits, the result
1265 // will have low zero bits.
1266 if (Opcode == Instruction::Add ||
1267 Opcode == Instruction::Sub ||
1268 Opcode == Instruction::And ||
1269 Opcode == Instruction::Or ||
1270 Opcode == Instruction::Mul) {
1271 Value *LL = LU->getOperand(0);
1272 Value *LR = LU->getOperand(1);
1273 // Find a recurrence.
1280 // Ok, we have a PHI of the form L op= R. Check for low
1282 computeKnownBits(R, Known2, Depth + 1, Q);
1284 // We need to take the minimum number of known bits
1285 KnownBits Known3(Known);
1286 computeKnownBits(L, Known3, Depth + 1, Q);
1288 Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
1289 Known3.countMinTrailingZeros()));
1291 if (DontImproveNonNegativePhiBits)
1294 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1295 if (OverflowOp && OverflowOp->hasNoSignedWrap()) {
1296 // If initial value of recurrence is nonnegative, and we are adding
1297 // a nonnegative number with nsw, the result can only be nonnegative
1298 // or poison value regardless of the number of times we execute the
1299 // add in phi recurrence. If initial value is negative and we are
1300 // adding a negative number with nsw, the result can only be
1301 // negative or poison value. Similar arguments apply to sub and mul.
1303 // (add non-negative, non-negative) --> non-negative
1304 // (add negative, negative) --> negative
1305 if (Opcode == Instruction::Add) {
1306 if (Known2.isNonNegative() && Known3.isNonNegative())
1307 Known.makeNonNegative();
1308 else if (Known2.isNegative() && Known3.isNegative())
1309 Known.makeNegative();
1312 // (sub nsw non-negative, negative) --> non-negative
1313 // (sub nsw negative, non-negative) --> negative
1314 else if (Opcode == Instruction::Sub && LL == I) {
1315 if (Known2.isNonNegative() && Known3.isNegative())
1316 Known.makeNonNegative();
1317 else if (Known2.isNegative() && Known3.isNonNegative())
1318 Known.makeNegative();
1321 // (mul nsw non-negative, non-negative) --> non-negative
1322 else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1323 Known3.isNonNegative())
1324 Known.makeNonNegative();
1332 // Unreachable blocks may have zero-operand PHI nodes.
1333 if (P->getNumIncomingValues() == 0)
1336 // Otherwise take the unions of the known bit sets of the operands,
1337 // taking conservative care to avoid excessive recursion.
1338 if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1339 // Skip if every incoming value references to ourself.
1340 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1343 Known.Zero.setAllBits();
1344 Known.One.setAllBits();
1345 for (Value *IncValue : P->incoming_values()) {
1346 // Skip direct self references.
1347 if (IncValue == P) continue;
1349 Known2 = KnownBits(BitWidth);
1350 // Recurse, but cap the recursion to one level, because we don't
1351 // want to waste time spinning around in loops.
1352 computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
1353 Known.Zero &= Known2.Zero;
1354 Known.One &= Known2.One;
1355 // If all bits have been ruled out, there's no need to check
1357 if (!Known.Zero && !Known.One)
1363 case Instruction::Call:
1364 case Instruction::Invoke:
1365 // If range metadata is attached to this call, set known bits from that,
1366 // and then intersect with known bits based on other properties of the
1368 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1369 computeKnownBitsFromRangeMetadata(*MD, Known);
1370 if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1371 computeKnownBits(RV, Known2, Depth + 1, Q);
1372 Known.Zero |= Known2.Zero;
1373 Known.One |= Known2.One;
1375 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1376 switch (II->getIntrinsicID()) {
1378 case Intrinsic::bitreverse:
1379 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1380 Known.Zero |= Known2.Zero.reverseBits();
1381 Known.One |= Known2.One.reverseBits();
1383 case Intrinsic::bswap:
1384 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1385 Known.Zero |= Known2.Zero.byteSwap();
1386 Known.One |= Known2.One.byteSwap();
1388 case Intrinsic::ctlz: {
1389 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1390 // If we have a known 1, its position is our upper bound.
1391 unsigned PossibleLZ = Known2.One.countLeadingZeros();
1392 // If this call is undefined for 0, the result will be less than 2^n.
1393 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1394 PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1395 unsigned LowBits = Log2_32(PossibleLZ)+1;
1396 Known.Zero.setBitsFrom(LowBits);
1399 case Intrinsic::cttz: {
1400 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1401 // If we have a known 1, its position is our upper bound.
1402 unsigned PossibleTZ = Known2.One.countTrailingZeros();
1403 // If this call is undefined for 0, the result will be less than 2^n.
1404 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1405 PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1406 unsigned LowBits = Log2_32(PossibleTZ)+1;
1407 Known.Zero.setBitsFrom(LowBits);
1410 case Intrinsic::ctpop: {
1411 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1412 // We can bound the space the count needs. Also, bits known to be zero
1413 // can't contribute to the population.
1414 unsigned BitsPossiblySet = Known2.countMaxPopulation();
1415 unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1416 Known.Zero.setBitsFrom(LowBits);
1417 // TODO: we could bound KnownOne using the lower bound on the number
1418 // of bits which might be set provided by popcnt KnownOne2.
1421 case Intrinsic::x86_sse42_crc32_64_64:
1422 Known.Zero.setBitsFrom(32);
1427 case Instruction::ExtractElement:
1428 // Look through extract element. At the moment we keep this simple and skip
1429 // tracking the specific element. But at least we might find information
1430 // valid for all elements of the vector (for example if vector is sign
1431 // extended, shifted, etc).
1432 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1434 case Instruction::ExtractValue:
1435 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1436 const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1437 if (EVI->getNumIndices() != 1) break;
1438 if (EVI->getIndices()[0] == 0) {
1439 switch (II->getIntrinsicID()) {
1441 case Intrinsic::uadd_with_overflow:
1442 case Intrinsic::sadd_with_overflow:
1443 computeKnownBitsAddSub(true, II->getArgOperand(0),
1444 II->getArgOperand(1), false, Known, Known2,
1447 case Intrinsic::usub_with_overflow:
1448 case Intrinsic::ssub_with_overflow:
1449 computeKnownBitsAddSub(false, II->getArgOperand(0),
1450 II->getArgOperand(1), false, Known, Known2,
1453 case Intrinsic::umul_with_overflow:
1454 case Intrinsic::smul_with_overflow:
1455 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1456 Known, Known2, Depth, Q);
1464 /// Determine which bits of V are known to be either zero or one and return
1466 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1467 KnownBits Known(getBitWidth(V->getType(), Q.DL));
1468 computeKnownBits(V, Known, Depth, Q);
1472 /// Determine which bits of V are known to be either zero or one and return
1473 /// them in the Known bit set.
1475 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1476 /// we cannot optimize based on the assumption that it is zero without changing
1477 /// it to be an explicit zero. If we don't change it to zero, other code could
1478 /// optimized based on the contradictory assumption that it is non-zero.
1479 /// Because instcombine aggressively folds operations with undef args anyway,
1480 /// this won't lose us code quality.
1482 /// This function is defined on values with integer type, values with pointer
1483 /// type, and vectors of integers. In the case
1484 /// where V is a vector, known zero, and known one values are the
1485 /// same width as the vector element, and the bit is set only if it is true
1486 /// for all of the elements in the vector.
1487 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1489 assert(V && "No Value?");
1490 assert(Depth <= MaxDepth && "Limit Search Depth");
1491 unsigned BitWidth = Known.getBitWidth();
1493 assert((V->getType()->isIntOrIntVectorTy() ||
1494 V->getType()->getScalarType()->isPointerTy()) &&
1495 "Not integer or pointer type!");
1496 assert((Q.DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1497 (!V->getType()->isIntOrIntVectorTy() ||
1498 V->getType()->getScalarSizeInBits() == BitWidth) &&
1499 "V and Known should have same BitWidth");
1503 if (match(V, m_APInt(C))) {
1504 // We know all of the bits for a scalar constant or a splat vector constant!
1506 Known.Zero = ~Known.One;
1509 // Null and aggregate-zero are all-zeros.
1510 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1514 // Handle a constant vector by taking the intersection of the known bits of
1516 if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1517 // We know that CDS must be a vector of integers. Take the intersection of
1519 Known.Zero.setAllBits(); Known.One.setAllBits();
1520 APInt Elt(BitWidth, 0);
1521 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1522 Elt = CDS->getElementAsInteger(i);
1529 if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1530 // We know that CV must be a vector of integers. Take the intersection of
1532 Known.Zero.setAllBits(); Known.One.setAllBits();
1533 APInt Elt(BitWidth, 0);
1534 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1535 Constant *Element = CV->getAggregateElement(i);
1536 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1541 Elt = ElementCI->getValue();
1548 // Start out not knowing anything.
1551 // We can't imply anything about undefs.
1552 if (isa<UndefValue>(V))
1555 // There's no point in looking through other users of ConstantData for
1556 // assumptions. Confirm that we've handled them all.
1557 assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1559 // Limit search depth.
1560 // All recursive calls that increase depth must come after this.
1561 if (Depth == MaxDepth)
1564 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1565 // the bits of its aliasee.
1566 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1567 if (!GA->isInterposable())
1568 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1572 if (const Operator *I = dyn_cast<Operator>(V))
1573 computeKnownBitsFromOperator(I, Known, Depth, Q);
1575 // Aligned pointers have trailing zeros - refine Known.Zero set
1576 if (V->getType()->isPointerTy()) {
1577 unsigned Align = V->getPointerAlignment(Q.DL);
1579 Known.Zero.setLowBits(countTrailingZeros(Align));
1582 // computeKnownBitsFromAssume strictly refines Known.
1583 // Therefore, we run them after computeKnownBitsFromOperator.
1585 // Check whether a nearby assume intrinsic can determine some known bits.
1586 computeKnownBitsFromAssume(V, Known, Depth, Q);
1588 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1591 /// Return true if the given value is known to have exactly one
1592 /// bit set when defined. For vectors return true if every element is known to
1593 /// be a power of two when defined. Supports values with integer or pointer
1594 /// types and vectors of integers.
1595 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1597 if (const Constant *C = dyn_cast<Constant>(V)) {
1598 if (C->isNullValue())
1601 const APInt *ConstIntOrConstSplatInt;
1602 if (match(C, m_APInt(ConstIntOrConstSplatInt)))
1603 return ConstIntOrConstSplatInt->isPowerOf2();
1606 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1607 // it is shifted off the end then the result is undefined.
1608 if (match(V, m_Shl(m_One(), m_Value())))
1611 // (signmask) >>l X is clearly a power of two if the one is not shifted off
1612 // the bottom. If it is shifted off the bottom then the result is undefined.
1613 if (match(V, m_LShr(m_SignMask(), m_Value())))
1616 // The remaining tests are all recursive, so bail out if we hit the limit.
1617 if (Depth++ == MaxDepth)
1620 Value *X = nullptr, *Y = nullptr;
1621 // A shift left or a logical shift right of a power of two is a power of two
1623 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1624 match(V, m_LShr(m_Value(X), m_Value()))))
1625 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1627 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1628 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1630 if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1631 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1632 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1634 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1635 // A power of two and'd with anything is a power of two or zero.
1636 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1637 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1639 // X & (-X) is always a power of two or zero.
1640 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1645 // Adding a power-of-two or zero to the same power-of-two or zero yields
1646 // either the original power-of-two, a larger power-of-two or zero.
1647 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1648 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1649 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1650 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1651 match(X, m_And(m_Value(), m_Specific(Y))))
1652 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1654 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1655 match(Y, m_And(m_Value(), m_Specific(X))))
1656 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1659 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1660 KnownBits LHSBits(BitWidth);
1661 computeKnownBits(X, LHSBits, Depth, Q);
1663 KnownBits RHSBits(BitWidth);
1664 computeKnownBits(Y, RHSBits, Depth, Q);
1665 // If i8 V is a power of two or zero:
1666 // ZeroBits: 1 1 1 0 1 1 1 1
1667 // ~ZeroBits: 0 0 0 1 0 0 0 0
1668 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1669 // If OrZero isn't set, we cannot give back a zero result.
1670 // Make sure either the LHS or RHS has a bit set.
1671 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1676 // An exact divide or right shift can only shift off zero bits, so the result
1677 // is a power of two only if the first operand is a power of two and not
1678 // copying a sign bit (sdiv int_min, 2).
1679 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1680 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1681 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1688 /// \brief Test whether a GEP's result is known to be non-null.
1690 /// Uses properties inherent in a GEP to try to determine whether it is known
1693 /// Currently this routine does not support vector GEPs.
1694 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1696 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1699 // FIXME: Support vector-GEPs.
1700 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1702 // If the base pointer is non-null, we cannot walk to a null address with an
1703 // inbounds GEP in address space zero.
1704 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1707 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1708 // If so, then the GEP cannot produce a null pointer, as doing so would
1709 // inherently violate the inbounds contract within address space zero.
1710 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1711 GTI != GTE; ++GTI) {
1712 // Struct types are easy -- they must always be indexed by a constant.
1713 if (StructType *STy = GTI.getStructTypeOrNull()) {
1714 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1715 unsigned ElementIdx = OpC->getZExtValue();
1716 const StructLayout *SL = Q.DL.getStructLayout(STy);
1717 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1718 if (ElementOffset > 0)
1723 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1724 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1727 // Fast path the constant operand case both for efficiency and so we don't
1728 // increment Depth when just zipping down an all-constant GEP.
1729 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1735 // We post-increment Depth here because while isKnownNonZero increments it
1736 // as well, when we pop back up that increment won't persist. We don't want
1737 // to recurse 10k times just because we have 10k GEP operands. We don't
1738 // bail completely out because we want to handle constant GEPs regardless
1740 if (Depth++ >= MaxDepth)
1743 if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1750 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1751 /// ensure that the value it's attached to is never Value? 'RangeType' is
1752 /// is the type of the value described by the range.
1753 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1754 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1755 assert(NumRanges >= 1);
1756 for (unsigned i = 0; i < NumRanges; ++i) {
1757 ConstantInt *Lower =
1758 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1759 ConstantInt *Upper =
1760 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1761 ConstantRange Range(Lower->getValue(), Upper->getValue());
1762 if (Range.contains(Value))
1768 /// Return true if the given value is known to be non-zero when defined. For
1769 /// vectors, return true if every element is known to be non-zero when
1770 /// defined. For pointers, if the context instruction and dominator tree are
1771 /// specified, perform context-sensitive analysis and return true if the
1772 /// pointer couldn't possibly be null at the specified instruction.
1773 /// Supports values with integer or pointer type and vectors of integers.
1774 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1775 if (auto *C = dyn_cast<Constant>(V)) {
1776 if (C->isNullValue())
1778 if (isa<ConstantInt>(C))
1779 // Must be non-zero due to null test above.
1782 // For constant vectors, check that all elements are undefined or known
1783 // non-zero to determine that the whole vector is known non-zero.
1784 if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1785 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1786 Constant *Elt = C->getAggregateElement(i);
1787 if (!Elt || Elt->isNullValue())
1789 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1798 if (auto *I = dyn_cast<Instruction>(V)) {
1799 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1800 // If the possible ranges don't contain zero, then the value is
1801 // definitely non-zero.
1802 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1803 const APInt ZeroValue(Ty->getBitWidth(), 0);
1804 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1810 // The remaining tests are all recursive, so bail out if we hit the limit.
1811 if (Depth++ >= MaxDepth)
1814 // Check for pointer simplifications.
1815 if (V->getType()->isPointerTy()) {
1816 if (isKnownNonNullAt(V, Q.CxtI, Q.DT))
1818 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1819 if (isGEPKnownNonNull(GEP, Depth, Q))
1823 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
1825 // X | Y != 0 if X != 0 or Y != 0.
1826 Value *X = nullptr, *Y = nullptr;
1827 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1828 return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
1830 // ext X != 0 if X != 0.
1831 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1832 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
1834 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1835 // if the lowest bit is shifted off the end.
1836 if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1837 // shl nuw can't remove any non-zero bits.
1838 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1839 if (BO->hasNoUnsignedWrap())
1840 return isKnownNonZero(X, Depth, Q);
1842 KnownBits Known(BitWidth);
1843 computeKnownBits(X, Known, Depth, Q);
1847 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1848 // defined if the sign bit is shifted off the end.
1849 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1850 // shr exact can only shift out zero bits.
1851 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1853 return isKnownNonZero(X, Depth, Q);
1855 KnownBits Known = computeKnownBits(X, Depth, Q);
1856 if (Known.isNegative())
1859 // If the shifter operand is a constant, and all of the bits shifted
1860 // out are known to be zero, and X is known non-zero then at least one
1861 // non-zero bit must remain.
1862 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1863 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1864 // Is there a known one in the portion not shifted out?
1865 if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
1867 // Are all the bits to be shifted out known zero?
1868 if (Known.countMinTrailingZeros() >= ShiftVal)
1869 return isKnownNonZero(X, Depth, Q);
1872 // div exact can only produce a zero if the dividend is zero.
1873 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1874 return isKnownNonZero(X, Depth, Q);
1877 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1878 KnownBits XKnown = computeKnownBits(X, Depth, Q);
1879 KnownBits YKnown = computeKnownBits(Y, Depth, Q);
1881 // If X and Y are both non-negative (as signed values) then their sum is not
1882 // zero unless both X and Y are zero.
1883 if (XKnown.isNonNegative() && YKnown.isNonNegative())
1884 if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
1887 // If X and Y are both negative (as signed values) then their sum is not
1888 // zero unless both X and Y equal INT_MIN.
1889 if (XKnown.isNegative() && YKnown.isNegative()) {
1890 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1891 // The sign bit of X is set. If some other bit is set then X is not equal
1893 if (XKnown.One.intersects(Mask))
1895 // The sign bit of Y is set. If some other bit is set then Y is not equal
1897 if (YKnown.One.intersects(Mask))
1901 // The sum of a non-negative number and a power of two is not zero.
1902 if (XKnown.isNonNegative() &&
1903 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
1905 if (YKnown.isNonNegative() &&
1906 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
1910 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1911 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1912 // If X and Y are non-zero then so is X * Y as long as the multiplication
1913 // does not overflow.
1914 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1915 isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
1918 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1919 else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
1920 if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
1921 isKnownNonZero(SI->getFalseValue(), Depth, Q))
1925 else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
1926 // Try and detect a recurrence that monotonically increases from a
1927 // starting value, as these are common as induction variables.
1928 if (PN->getNumIncomingValues() == 2) {
1929 Value *Start = PN->getIncomingValue(0);
1930 Value *Induction = PN->getIncomingValue(1);
1931 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
1932 std::swap(Start, Induction);
1933 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
1934 if (!C->isZero() && !C->isNegative()) {
1936 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
1937 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
1943 // Check if all incoming values are non-zero constant.
1944 bool AllNonZeroConstants = all_of(PN->operands(), [](Value *V) {
1945 return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZeroValue();
1947 if (AllNonZeroConstants)
1951 KnownBits Known(BitWidth);
1952 computeKnownBits(V, Known, Depth, Q);
1953 return Known.One != 0;
1956 /// Return true if V2 == V1 + X, where X is known non-zero.
1957 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
1958 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
1959 if (!BO || BO->getOpcode() != Instruction::Add)
1961 Value *Op = nullptr;
1962 if (V2 == BO->getOperand(0))
1963 Op = BO->getOperand(1);
1964 else if (V2 == BO->getOperand(1))
1965 Op = BO->getOperand(0);
1968 return isKnownNonZero(Op, 0, Q);
1971 /// Return true if it is known that V1 != V2.
1972 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
1973 if (V1->getType()->isVectorTy() || V1 == V2)
1975 if (V1->getType() != V2->getType())
1976 // We can't look through casts yet.
1978 if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
1981 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
1982 // Are any known bits in V1 contradictory to known bits in V2? If V1
1983 // has a known zero where V2 has a known one, they must not be equal.
1984 auto BitWidth = Ty->getBitWidth();
1985 KnownBits Known1(BitWidth);
1986 computeKnownBits(V1, Known1, 0, Q);
1987 KnownBits Known2(BitWidth);
1988 computeKnownBits(V2, Known2, 0, Q);
1990 APInt OppositeBits = (Known1.Zero & Known2.One) |
1991 (Known2.Zero & Known1.One);
1992 if (OppositeBits.getBoolValue())
1998 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1999 /// simplify operations downstream. Mask is known to be zero for bits that V
2002 /// This function is defined on values with integer type, values with pointer
2003 /// type, and vectors of integers. In the case
2004 /// where V is a vector, the mask, known zero, and known one values are the
2005 /// same width as the vector element, and the bit is set only if it is true
2006 /// for all of the elements in the vector.
2007 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2009 KnownBits Known(Mask.getBitWidth());
2010 computeKnownBits(V, Known, Depth, Q);
2011 return Mask.isSubsetOf(Known.Zero);
2014 /// For vector constants, loop over the elements and find the constant with the
2015 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2016 /// or if any element was not analyzed; otherwise, return the count for the
2017 /// element with the minimum number of sign bits.
2018 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2020 const auto *CV = dyn_cast<Constant>(V);
2021 if (!CV || !CV->getType()->isVectorTy())
2024 unsigned MinSignBits = TyBits;
2025 unsigned NumElts = CV->getType()->getVectorNumElements();
2026 for (unsigned i = 0; i != NumElts; ++i) {
2027 // If we find a non-ConstantInt, bail out.
2028 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2032 // If the sign bit is 1, flip the bits, so we always count leading zeros.
2033 APInt EltVal = Elt->getValue();
2034 if (EltVal.isNegative())
2036 MinSignBits = std::min(MinSignBits, EltVal.countLeadingZeros());
2042 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2045 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2047 unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2048 assert(Result > 0 && "At least one sign bit needs to be present!");
2052 /// Return the number of times the sign bit of the register is replicated into
2053 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2054 /// (itself), but other cases can give us information. For example, immediately
2055 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2056 /// other, so we return 3. For vectors, return the number of sign bits for the
2057 /// vector element with the mininum number of known sign bits.
2058 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2061 // We return the minimum number of sign bits that are guaranteed to be present
2062 // in V, so for undef we have to conservatively return 1. We don't have the
2063 // same behavior for poison though -- that's a FIXME today.
2065 unsigned TyBits = Q.DL.getTypeSizeInBits(V->getType()->getScalarType());
2067 unsigned FirstAnswer = 1;
2069 // Note that ConstantInt is handled by the general computeKnownBits case
2072 if (Depth == MaxDepth)
2073 return 1; // Limit search depth.
2075 const Operator *U = dyn_cast<Operator>(V);
2076 switch (Operator::getOpcode(V)) {
2078 case Instruction::SExt:
2079 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2080 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2082 case Instruction::SDiv: {
2083 const APInt *Denominator;
2084 // sdiv X, C -> adds log(C) sign bits.
2085 if (match(U->getOperand(1), m_APInt(Denominator))) {
2087 // Ignore non-positive denominator.
2088 if (!Denominator->isStrictlyPositive())
2091 // Calculate the incoming numerator bits.
2092 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2094 // Add floor(log(C)) bits to the numerator bits.
2095 return std::min(TyBits, NumBits + Denominator->logBase2());
2100 case Instruction::SRem: {
2101 const APInt *Denominator;
2102 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2103 // positive constant. This let us put a lower bound on the number of sign
2105 if (match(U->getOperand(1), m_APInt(Denominator))) {
2107 // Ignore non-positive denominator.
2108 if (!Denominator->isStrictlyPositive())
2111 // Calculate the incoming numerator bits. SRem by a positive constant
2112 // can't lower the number of sign bits.
2114 ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2116 // Calculate the leading sign bit constraints by examining the
2117 // denominator. Given that the denominator is positive, there are two
2120 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2121 // (1 << ceilLogBase2(C)).
2123 // 2. the numerator is negative. Then the result range is (-C,0] and
2124 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2126 // Thus a lower bound on the number of sign bits is `TyBits -
2127 // ceilLogBase2(C)`.
2129 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2130 return std::max(NumrBits, ResBits);
2135 case Instruction::AShr: {
2136 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2137 // ashr X, C -> adds C sign bits. Vectors too.
2139 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2140 unsigned ShAmtLimited = ShAmt->getZExtValue();
2141 if (ShAmtLimited >= TyBits)
2142 break; // Bad shift.
2143 Tmp += ShAmtLimited;
2144 if (Tmp > TyBits) Tmp = TyBits;
2148 case Instruction::Shl: {
2150 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2151 // shl destroys sign bits.
2152 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2153 Tmp2 = ShAmt->getZExtValue();
2154 if (Tmp2 >= TyBits || // Bad shift.
2155 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2160 case Instruction::And:
2161 case Instruction::Or:
2162 case Instruction::Xor: // NOT is handled here.
2163 // Logical binary ops preserve the number of sign bits at the worst.
2164 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2166 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2167 FirstAnswer = std::min(Tmp, Tmp2);
2168 // We computed what we know about the sign bits as our first
2169 // answer. Now proceed to the generic code that uses
2170 // computeKnownBits, and pick whichever answer is better.
2174 case Instruction::Select:
2175 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2176 if (Tmp == 1) return 1; // Early out.
2177 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2178 return std::min(Tmp, Tmp2);
2180 case Instruction::Add:
2181 // Add can have at most one carry bit. Thus we know that the output
2182 // is, at worst, one more bit than the inputs.
2183 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2184 if (Tmp == 1) return 1; // Early out.
2186 // Special case decrementing a value (ADD X, -1):
2187 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2188 if (CRHS->isAllOnesValue()) {
2189 KnownBits Known(TyBits);
2190 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2192 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2194 if ((Known.Zero | 1).isAllOnesValue())
2197 // If we are subtracting one from a positive number, there is no carry
2198 // out of the result.
2199 if (Known.isNonNegative())
2203 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2204 if (Tmp2 == 1) return 1;
2205 return std::min(Tmp, Tmp2)-1;
2207 case Instruction::Sub:
2208 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2209 if (Tmp2 == 1) return 1;
2212 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2213 if (CLHS->isNullValue()) {
2214 KnownBits Known(TyBits);
2215 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2216 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2218 if ((Known.Zero | 1).isAllOnesValue())
2221 // If the input is known to be positive (the sign bit is known clear),
2222 // the output of the NEG has the same number of sign bits as the input.
2223 if (Known.isNonNegative())
2226 // Otherwise, we treat this like a SUB.
2229 // Sub can have at most one carry bit. Thus we know that the output
2230 // is, at worst, one more bit than the inputs.
2231 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2232 if (Tmp == 1) return 1; // Early out.
2233 return std::min(Tmp, Tmp2)-1;
2235 case Instruction::PHI: {
2236 const PHINode *PN = cast<PHINode>(U);
2237 unsigned NumIncomingValues = PN->getNumIncomingValues();
2238 // Don't analyze large in-degree PHIs.
2239 if (NumIncomingValues > 4) break;
2240 // Unreachable blocks may have zero-operand PHI nodes.
2241 if (NumIncomingValues == 0) break;
2243 // Take the minimum of all incoming values. This can't infinitely loop
2244 // because of our depth threshold.
2245 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2246 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2247 if (Tmp == 1) return Tmp;
2249 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2254 case Instruction::Trunc:
2255 // FIXME: it's tricky to do anything useful for this, but it is an important
2256 // case for targets like X86.
2259 case Instruction::ExtractElement:
2260 // Look through extract element. At the moment we keep this simple and skip
2261 // tracking the specific element. But at least we might find information
2262 // valid for all elements of the vector (for example if vector is sign
2263 // extended, shifted, etc).
2264 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2267 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2268 // use this information.
2270 // If we can examine all elements of a vector constant successfully, we're
2271 // done (we can't do any better than that). If not, keep trying.
2272 if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2275 KnownBits Known(TyBits);
2276 computeKnownBits(V, Known, Depth, Q);
2278 // If we know that the sign bit is either zero or one, determine the number of
2279 // identical bits in the top of the input value.
2280 return std::max(FirstAnswer, Known.countMinSignBits());
2283 /// This function computes the integer multiple of Base that equals V.
2284 /// If successful, it returns true and returns the multiple in
2285 /// Multiple. If unsuccessful, it returns false. It looks
2286 /// through SExt instructions only if LookThroughSExt is true.
2287 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2288 bool LookThroughSExt, unsigned Depth) {
2289 const unsigned MaxDepth = 6;
2291 assert(V && "No Value?");
2292 assert(Depth <= MaxDepth && "Limit Search Depth");
2293 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2295 Type *T = V->getType();
2297 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2307 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2308 Constant *BaseVal = ConstantInt::get(T, Base);
2309 if (CO && CO == BaseVal) {
2311 Multiple = ConstantInt::get(T, 1);
2315 if (CI && CI->getZExtValue() % Base == 0) {
2316 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2320 if (Depth == MaxDepth) return false; // Limit search depth.
2322 Operator *I = dyn_cast<Operator>(V);
2323 if (!I) return false;
2325 switch (I->getOpcode()) {
2327 case Instruction::SExt:
2328 if (!LookThroughSExt) return false;
2329 // otherwise fall through to ZExt
2330 case Instruction::ZExt:
2331 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2332 LookThroughSExt, Depth+1);
2333 case Instruction::Shl:
2334 case Instruction::Mul: {
2335 Value *Op0 = I->getOperand(0);
2336 Value *Op1 = I->getOperand(1);
2338 if (I->getOpcode() == Instruction::Shl) {
2339 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2340 if (!Op1CI) return false;
2341 // Turn Op0 << Op1 into Op0 * 2^Op1
2342 APInt Op1Int = Op1CI->getValue();
2343 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2344 APInt API(Op1Int.getBitWidth(), 0);
2345 API.setBit(BitToSet);
2346 Op1 = ConstantInt::get(V->getContext(), API);
2349 Value *Mul0 = nullptr;
2350 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2351 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2352 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2353 if (Op1C->getType()->getPrimitiveSizeInBits() <
2354 MulC->getType()->getPrimitiveSizeInBits())
2355 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2356 if (Op1C->getType()->getPrimitiveSizeInBits() >
2357 MulC->getType()->getPrimitiveSizeInBits())
2358 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2360 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2361 Multiple = ConstantExpr::getMul(MulC, Op1C);
2365 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2366 if (Mul0CI->getValue() == 1) {
2367 // V == Base * Op1, so return Op1
2373 Value *Mul1 = nullptr;
2374 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2375 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2376 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2377 if (Op0C->getType()->getPrimitiveSizeInBits() <
2378 MulC->getType()->getPrimitiveSizeInBits())
2379 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2380 if (Op0C->getType()->getPrimitiveSizeInBits() >
2381 MulC->getType()->getPrimitiveSizeInBits())
2382 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2384 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2385 Multiple = ConstantExpr::getMul(MulC, Op0C);
2389 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2390 if (Mul1CI->getValue() == 1) {
2391 // V == Base * Op0, so return Op0
2399 // We could not determine if V is a multiple of Base.
2403 Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
2404 const TargetLibraryInfo *TLI) {
2405 const Function *F = ICS.getCalledFunction();
2407 return Intrinsic::not_intrinsic;
2409 if (F->isIntrinsic())
2410 return F->getIntrinsicID();
2413 return Intrinsic::not_intrinsic;
2416 // We're going to make assumptions on the semantics of the functions, check
2417 // that the target knows that it's available in this environment and it does
2418 // not have local linkage.
2419 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2420 return Intrinsic::not_intrinsic;
2422 if (!ICS.onlyReadsMemory())
2423 return Intrinsic::not_intrinsic;
2425 // Otherwise check if we have a call to a function that can be turned into a
2426 // vector intrinsic.
2433 return Intrinsic::sin;
2437 return Intrinsic::cos;
2441 return Intrinsic::exp;
2445 return Intrinsic::exp2;
2449 return Intrinsic::log;
2451 case LibFunc_log10f:
2452 case LibFunc_log10l:
2453 return Intrinsic::log10;
2457 return Intrinsic::log2;
2461 return Intrinsic::fabs;
2465 return Intrinsic::minnum;
2469 return Intrinsic::maxnum;
2470 case LibFunc_copysign:
2471 case LibFunc_copysignf:
2472 case LibFunc_copysignl:
2473 return Intrinsic::copysign;
2475 case LibFunc_floorf:
2476 case LibFunc_floorl:
2477 return Intrinsic::floor;
2481 return Intrinsic::ceil;
2483 case LibFunc_truncf:
2484 case LibFunc_truncl:
2485 return Intrinsic::trunc;
2489 return Intrinsic::rint;
2490 case LibFunc_nearbyint:
2491 case LibFunc_nearbyintf:
2492 case LibFunc_nearbyintl:
2493 return Intrinsic::nearbyint;
2495 case LibFunc_roundf:
2496 case LibFunc_roundl:
2497 return Intrinsic::round;
2501 return Intrinsic::pow;
2505 if (ICS->hasNoNaNs())
2506 return Intrinsic::sqrt;
2507 return Intrinsic::not_intrinsic;
2510 return Intrinsic::not_intrinsic;
2513 /// Return true if we can prove that the specified FP value is never equal to
2516 /// NOTE: this function will need to be revisited when we support non-default
2519 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
2521 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2522 return !CFP->getValueAPF().isNegZero();
2524 if (Depth == MaxDepth)
2525 return false; // Limit search depth.
2527 const Operator *I = dyn_cast<Operator>(V);
2528 if (!I) return false;
2530 // Check if the nsz fast-math flag is set
2531 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2532 if (FPO->hasNoSignedZeros())
2535 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2536 if (I->getOpcode() == Instruction::FAdd)
2537 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2538 if (CFP->isNullValue())
2541 // sitofp and uitofp turn into +0.0 for zero.
2542 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2545 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2546 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2550 // sqrt(-0.0) = -0.0, no other negative results are possible.
2551 case Intrinsic::sqrt:
2552 return CannotBeNegativeZero(CI->getArgOperand(0), TLI, Depth + 1);
2554 case Intrinsic::fabs:
2562 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2563 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2564 /// bit despite comparing equal.
2565 static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
2566 const TargetLibraryInfo *TLI,
2569 // TODO: This function does not do the right thing when SignBitOnly is true
2570 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2571 // which flips the sign bits of NaNs. See
2572 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2574 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2575 return !CFP->getValueAPF().isNegative() ||
2576 (!SignBitOnly && CFP->getValueAPF().isZero());
2579 if (Depth == MaxDepth)
2580 return false; // Limit search depth.
2582 const Operator *I = dyn_cast<Operator>(V);
2586 switch (I->getOpcode()) {
2589 // Unsigned integers are always nonnegative.
2590 case Instruction::UIToFP:
2592 case Instruction::FMul:
2593 // x*x is always non-negative or a NaN.
2594 if (I->getOperand(0) == I->getOperand(1) &&
2595 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2599 case Instruction::FAdd:
2600 case Instruction::FDiv:
2601 case Instruction::FRem:
2602 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2604 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2606 case Instruction::Select:
2607 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2609 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2611 case Instruction::FPExt:
2612 case Instruction::FPTrunc:
2613 // Widening/narrowing never change sign.
2614 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2616 case Instruction::Call:
2617 const auto *CI = cast<CallInst>(I);
2618 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2622 case Intrinsic::maxnum:
2623 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2625 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2627 case Intrinsic::minnum:
2628 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2630 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2632 case Intrinsic::exp:
2633 case Intrinsic::exp2:
2634 case Intrinsic::fabs:
2637 case Intrinsic::sqrt:
2638 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
2641 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2642 CannotBeNegativeZero(CI->getOperand(0), TLI));
2644 case Intrinsic::powi:
2645 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2646 // powi(x,n) is non-negative if n is even.
2647 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2650 // TODO: This is not correct. Given that exp is an integer, here are the
2651 // ways that pow can return a negative value:
2653 // pow(x, exp) --> negative if exp is odd and x is negative.
2654 // pow(-0, exp) --> -inf if exp is negative odd.
2655 // pow(-0, exp) --> -0 if exp is positive odd.
2656 // pow(-inf, exp) --> -0 if exp is negative odd.
2657 // pow(-inf, exp) --> -inf if exp is positive odd.
2659 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2660 // but we must return false if x == -0. Unfortunately we do not currently
2661 // have a way of expressing this constraint. See details in
2662 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2663 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2666 case Intrinsic::fma:
2667 case Intrinsic::fmuladd:
2668 // x*x+y is non-negative if y is non-negative.
2669 return I->getOperand(0) == I->getOperand(1) &&
2670 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
2671 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2679 bool llvm::CannotBeOrderedLessThanZero(const Value *V,
2680 const TargetLibraryInfo *TLI) {
2681 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
2684 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
2685 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
2688 /// If the specified value can be set by repeating the same byte in memory,
2689 /// return the i8 value that it is represented with. This is
2690 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2691 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2692 /// byte store (e.g. i16 0x1234), return null.
2693 Value *llvm::isBytewiseValue(Value *V) {
2694 // All byte-wide stores are splatable, even of arbitrary variables.
2695 if (V->getType()->isIntegerTy(8)) return V;
2697 // Handle 'null' ConstantArrayZero etc.
2698 if (Constant *C = dyn_cast<Constant>(V))
2699 if (C->isNullValue())
2700 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2702 // Constant float and double values can be handled as integer values if the
2703 // corresponding integer value is "byteable". An important case is 0.0.
2704 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2705 if (CFP->getType()->isFloatTy())
2706 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2707 if (CFP->getType()->isDoubleTy())
2708 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2709 // Don't handle long double formats, which have strange constraints.
2712 // We can handle constant integers that are multiple of 8 bits.
2713 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2714 if (CI->getBitWidth() % 8 == 0) {
2715 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2717 if (!CI->getValue().isSplat(8))
2719 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2723 // A ConstantDataArray/Vector is splatable if all its members are equal and
2725 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2726 Value *Elt = CA->getElementAsConstant(0);
2727 Value *Val = isBytewiseValue(Elt);
2731 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2732 if (CA->getElementAsConstant(I) != Elt)
2738 // Conceptually, we could handle things like:
2739 // %a = zext i8 %X to i16
2740 // %b = shl i16 %a, 8
2741 // %c = or i16 %a, %b
2742 // but until there is an example that actually needs this, it doesn't seem
2743 // worth worrying about.
2748 // This is the recursive version of BuildSubAggregate. It takes a few different
2749 // arguments. Idxs is the index within the nested struct From that we are
2750 // looking at now (which is of type IndexedType). IdxSkip is the number of
2751 // indices from Idxs that should be left out when inserting into the resulting
2752 // struct. To is the result struct built so far, new insertvalue instructions
2754 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2755 SmallVectorImpl<unsigned> &Idxs,
2757 Instruction *InsertBefore) {
2758 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2760 // Save the original To argument so we can modify it
2762 // General case, the type indexed by Idxs is a struct
2763 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2764 // Process each struct element recursively
2767 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2771 // Couldn't find any inserted value for this index? Cleanup
2772 while (PrevTo != OrigTo) {
2773 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2774 PrevTo = Del->getAggregateOperand();
2775 Del->eraseFromParent();
2777 // Stop processing elements
2781 // If we successfully found a value for each of our subaggregates
2785 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2786 // the struct's elements had a value that was inserted directly. In the latter
2787 // case, perhaps we can't determine each of the subelements individually, but
2788 // we might be able to find the complete struct somewhere.
2790 // Find the value that is at that particular spot
2791 Value *V = FindInsertedValue(From, Idxs);
2796 // Insert the value in the new (sub) aggregrate
2797 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2798 "tmp", InsertBefore);
2801 // This helper takes a nested struct and extracts a part of it (which is again a
2802 // struct) into a new value. For example, given the struct:
2803 // { a, { b, { c, d }, e } }
2804 // and the indices "1, 1" this returns
2807 // It does this by inserting an insertvalue for each element in the resulting
2808 // struct, as opposed to just inserting a single struct. This will only work if
2809 // each of the elements of the substruct are known (ie, inserted into From by an
2810 // insertvalue instruction somewhere).
2812 // All inserted insertvalue instructions are inserted before InsertBefore
2813 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2814 Instruction *InsertBefore) {
2815 assert(InsertBefore && "Must have someplace to insert!");
2816 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2818 Value *To = UndefValue::get(IndexedType);
2819 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2820 unsigned IdxSkip = Idxs.size();
2822 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2825 /// Given an aggregrate and an sequence of indices, see if
2826 /// the scalar value indexed is already around as a register, for example if it
2827 /// were inserted directly into the aggregrate.
2829 /// If InsertBefore is not null, this function will duplicate (modified)
2830 /// insertvalues when a part of a nested struct is extracted.
2831 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2832 Instruction *InsertBefore) {
2833 // Nothing to index? Just return V then (this is useful at the end of our
2835 if (idx_range.empty())
2837 // We have indices, so V should have an indexable type.
2838 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2839 "Not looking at a struct or array?");
2840 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2841 "Invalid indices for type?");
2843 if (Constant *C = dyn_cast<Constant>(V)) {
2844 C = C->getAggregateElement(idx_range[0]);
2845 if (!C) return nullptr;
2846 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2849 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2850 // Loop the indices for the insertvalue instruction in parallel with the
2851 // requested indices
2852 const unsigned *req_idx = idx_range.begin();
2853 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2854 i != e; ++i, ++req_idx) {
2855 if (req_idx == idx_range.end()) {
2856 // We can't handle this without inserting insertvalues
2860 // The requested index identifies a part of a nested aggregate. Handle
2861 // this specially. For example,
2862 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2863 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2864 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2865 // This can be changed into
2866 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2867 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2868 // which allows the unused 0,0 element from the nested struct to be
2870 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2874 // This insert value inserts something else than what we are looking for.
2875 // See if the (aggregate) value inserted into has the value we are
2876 // looking for, then.
2878 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2881 // If we end up here, the indices of the insertvalue match with those
2882 // requested (though possibly only partially). Now we recursively look at
2883 // the inserted value, passing any remaining indices.
2884 return FindInsertedValue(I->getInsertedValueOperand(),
2885 makeArrayRef(req_idx, idx_range.end()),
2889 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2890 // If we're extracting a value from an aggregate that was extracted from
2891 // something else, we can extract from that something else directly instead.
2892 // However, we will need to chain I's indices with the requested indices.
2894 // Calculate the number of indices required
2895 unsigned size = I->getNumIndices() + idx_range.size();
2896 // Allocate some space to put the new indices in
2897 SmallVector<unsigned, 5> Idxs;
2899 // Add indices from the extract value instruction
2900 Idxs.append(I->idx_begin(), I->idx_end());
2902 // Add requested indices
2903 Idxs.append(idx_range.begin(), idx_range.end());
2905 assert(Idxs.size() == size
2906 && "Number of indices added not correct?");
2908 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2910 // Otherwise, we don't know (such as, extracting from a function return value
2911 // or load instruction)
2915 /// Analyze the specified pointer to see if it can be expressed as a base
2916 /// pointer plus a constant offset. Return the base and offset to the caller.
2917 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2918 const DataLayout &DL) {
2919 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2920 APInt ByteOffset(BitWidth, 0);
2922 // We walk up the defs but use a visited set to handle unreachable code. In
2923 // that case, we stop after accumulating the cycle once (not that it
2925 SmallPtrSet<Value *, 16> Visited;
2926 while (Visited.insert(Ptr).second) {
2927 if (Ptr->getType()->isVectorTy())
2930 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2931 // If one of the values we have visited is an addrspacecast, then
2932 // the pointer type of this GEP may be different from the type
2933 // of the Ptr parameter which was passed to this function. This
2934 // means when we construct GEPOffset, we need to use the size
2935 // of GEP's pointer type rather than the size of the original
2937 APInt GEPOffset(DL.getPointerTypeSizeInBits(Ptr->getType()), 0);
2938 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2941 ByteOffset += GEPOffset.getSExtValue();
2943 Ptr = GEP->getPointerOperand();
2944 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2945 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2946 Ptr = cast<Operator>(Ptr)->getOperand(0);
2947 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2948 if (GA->isInterposable())
2950 Ptr = GA->getAliasee();
2955 Offset = ByteOffset.getSExtValue();
2959 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP,
2960 unsigned CharSize) {
2961 // Make sure the GEP has exactly three arguments.
2962 if (GEP->getNumOperands() != 3)
2965 // Make sure the index-ee is a pointer to array of \p CharSize integers.
2967 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
2968 if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
2971 // Check to make sure that the first operand of the GEP is an integer and
2972 // has value 0 so that we are sure we're indexing into the initializer.
2973 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2974 if (!FirstIdx || !FirstIdx->isZero())
2980 bool llvm::getConstantDataArrayInfo(const Value *V,
2981 ConstantDataArraySlice &Slice,
2982 unsigned ElementSize, uint64_t Offset) {
2985 // Look through bitcast instructions and geps.
2986 V = V->stripPointerCasts();
2988 // If the value is a GEP instruction or constant expression, treat it as an
2990 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2991 // The GEP operator should be based on a pointer to string constant, and is
2992 // indexing into the string constant.
2993 if (!isGEPBasedOnPointerToString(GEP, ElementSize))
2996 // If the second index isn't a ConstantInt, then this is a variable index
2997 // into the array. If this occurs, we can't say anything meaningful about
2999 uint64_t StartIdx = 0;
3000 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3001 StartIdx = CI->getZExtValue();
3004 return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
3008 // The GEP instruction, constant or instruction, must reference a global
3009 // variable that is a constant and is initialized. The referenced constant
3010 // initializer is the array that we'll use for optimization.
3011 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3012 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3015 const ConstantDataArray *Array;
3017 if (GV->getInitializer()->isNullValue()) {
3018 Type *GVTy = GV->getValueType();
3019 if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
3020 // A zeroinitializer for the array; There is no ConstantDataArray.
3023 const DataLayout &DL = GV->getParent()->getDataLayout();
3024 uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
3025 uint64_t Length = SizeInBytes / (ElementSize / 8);
3026 if (Length <= Offset)
3029 Slice.Array = nullptr;
3031 Slice.Length = Length - Offset;
3035 // This must be a ConstantDataArray.
3036 Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3039 ArrayTy = Array->getType();
3041 if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
3044 uint64_t NumElts = ArrayTy->getArrayNumElements();
3045 if (Offset > NumElts)
3048 Slice.Array = Array;
3049 Slice.Offset = Offset;
3050 Slice.Length = NumElts - Offset;
3054 /// This function computes the length of a null-terminated C string pointed to
3055 /// by V. If successful, it returns true and returns the string in Str.
3056 /// If unsuccessful, it returns false.
3057 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
3058 uint64_t Offset, bool TrimAtNul) {
3059 ConstantDataArraySlice Slice;
3060 if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
3063 if (Slice.Array == nullptr) {
3068 if (Slice.Length == 1) {
3069 Str = StringRef("", 1);
3072 // We cannot instantiate a StringRef as we do not have an apropriate string
3077 // Start out with the entire array in the StringRef.
3078 Str = Slice.Array->getAsString();
3079 // Skip over 'offset' bytes.
3080 Str = Str.substr(Slice.Offset);
3083 // Trim off the \0 and anything after it. If the array is not nul
3084 // terminated, we just return the whole end of string. The client may know
3085 // some other way that the string is length-bound.
3086 Str = Str.substr(0, Str.find('\0'));
3091 // These next two are very similar to the above, but also look through PHI
3093 // TODO: See if we can integrate these two together.
3095 /// If we can compute the length of the string pointed to by
3096 /// the specified pointer, return 'len+1'. If we can't, return 0.
3097 static uint64_t GetStringLengthH(const Value *V,
3098 SmallPtrSetImpl<const PHINode*> &PHIs,
3099 unsigned CharSize) {
3100 // Look through noop bitcast instructions.
3101 V = V->stripPointerCasts();
3103 // If this is a PHI node, there are two cases: either we have already seen it
3105 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3106 if (!PHIs.insert(PN).second)
3107 return ~0ULL; // already in the set.
3109 // If it was new, see if all the input strings are the same length.
3110 uint64_t LenSoFar = ~0ULL;
3111 for (Value *IncValue : PN->incoming_values()) {
3112 uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
3113 if (Len == 0) return 0; // Unknown length -> unknown.
3115 if (Len == ~0ULL) continue;
3117 if (Len != LenSoFar && LenSoFar != ~0ULL)
3118 return 0; // Disagree -> unknown.
3122 // Success, all agree.
3126 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3127 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3128 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
3129 if (Len1 == 0) return 0;
3130 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
3131 if (Len2 == 0) return 0;
3132 if (Len1 == ~0ULL) return Len2;
3133 if (Len2 == ~0ULL) return Len1;
3134 if (Len1 != Len2) return 0;
3138 // Otherwise, see if we can read the string.
3139 ConstantDataArraySlice Slice;
3140 if (!getConstantDataArrayInfo(V, Slice, CharSize))
3143 if (Slice.Array == nullptr)
3146 // Search for nul characters
3147 unsigned NullIndex = 0;
3148 for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
3149 if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
3153 return NullIndex + 1;
3156 /// If we can compute the length of the string pointed to by
3157 /// the specified pointer, return 'len+1'. If we can't, return 0.
3158 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
3159 if (!V->getType()->isPointerTy()) return 0;
3161 SmallPtrSet<const PHINode*, 32> PHIs;
3162 uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
3163 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3164 // an empty string as a length.
3165 return Len == ~0ULL ? 1 : Len;
3168 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
3169 /// previous iteration of the loop was referring to the same object as \p PN.
3170 static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
3171 const LoopInfo *LI) {
3172 // Find the loop-defined value.
3173 Loop *L = LI->getLoopFor(PN->getParent());
3174 if (PN->getNumIncomingValues() != 2)
3177 // Find the value from previous iteration.
3178 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3179 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3180 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3181 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3184 // If a new pointer is loaded in the loop, the pointer references a different
3185 // object in every iteration. E.g.:
3189 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3190 if (!L->isLoopInvariant(Load->getPointerOperand()))
3195 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3196 unsigned MaxLookup) {
3197 if (!V->getType()->isPointerTy())
3199 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3200 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3201 V = GEP->getPointerOperand();
3202 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3203 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3204 V = cast<Operator>(V)->getOperand(0);
3205 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3206 if (GA->isInterposable())
3208 V = GA->getAliasee();
3209 } else if (isa<AllocaInst>(V)) {
3210 // An alloca can't be further simplified.
3213 if (auto CS = CallSite(V))
3214 if (Value *RV = CS.getReturnedArgOperand()) {
3219 // See if InstructionSimplify knows any relevant tricks.
3220 if (Instruction *I = dyn_cast<Instruction>(V))
3221 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3222 if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3229 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3234 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3235 const DataLayout &DL, LoopInfo *LI,
3236 unsigned MaxLookup) {
3237 SmallPtrSet<Value *, 4> Visited;
3238 SmallVector<Value *, 4> Worklist;
3239 Worklist.push_back(V);
3241 Value *P = Worklist.pop_back_val();
3242 P = GetUnderlyingObject(P, DL, MaxLookup);
3244 if (!Visited.insert(P).second)
3247 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3248 Worklist.push_back(SI->getTrueValue());
3249 Worklist.push_back(SI->getFalseValue());
3253 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3254 // If this PHI changes the underlying object in every iteration of the
3255 // loop, don't look through it. Consider:
3258 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3262 // Prev is tracking Curr one iteration behind so they refer to different
3263 // underlying objects.
3264 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3265 isSameUnderlyingObjectInLoop(PN, LI))
3266 for (Value *IncValue : PN->incoming_values())
3267 Worklist.push_back(IncValue);
3271 Objects.push_back(P);
3272 } while (!Worklist.empty());
3275 /// Return true if the only users of this pointer are lifetime markers.
3276 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3277 for (const User *U : V->users()) {
3278 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3279 if (!II) return false;
3281 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3282 II->getIntrinsicID() != Intrinsic::lifetime_end)
3288 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3289 const Instruction *CtxI,
3290 const DominatorTree *DT) {
3291 const Operator *Inst = dyn_cast<Operator>(V);
3295 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3296 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3300 switch (Inst->getOpcode()) {
3303 case Instruction::UDiv:
3304 case Instruction::URem: {
3305 // x / y is undefined if y == 0.
3307 if (match(Inst->getOperand(1), m_APInt(V)))
3311 case Instruction::SDiv:
3312 case Instruction::SRem: {
3313 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3314 const APInt *Numerator, *Denominator;
3315 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3317 // We cannot hoist this division if the denominator is 0.
3318 if (*Denominator == 0)
3320 // It's safe to hoist if the denominator is not 0 or -1.
3321 if (*Denominator != -1)
3323 // At this point we know that the denominator is -1. It is safe to hoist as
3324 // long we know that the numerator is not INT_MIN.
3325 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3326 return !Numerator->isMinSignedValue();
3327 // The numerator *might* be MinSignedValue.
3330 case Instruction::Load: {
3331 const LoadInst *LI = cast<LoadInst>(Inst);
3332 if (!LI->isUnordered() ||
3333 // Speculative load may create a race that did not exist in the source.
3334 LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3335 // Speculative load may load data from dirty regions.
3336 LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress))
3338 const DataLayout &DL = LI->getModule()->getDataLayout();
3339 return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
3340 LI->getAlignment(), DL, CtxI, DT);
3342 case Instruction::Call: {
3343 auto *CI = cast<const CallInst>(Inst);
3344 const Function *Callee = CI->getCalledFunction();
3346 // The called function could have undefined behavior or side-effects, even
3347 // if marked readnone nounwind.
3348 return Callee && Callee->isSpeculatable();
3350 case Instruction::VAArg:
3351 case Instruction::Alloca:
3352 case Instruction::Invoke:
3353 case Instruction::PHI:
3354 case Instruction::Store:
3355 case Instruction::Ret:
3356 case Instruction::Br:
3357 case Instruction::IndirectBr:
3358 case Instruction::Switch:
3359 case Instruction::Unreachable:
3360 case Instruction::Fence:
3361 case Instruction::AtomicRMW:
3362 case Instruction::AtomicCmpXchg:
3363 case Instruction::LandingPad:
3364 case Instruction::Resume:
3365 case Instruction::CatchSwitch:
3366 case Instruction::CatchPad:
3367 case Instruction::CatchRet:
3368 case Instruction::CleanupPad:
3369 case Instruction::CleanupRet:
3370 return false; // Misc instructions which have effects
3374 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3375 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3378 /// Return true if we know that the specified value is never null.
3379 bool llvm::isKnownNonNull(const Value *V) {
3380 assert(V->getType()->isPointerTy() && "V must be pointer type");
3382 // Alloca never returns null, malloc might.
3383 if (isa<AllocaInst>(V)) return true;
3385 // A byval, inalloca, or nonnull argument is never null.
3386 if (const Argument *A = dyn_cast<Argument>(V))
3387 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3389 // A global variable in address space 0 is non null unless extern weak
3390 // or an absolute symbol reference. Other address spaces may have null as a
3391 // valid address for a global, so we can't assume anything.
3392 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3393 return !GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
3394 GV->getType()->getAddressSpace() == 0;
3396 // A Load tagged with nonnull metadata is never null.
3397 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3398 return LI->getMetadata(LLVMContext::MD_nonnull);
3400 if (auto CS = ImmutableCallSite(V))
3401 if (CS.isReturnNonNull())
3407 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3408 const Instruction *CtxI,
3409 const DominatorTree *DT) {
3410 assert(V->getType()->isPointerTy() && "V must be pointer type");
3411 assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
3412 assert(CtxI && "Context instruction required for analysis");
3413 assert(DT && "Dominator tree required for analysis");
3415 unsigned NumUsesExplored = 0;
3416 for (auto *U : V->users()) {
3417 // Avoid massive lists
3418 if (NumUsesExplored >= DomConditionsMaxUses)
3422 // If the value is used as an argument to a call or invoke, then argument
3423 // attributes may provide an answer about null-ness.
3424 if (auto CS = ImmutableCallSite(U))
3425 if (auto *CalledFunc = CS.getCalledFunction())
3426 for (const Argument &Arg : CalledFunc->args())
3427 if (CS.getArgOperand(Arg.getArgNo()) == V &&
3428 Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
3431 // Consider only compare instructions uniquely controlling a branch
3432 CmpInst::Predicate Pred;
3433 if (!match(const_cast<User *>(U),
3434 m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
3435 (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
3438 for (auto *CmpU : U->users()) {
3439 if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) {
3440 assert(BI->isConditional() && "uses a comparison!");
3442 BasicBlock *NonNullSuccessor =
3443 BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
3444 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3445 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3447 } else if (Pred == ICmpInst::ICMP_NE &&
3448 match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) &&
3449 DT->dominates(cast<Instruction>(CmpU), CtxI)) {
3458 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3459 const DominatorTree *DT) {
3460 if (isa<ConstantPointerNull>(V) || isa<UndefValue>(V))
3463 if (isKnownNonNull(V))
3469 return ::isKnownNonNullFromDominatingCondition(V, CtxI, DT);
3472 OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS,
3474 const DataLayout &DL,
3475 AssumptionCache *AC,
3476 const Instruction *CxtI,
3477 const DominatorTree *DT) {
3478 // Multiplying n * m significant bits yields a result of n + m significant
3479 // bits. If the total number of significant bits does not exceed the
3480 // result bit width (minus 1), there is no overflow.
3481 // This means if we have enough leading zero bits in the operands
3482 // we can guarantee that the result does not overflow.
3483 // Ref: "Hacker's Delight" by Henry Warren
3484 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3485 KnownBits LHSKnown(BitWidth);
3486 KnownBits RHSKnown(BitWidth);
3487 computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3488 computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3489 // Note that underestimating the number of zero bits gives a more
3490 // conservative answer.
3491 unsigned ZeroBits = LHSKnown.countMinLeadingZeros() +
3492 RHSKnown.countMinLeadingZeros();
3493 // First handle the easy case: if we have enough zero bits there's
3494 // definitely no overflow.
3495 if (ZeroBits >= BitWidth)
3496 return OverflowResult::NeverOverflows;
3498 // Get the largest possible values for each operand.
3499 APInt LHSMax = ~LHSKnown.Zero;
3500 APInt RHSMax = ~RHSKnown.Zero;
3502 // We know the multiply operation doesn't overflow if the maximum values for
3503 // each operand will not overflow after we multiply them together.
3505 (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3507 return OverflowResult::NeverOverflows;
3509 // We know it always overflows if multiplying the smallest possible values for
3510 // the operands also results in overflow.
3512 (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3514 return OverflowResult::AlwaysOverflows;
3516 return OverflowResult::MayOverflow;
3519 OverflowResult llvm::computeOverflowForUnsignedAdd(const Value *LHS,
3521 const DataLayout &DL,
3522 AssumptionCache *AC,
3523 const Instruction *CxtI,
3524 const DominatorTree *DT) {
3525 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3526 if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
3527 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3529 if (LHSKnown.isNegative() && RHSKnown.isNegative()) {
3530 // The sign bit is set in both cases: this MUST overflow.
3531 // Create a simple add instruction, and insert it into the struct.
3532 return OverflowResult::AlwaysOverflows;
3535 if (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()) {
3536 // The sign bit is clear in both cases: this CANNOT overflow.
3537 // Create a simple add instruction, and insert it into the struct.
3538 return OverflowResult::NeverOverflows;
3542 return OverflowResult::MayOverflow;
3545 /// \brief Return true if we can prove that adding the two values of the
3546 /// knownbits will not overflow.
3547 /// Otherwise return false.
3548 static bool checkRippleForSignedAdd(const KnownBits &LHSKnown,
3549 const KnownBits &RHSKnown) {
3550 // Addition of two 2's complement numbers having opposite signs will never
3552 if ((LHSKnown.isNegative() && RHSKnown.isNonNegative()) ||
3553 (LHSKnown.isNonNegative() && RHSKnown.isNegative()))
3556 // If either of the values is known to be non-negative, adding them can only
3557 // overflow if the second is also non-negative, so we can assume that.
3558 // Two non-negative numbers will only overflow if there is a carry to the
3559 // sign bit, so we can check if even when the values are as big as possible
3560 // there is no overflow to the sign bit.
3561 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) {
3562 APInt MaxLHS = ~LHSKnown.Zero;
3563 MaxLHS.clearSignBit();
3564 APInt MaxRHS = ~RHSKnown.Zero;
3565 MaxRHS.clearSignBit();
3566 APInt Result = std::move(MaxLHS) + std::move(MaxRHS);
3567 return Result.isSignBitClear();
3570 // If either of the values is known to be negative, adding them can only
3571 // overflow if the second is also negative, so we can assume that.
3572 // Two negative number will only overflow if there is no carry to the sign
3573 // bit, so we can check if even when the values are as small as possible
3574 // there is overflow to the sign bit.
3575 if (LHSKnown.isNegative() || RHSKnown.isNegative()) {
3576 APInt MinLHS = LHSKnown.One;
3577 MinLHS.clearSignBit();
3578 APInt MinRHS = RHSKnown.One;
3579 MinRHS.clearSignBit();
3580 APInt Result = std::move(MinLHS) + std::move(MinRHS);
3581 return Result.isSignBitSet();
3584 // If we reached here it means that we know nothing about the sign bits.
3585 // In this case we can't know if there will be an overflow, since by
3586 // changing the sign bits any two values can be made to overflow.
3590 static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
3592 const AddOperator *Add,
3593 const DataLayout &DL,
3594 AssumptionCache *AC,
3595 const Instruction *CxtI,
3596 const DominatorTree *DT) {
3597 if (Add && Add->hasNoSignedWrap()) {
3598 return OverflowResult::NeverOverflows;
3601 // If LHS and RHS each have at least two sign bits, the addition will look
3607 // If the carry into the most significant position is 0, X and Y can't both
3608 // be 1 and therefore the carry out of the addition is also 0.
3610 // If the carry into the most significant position is 1, X and Y can't both
3611 // be 0 and therefore the carry out of the addition is also 1.
3613 // Since the carry into the most significant position is always equal to
3614 // the carry out of the addition, there is no signed overflow.
3615 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
3616 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
3617 return OverflowResult::NeverOverflows;
3619 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3620 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3622 if (checkRippleForSignedAdd(LHSKnown, RHSKnown))
3623 return OverflowResult::NeverOverflows;
3625 // The remaining code needs Add to be available. Early returns if not so.
3627 return OverflowResult::MayOverflow;
3629 // If the sign of Add is the same as at least one of the operands, this add
3630 // CANNOT overflow. This is particularly useful when the sum is
3631 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3633 bool LHSOrRHSKnownNonNegative =
3634 (LHSKnown.isNonNegative() || RHSKnown.isNonNegative());
3635 bool LHSOrRHSKnownNegative =
3636 (LHSKnown.isNegative() || RHSKnown.isNegative());
3637 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3638 KnownBits AddKnown = computeKnownBits(Add, DL, /*Depth=*/0, AC, CxtI, DT);
3639 if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
3640 (AddKnown.isNegative() && LHSOrRHSKnownNegative)) {
3641 return OverflowResult::NeverOverflows;
3645 return OverflowResult::MayOverflow;
3648 bool llvm::isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
3649 const DominatorTree &DT) {
3651 auto IID = II->getIntrinsicID();
3652 assert((IID == Intrinsic::sadd_with_overflow ||
3653 IID == Intrinsic::uadd_with_overflow ||
3654 IID == Intrinsic::ssub_with_overflow ||
3655 IID == Intrinsic::usub_with_overflow ||
3656 IID == Intrinsic::smul_with_overflow ||
3657 IID == Intrinsic::umul_with_overflow) &&
3658 "Not an overflow intrinsic!");
3661 SmallVector<const BranchInst *, 2> GuardingBranches;
3662 SmallVector<const ExtractValueInst *, 2> Results;
3664 for (const User *U : II->users()) {
3665 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
3666 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
3668 if (EVI->getIndices()[0] == 0)
3669 Results.push_back(EVI);
3671 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
3673 for (const auto *U : EVI->users())
3674 if (const auto *B = dyn_cast<BranchInst>(U)) {
3675 assert(B->isConditional() && "How else is it using an i1?");
3676 GuardingBranches.push_back(B);
3680 // We are using the aggregate directly in a way we don't want to analyze
3681 // here (storing it to a global, say).
3686 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
3687 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
3688 if (!NoWrapEdge.isSingleEdge())
3691 // Check if all users of the add are provably no-wrap.
3692 for (const auto *Result : Results) {
3693 // If the extractvalue itself is not executed on overflow, the we don't
3694 // need to check each use separately, since domination is transitive.
3695 if (DT.dominates(NoWrapEdge, Result->getParent()))
3698 for (auto &RU : Result->uses())
3699 if (!DT.dominates(NoWrapEdge, RU))
3706 return any_of(GuardingBranches, AllUsesGuardedByBranch);
3710 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
3711 const DataLayout &DL,
3712 AssumptionCache *AC,
3713 const Instruction *CxtI,
3714 const DominatorTree *DT) {
3715 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3716 Add, DL, AC, CxtI, DT);
3719 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
3721 const DataLayout &DL,
3722 AssumptionCache *AC,
3723 const Instruction *CxtI,
3724 const DominatorTree *DT) {
3725 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3728 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3729 // A memory operation returns normally if it isn't volatile. A volatile
3730 // operation is allowed to trap.
3732 // An atomic operation isn't guaranteed to return in a reasonable amount of
3733 // time because it's possible for another thread to interfere with it for an
3734 // arbitrary length of time, but programs aren't allowed to rely on that.
3735 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
3736 return !LI->isVolatile();
3737 if (const StoreInst *SI = dyn_cast<StoreInst>(I))
3738 return !SI->isVolatile();
3739 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
3740 return !CXI->isVolatile();
3741 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
3742 return !RMWI->isVolatile();
3743 if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
3744 return !MII->isVolatile();
3746 // If there is no successor, then execution can't transfer to it.
3747 if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
3748 return !CRI->unwindsToCaller();
3749 if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
3750 return !CatchSwitch->unwindsToCaller();
3751 if (isa<ResumeInst>(I))
3753 if (isa<ReturnInst>(I))
3755 if (isa<UnreachableInst>(I))
3758 // Calls can throw, or contain an infinite loop, or kill the process.
3759 if (auto CS = ImmutableCallSite(I)) {
3760 // Call sites that throw have implicit non-local control flow.
3761 if (!CS.doesNotThrow())
3764 // Non-throwing call sites can loop infinitely, call exit/pthread_exit
3765 // etc. and thus not return. However, LLVM already assumes that
3767 // - Thread exiting actions are modeled as writes to memory invisible to
3770 // - Loops that don't have side effects (side effects are volatile/atomic
3771 // stores and IO) always terminate (see http://llvm.org/PR965).
3772 // Furthermore IO itself is also modeled as writes to memory invisible to
3775 // We rely on those assumptions here, and use the memory effects of the call
3776 // target as a proxy for checking that it always returns.
3778 // FIXME: This isn't aggressive enough; a call which only writes to a global
3779 // is guaranteed to return.
3780 return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
3781 match(I, m_Intrinsic<Intrinsic::assume>());
3784 // Other instructions return normally.
3788 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3790 // The loop header is guaranteed to be executed for every iteration.
3792 // FIXME: Relax this constraint to cover all basic blocks that are
3793 // guaranteed to be executed at every iteration.
3794 if (I->getParent() != L->getHeader()) return false;
3796 for (const Instruction &LI : *L->getHeader()) {
3797 if (&LI == I) return true;
3798 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3800 llvm_unreachable("Instruction not contained in its own parent basic block.");
3803 bool llvm::propagatesFullPoison(const Instruction *I) {
3804 switch (I->getOpcode()) {
3805 case Instruction::Add:
3806 case Instruction::Sub:
3807 case Instruction::Xor:
3808 case Instruction::Trunc:
3809 case Instruction::BitCast:
3810 case Instruction::AddrSpaceCast:
3811 case Instruction::Mul:
3812 case Instruction::Shl:
3813 case Instruction::GetElementPtr:
3814 // These operations all propagate poison unconditionally. Note that poison
3815 // is not any particular value, so xor or subtraction of poison with
3816 // itself still yields poison, not zero.
3819 case Instruction::AShr:
3820 case Instruction::SExt:
3821 // For these operations, one bit of the input is replicated across
3822 // multiple output bits. A replicated poison bit is still poison.
3825 case Instruction::ICmp:
3826 // Comparing poison with any value yields poison. This is why, for
3827 // instance, x s< (x +nsw 1) can be folded to true.
3835 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3836 switch (I->getOpcode()) {
3837 case Instruction::Store:
3838 return cast<StoreInst>(I)->getPointerOperand();
3840 case Instruction::Load:
3841 return cast<LoadInst>(I)->getPointerOperand();
3843 case Instruction::AtomicCmpXchg:
3844 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3846 case Instruction::AtomicRMW:
3847 return cast<AtomicRMWInst>(I)->getPointerOperand();
3849 case Instruction::UDiv:
3850 case Instruction::SDiv:
3851 case Instruction::URem:
3852 case Instruction::SRem:
3853 return I->getOperand(1);
3860 bool llvm::programUndefinedIfFullPoison(const Instruction *PoisonI) {
3861 // We currently only look for uses of poison values within the same basic
3862 // block, as that makes it easier to guarantee that the uses will be
3863 // executed given that PoisonI is executed.
3865 // FIXME: Expand this to consider uses beyond the same basic block. To do
3866 // this, look out for the distinction between post-dominance and strong
3868 const BasicBlock *BB = PoisonI->getParent();
3870 // Set of instructions that we have proved will yield poison if PoisonI
3872 SmallSet<const Value *, 16> YieldsPoison;
3873 SmallSet<const BasicBlock *, 4> Visited;
3874 YieldsPoison.insert(PoisonI);
3875 Visited.insert(PoisonI->getParent());
3877 BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
3880 while (Iter++ < MaxDepth) {
3881 for (auto &I : make_range(Begin, End)) {
3882 if (&I != PoisonI) {
3883 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
3884 if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
3886 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
3890 // Mark poison that propagates from I through uses of I.
3891 if (YieldsPoison.count(&I)) {
3892 for (const User *User : I.users()) {
3893 const Instruction *UserI = cast<Instruction>(User);
3894 if (propagatesFullPoison(UserI))
3895 YieldsPoison.insert(User);
3900 if (auto *NextBB = BB->getSingleSuccessor()) {
3901 if (Visited.insert(NextBB).second) {
3903 Begin = BB->getFirstNonPHI()->getIterator();
3914 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
3918 if (auto *C = dyn_cast<ConstantFP>(V))
3923 static bool isKnownNonZero(const Value *V) {
3924 if (auto *C = dyn_cast<ConstantFP>(V))
3925 return !C->isZero();
3929 /// Match non-obvious integer minimum and maximum sequences.
3930 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
3931 Value *CmpLHS, Value *CmpRHS,
3932 Value *TrueVal, Value *FalseVal,
3933 Value *&LHS, Value *&RHS) {
3934 // Assume success. If there's no match, callers should not use these anyway.
3938 // Recognize variations of:
3939 // CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
3941 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
3944 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
3945 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
3946 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
3947 return {SPF_SMAX, SPNB_NA, false};
3949 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
3950 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
3951 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
3952 return {SPF_SMIN, SPNB_NA, false};
3954 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
3955 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
3956 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
3957 return {SPF_UMAX, SPNB_NA, false};
3959 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
3960 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
3961 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
3962 return {SPF_UMIN, SPNB_NA, false};
3965 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
3966 return {SPF_UNKNOWN, SPNB_NA, false};
3969 // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
3970 // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
3971 if (match(TrueVal, m_Zero()) &&
3972 match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
3973 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
3976 // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
3977 // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
3978 if (match(FalseVal, m_Zero()) &&
3979 match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
3980 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
3982 if (!match(CmpRHS, m_APInt(C1)))
3983 return {SPF_UNKNOWN, SPNB_NA, false};
3985 // An unsigned min/max can be written with a signed compare.
3987 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
3988 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
3989 // Is the sign bit set?
3990 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
3991 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
3992 if (Pred == CmpInst::ICMP_SLT && *C1 == 0 && C2->isMaxSignedValue())
3993 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
3995 // Is the sign bit clear?
3996 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
3997 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
3998 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
3999 C2->isMinSignedValue())
4000 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4003 // Look through 'not' ops to find disguised signed min/max.
4004 // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
4005 // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
4006 if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
4007 match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
4008 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4010 // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
4011 // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
4012 if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
4013 match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
4014 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4016 return {SPF_UNKNOWN, SPNB_NA, false};
4019 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
4021 Value *CmpLHS, Value *CmpRHS,
4022 Value *TrueVal, Value *FalseVal,
4023 Value *&LHS, Value *&RHS) {
4027 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
4028 // return inconsistent results between implementations.
4029 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4030 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4031 // Therefore we behave conservatively and only proceed if at least one of the
4032 // operands is known to not be zero, or if we don't care about signed zeroes.
4035 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
4036 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
4037 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4038 !isKnownNonZero(CmpRHS))
4039 return {SPF_UNKNOWN, SPNB_NA, false};
4042 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4043 bool Ordered = false;
4045 // When given one NaN and one non-NaN input:
4046 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4047 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4048 // ordered comparison fails), which could be NaN or non-NaN.
4049 // so here we discover exactly what NaN behavior is required/accepted.
4050 if (CmpInst::isFPPredicate(Pred)) {
4051 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4052 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4054 if (LHSSafe && RHSSafe) {
4055 // Both operands are known non-NaN.
4056 NaNBehavior = SPNB_RETURNS_ANY;
4057 } else if (CmpInst::isOrdered(Pred)) {
4058 // An ordered comparison will return false when given a NaN, so it
4062 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
4063 NaNBehavior = SPNB_RETURNS_NAN;
4065 NaNBehavior = SPNB_RETURNS_OTHER;
4067 // Completely unsafe.
4068 return {SPF_UNKNOWN, SPNB_NA, false};
4071 // An unordered comparison will return true when given a NaN, so it
4074 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
4075 NaNBehavior = SPNB_RETURNS_OTHER;
4077 NaNBehavior = SPNB_RETURNS_NAN;
4079 // Completely unsafe.
4080 return {SPF_UNKNOWN, SPNB_NA, false};
4084 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
4085 std::swap(CmpLHS, CmpRHS);
4086 Pred = CmpInst::getSwappedPredicate(Pred);
4087 if (NaNBehavior == SPNB_RETURNS_NAN)
4088 NaNBehavior = SPNB_RETURNS_OTHER;
4089 else if (NaNBehavior == SPNB_RETURNS_OTHER)
4090 NaNBehavior = SPNB_RETURNS_NAN;
4094 // ([if]cmp X, Y) ? X : Y
4095 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
4097 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4098 case ICmpInst::ICMP_UGT:
4099 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4100 case ICmpInst::ICMP_SGT:
4101 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4102 case ICmpInst::ICMP_ULT:
4103 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4104 case ICmpInst::ICMP_SLT:
4105 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4106 case FCmpInst::FCMP_UGT:
4107 case FCmpInst::FCMP_UGE:
4108 case FCmpInst::FCMP_OGT:
4109 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4110 case FCmpInst::FCMP_ULT:
4111 case FCmpInst::FCMP_ULE:
4112 case FCmpInst::FCMP_OLT:
4113 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4118 if (match(CmpRHS, m_APInt(C1))) {
4119 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
4120 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
4122 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
4123 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
4124 if (Pred == ICmpInst::ICMP_SGT && (*C1 == 0 || C1->isAllOnesValue())) {
4125 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
4128 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
4129 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
4130 if (Pred == ICmpInst::ICMP_SLT && (*C1 == 0 || *C1 == 1)) {
4131 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
4136 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4139 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4140 Instruction::CastOps *CastOp) {
4141 auto *Cast1 = dyn_cast<CastInst>(V1);
4145 *CastOp = Cast1->getOpcode();
4146 Type *SrcTy = Cast1->getSrcTy();
4147 if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4148 // If V1 and V2 are both the same cast from the same type, look through V1.
4149 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4150 return Cast2->getOperand(0);
4154 auto *C = dyn_cast<Constant>(V2);
4158 Constant *CastedTo = nullptr;
4160 case Instruction::ZExt:
4161 if (CmpI->isUnsigned())
4162 CastedTo = ConstantExpr::getTrunc(C, SrcTy);
4164 case Instruction::SExt:
4165 if (CmpI->isSigned())
4166 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
4168 case Instruction::Trunc:
4169 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
4171 case Instruction::FPTrunc:
4172 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
4174 case Instruction::FPExt:
4175 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
4177 case Instruction::FPToUI:
4178 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
4180 case Instruction::FPToSI:
4181 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
4183 case Instruction::UIToFP:
4184 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
4186 case Instruction::SIToFP:
4187 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
4196 // Make sure the cast doesn't lose any information.
4197 Constant *CastedBack =
4198 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
4199 if (CastedBack != C)
4205 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
4206 Instruction::CastOps *CastOp) {
4207 SelectInst *SI = dyn_cast<SelectInst>(V);
4208 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4210 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4211 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4213 CmpInst::Predicate Pred = CmpI->getPredicate();
4214 Value *CmpLHS = CmpI->getOperand(0);
4215 Value *CmpRHS = CmpI->getOperand(1);
4216 Value *TrueVal = SI->getTrueValue();
4217 Value *FalseVal = SI->getFalseValue();
4219 if (isa<FPMathOperator>(CmpI))
4220 FMF = CmpI->getFastMathFlags();
4223 if (CmpI->isEquality())
4224 return {SPF_UNKNOWN, SPNB_NA, false};
4226 // Deal with type mismatches.
4227 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4228 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
4229 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4230 cast<CastInst>(TrueVal)->getOperand(0), C,
4232 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
4233 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4234 C, cast<CastInst>(FalseVal)->getOperand(0),
4237 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4241 /// Return true if "icmp Pred LHS RHS" is always true.
4242 static bool isTruePredicate(CmpInst::Predicate Pred,
4243 const Value *LHS, const Value *RHS,
4244 const DataLayout &DL, unsigned Depth,
4245 AssumptionCache *AC, const Instruction *CxtI,
4246 const DominatorTree *DT) {
4247 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
4248 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
4255 case CmpInst::ICMP_SLE: {
4258 // LHS s<= LHS +_{nsw} C if C >= 0
4259 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
4260 return !C->isNegative();
4264 case CmpInst::ICMP_ULE: {
4267 // LHS u<= LHS +_{nuw} C for any C
4268 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
4271 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
4272 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
4274 const APInt *&CA, const APInt *&CB) {
4275 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
4276 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
4279 // If X & C == 0 then (X | C) == X +_{nuw} C
4280 if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
4281 match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
4282 KnownBits Known(CA->getBitWidth());
4283 computeKnownBits(X, Known, DL, Depth + 1, AC, CxtI, DT);
4285 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
4293 const APInt *CLHS, *CRHS;
4294 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
4295 return CLHS->ule(*CRHS);
4302 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
4303 /// ALHS ARHS" is true. Otherwise, return None.
4304 static Optional<bool>
4305 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
4306 const Value *ARHS, const Value *BLHS,
4307 const Value *BRHS, const DataLayout &DL,
4308 unsigned Depth, AssumptionCache *AC,
4309 const Instruction *CxtI, const DominatorTree *DT) {
4314 case CmpInst::ICMP_SLT:
4315 case CmpInst::ICMP_SLE:
4316 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth, AC, CxtI,
4318 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4322 case CmpInst::ICMP_ULT:
4323 case CmpInst::ICMP_ULE:
4324 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth, AC, CxtI,
4326 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4332 /// Return true if the operands of the two compares match. IsSwappedOps is true
4333 /// when the operands match, but are swapped.
4334 static bool isMatchingOps(const Value *ALHS, const Value *ARHS,
4335 const Value *BLHS, const Value *BRHS,
4336 bool &IsSwappedOps) {
4338 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
4339 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
4340 return IsMatchingOps || IsSwappedOps;
4343 /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is
4344 /// true. Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS
4345 /// BRHS" is false. Otherwise, return None if we can't infer anything.
4346 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
4349 CmpInst::Predicate BPred,
4352 bool IsSwappedOps) {
4353 // Canonicalize the operands so they're matching.
4355 std::swap(BLHS, BRHS);
4356 BPred = ICmpInst::getSwappedPredicate(BPred);
4358 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
4360 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
4366 /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is
4367 /// true. Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS
4368 /// C2" is false. Otherwise, return None if we can't infer anything.
4369 static Optional<bool>
4370 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, const Value *ALHS,
4371 const ConstantInt *C1,
4372 CmpInst::Predicate BPred,
4373 const Value *BLHS, const ConstantInt *C2) {
4374 assert(ALHS == BLHS && "LHS operands must match.");
4375 ConstantRange DomCR =
4376 ConstantRange::makeExactICmpRegion(APred, C1->getValue());
4378 ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
4379 ConstantRange Intersection = DomCR.intersectWith(CR);
4380 ConstantRange Difference = DomCR.difference(CR);
4381 if (Intersection.isEmptySet())
4383 if (Difference.isEmptySet())
4388 Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
4389 const DataLayout &DL, bool InvertAPred,
4390 unsigned Depth, AssumptionCache *AC,
4391 const Instruction *CxtI,
4392 const DominatorTree *DT) {
4393 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for example.
4394 if (LHS->getType() != RHS->getType())
4397 Type *OpTy = LHS->getType();
4398 assert(OpTy->getScalarType()->isIntegerTy(1));
4400 // LHS ==> RHS by definition
4401 if (!InvertAPred && LHS == RHS)
4404 if (OpTy->isVectorTy())
4405 // TODO: extending the code below to handle vectors
4407 assert(OpTy->isIntegerTy(1) && "implied by above");
4409 ICmpInst::Predicate APred, BPred;
4413 if (!match(LHS, m_ICmp(APred, m_Value(ALHS), m_Value(ARHS))) ||
4414 !match(RHS, m_ICmp(BPred, m_Value(BLHS), m_Value(BRHS))))
4418 APred = CmpInst::getInversePredicate(APred);
4420 // Can we infer anything when the two compares have matching operands?
4422 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) {
4423 if (Optional<bool> Implication = isImpliedCondMatchingOperands(
4424 APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps))
4426 // No amount of additional analysis will infer the second condition, so
4431 // Can we infer anything when the LHS operands match and the RHS operands are
4432 // constants (not necessarily matching)?
4433 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
4434 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
4435 APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS,
4436 cast<ConstantInt>(BRHS)))
4438 // No amount of additional analysis will infer the second condition, so
4444 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth, AC,