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/Dominators.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/GlobalAlias.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/Statepoint.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/KnownBits.h"
42 #include "llvm/Support/MathExtras.h"
47 using namespace llvm::PatternMatch;
49 const unsigned MaxDepth = 6;
51 // Controls the number of uses of the value searched for possible
52 // dominating comparisons.
53 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
54 cl::Hidden, cl::init(20));
56 // This optimization is known to cause performance regressions is some cases,
57 // keep it under a temporary flag for now.
59 DontImproveNonNegativePhiBits("dont-improve-non-negative-phi-bits",
60 cl::Hidden, cl::init(true));
62 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
63 /// 0). For vector types, returns the element type's bitwidth.
64 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
65 if (unsigned BitWidth = Ty->getScalarSizeInBits())
68 return DL.getPointerTypeSizeInBits(Ty);
72 // Simplifying using an assume can only be done in a particular control-flow
73 // context (the context instruction provides that context). If an assume and
74 // the context instruction are not in the same block then the DT helps in
75 // figuring out if we can use it.
79 const Instruction *CxtI;
80 const DominatorTree *DT;
81 // Unlike the other analyses, this may be a nullptr because not all clients
82 // provide it currently.
83 OptimizationRemarkEmitter *ORE;
85 /// Set of assumptions that should be excluded from further queries.
86 /// This is because of the potential for mutual recursion to cause
87 /// computeKnownBits to repeatedly visit the same assume intrinsic. The
88 /// classic case of this is assume(x = y), which will attempt to determine
89 /// bits in x from bits in y, which will attempt to determine bits in y from
90 /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
91 /// isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
92 /// isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so
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 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
147 const DataLayout &DL,
148 AssumptionCache *AC, const Instruction *CxtI,
149 const DominatorTree *DT) {
150 assert(LHS->getType() == RHS->getType() &&
151 "LHS and RHS should have the same type");
152 assert(LHS->getType()->isIntOrIntVectorTy() &&
153 "LHS and RHS should be integers");
154 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
155 KnownBits LHSKnown(IT->getBitWidth());
156 KnownBits RHSKnown(IT->getBitWidth());
157 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT);
158 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT);
159 return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
162 static void ComputeSignBit(const Value *V, bool &KnownZero, bool &KnownOne,
163 unsigned Depth, const Query &Q);
165 void llvm::ComputeSignBit(const Value *V, bool &KnownZero, bool &KnownOne,
166 const DataLayout &DL, unsigned Depth,
167 AssumptionCache *AC, const Instruction *CxtI,
168 const DominatorTree *DT) {
169 ::ComputeSignBit(V, KnownZero, KnownOne, Depth,
170 Query(DL, AC, safeCxtI(V, CxtI), DT));
173 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
176 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
178 unsigned Depth, AssumptionCache *AC,
179 const Instruction *CxtI,
180 const DominatorTree *DT) {
181 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
182 Query(DL, AC, safeCxtI(V, CxtI), DT));
185 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
187 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
188 AssumptionCache *AC, const Instruction *CxtI,
189 const DominatorTree *DT) {
190 return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
193 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
195 AssumptionCache *AC, const Instruction *CxtI,
196 const DominatorTree *DT) {
197 bool NonNegative, Negative;
198 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
202 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
203 AssumptionCache *AC, const Instruction *CxtI,
204 const DominatorTree *DT) {
205 if (auto *CI = dyn_cast<ConstantInt>(V))
206 return CI->getValue().isStrictlyPositive();
208 // TODO: We'd doing two recursive queries here. We should factor this such
209 // that only a single query is needed.
210 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) &&
211 isKnownNonZero(V, DL, Depth, AC, CxtI, DT);
214 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
215 AssumptionCache *AC, const Instruction *CxtI,
216 const DominatorTree *DT) {
217 bool NonNegative, Negative;
218 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT);
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.Zero.isSignBitSet() && Known2.Zero.isSignBitSet())
300 KnownOut.Zero.setSignBit();
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.One.isSignBitSet() && Known2.One.isSignBitSet())
304 KnownOut.One.setSignBit();
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.Zero.isSignBitSet();
325 bool isKnownNonNegativeOp0 = Known2.Zero.isSignBitSet();
326 bool isKnownNegativeOp1 = Known.One.isSignBitSet();
327 bool isKnownNegativeOp0 = Known2.One.isSignBitSet();
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 Known.One.clearAllBits();
346 unsigned TrailZ = Known.Zero.countTrailingOnes() +
347 Known2.Zero.countTrailingOnes();
348 unsigned LeadZ = std::max(Known.Zero.countLeadingOnes() +
349 Known2.Zero.countLeadingOnes(),
350 BitWidth) - BitWidth;
352 TrailZ = std::min(TrailZ, BitWidth);
353 LeadZ = std::min(LeadZ, BitWidth);
354 Known.Zero.clearAllBits();
355 Known.Zero.setLowBits(TrailZ);
356 Known.Zero.setHighBits(LeadZ);
358 // Only make use of no-wrap flags if we failed to compute the sign bit
359 // directly. This matters if the multiplication always overflows, in
360 // which case we prefer to follow the result of the direct computation,
361 // though as the program is invoking undefined behaviour we can choose
362 // whatever we like here.
363 if (isKnownNonNegative && !Known.One.isSignBitSet())
364 Known.Zero.setSignBit();
365 else if (isKnownNegative && !Known.Zero.isSignBitSet())
366 Known.One.setSignBit();
369 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
372 unsigned BitWidth = KnownZero.getBitWidth();
373 unsigned NumRanges = Ranges.getNumOperands() / 2;
374 assert(NumRanges >= 1);
376 KnownZero.setAllBits();
377 KnownOne.setAllBits();
379 for (unsigned i = 0; i < NumRanges; ++i) {
381 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
383 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
384 ConstantRange Range(Lower->getValue(), Upper->getValue());
386 // The first CommonPrefixBits of all values in Range are equal.
387 unsigned CommonPrefixBits =
388 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
390 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
391 KnownOne &= Range.getUnsignedMax() & Mask;
392 KnownZero &= ~Range.getUnsignedMax() & Mask;
396 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
397 SmallVector<const Value *, 16> WorkSet(1, I);
398 SmallPtrSet<const Value *, 32> Visited;
399 SmallPtrSet<const Value *, 16> EphValues;
401 // The instruction defining an assumption's condition itself is always
402 // considered ephemeral to that assumption (even if it has other
403 // non-ephemeral users). See r246696's test case for an example.
404 if (is_contained(I->operands(), E))
407 while (!WorkSet.empty()) {
408 const Value *V = WorkSet.pop_back_val();
409 if (!Visited.insert(V).second)
412 // If all uses of this value are ephemeral, then so is this value.
413 if (all_of(V->users(), [&](const User *U) { return EphValues.count(U); })) {
418 if (const User *U = dyn_cast<User>(V))
419 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
421 if (isSafeToSpeculativelyExecute(*J))
422 WorkSet.push_back(*J);
430 // Is this an intrinsic that cannot be speculated but also cannot trap?
431 static bool isAssumeLikeIntrinsic(const Instruction *I) {
432 if (const CallInst *CI = dyn_cast<CallInst>(I))
433 if (Function *F = CI->getCalledFunction())
434 switch (F->getIntrinsicID()) {
436 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
437 case Intrinsic::assume:
438 case Intrinsic::dbg_declare:
439 case Intrinsic::dbg_value:
440 case Intrinsic::invariant_start:
441 case Intrinsic::invariant_end:
442 case Intrinsic::lifetime_start:
443 case Intrinsic::lifetime_end:
444 case Intrinsic::objectsize:
445 case Intrinsic::ptr_annotation:
446 case Intrinsic::var_annotation:
453 bool llvm::isValidAssumeForContext(const Instruction *Inv,
454 const Instruction *CxtI,
455 const DominatorTree *DT) {
457 // There are two restrictions on the use of an assume:
458 // 1. The assume must dominate the context (or the control flow must
459 // reach the assume whenever it reaches the context).
460 // 2. The context must not be in the assume's set of ephemeral values
461 // (otherwise we will use the assume to prove that the condition
462 // feeding the assume is trivially true, thus causing the removal of
466 if (DT->dominates(Inv, CxtI))
468 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
469 // We don't have a DT, but this trivially dominates.
473 // With or without a DT, the only remaining case we will check is if the
474 // instructions are in the same BB. Give up if that is not the case.
475 if (Inv->getParent() != CxtI->getParent())
478 // If we have a dom tree, then we now know that the assume doens't dominate
479 // the other instruction. If we don't have a dom tree then we can check if
480 // the assume is first in the BB.
482 // Search forward from the assume until we reach the context (or the end
483 // of the block); the common case is that the assume will come first.
484 for (auto I = std::next(BasicBlock::const_iterator(Inv)),
485 IE = Inv->getParent()->end(); I != IE; ++I)
490 // The context comes first, but they're both in the same block. Make sure
491 // there is nothing in between that might interrupt the control flow.
492 for (BasicBlock::const_iterator I =
493 std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
495 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
498 return !isEphemeralValueOf(Inv, CxtI);
501 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
502 unsigned Depth, const Query &Q) {
503 // Use of assumptions is context-sensitive. If we don't have a context, we
505 if (!Q.AC || !Q.CxtI)
508 unsigned BitWidth = Known.getBitWidth();
510 // Note that the patterns below need to be kept in sync with the code
511 // in AssumptionCache::updateAffectedValues.
513 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
516 CallInst *I = cast<CallInst>(AssumeVH);
517 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
518 "Got assumption for the wrong function!");
522 // Warning: This loop can end up being somewhat performance sensetive.
523 // We're running this loop for once for each value queried resulting in a
524 // runtime of ~O(#assumes * #values).
526 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
527 "must be an assume intrinsic");
529 Value *Arg = I->getArgOperand(0);
531 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
532 assert(BitWidth == 1 && "assume operand is not i1?");
533 Known.Zero.clearAllBits();
534 Known.One.setAllBits();
537 if (match(Arg, m_Not(m_Specific(V))) &&
538 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
539 assert(BitWidth == 1 && "assume operand is not i1?");
540 Known.Zero.setAllBits();
541 Known.One.clearAllBits();
545 // The remaining tests are all recursive, so bail out if we hit the limit.
546 if (Depth == MaxDepth)
550 auto m_V = m_CombineOr(m_Specific(V),
551 m_CombineOr(m_PtrToInt(m_Specific(V)),
552 m_BitCast(m_Specific(V))));
554 CmpInst::Predicate Pred;
557 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
558 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
559 KnownBits RHSKnown(BitWidth);
560 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
561 Known.Zero |= RHSKnown.Zero;
562 Known.One |= RHSKnown.One;
564 } else if (match(Arg,
565 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
566 Pred == ICmpInst::ICMP_EQ &&
567 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
568 KnownBits RHSKnown(BitWidth);
569 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
570 KnownBits MaskKnown(BitWidth);
571 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
573 // For those bits in the mask that are known to be one, we can propagate
574 // known bits from the RHS to V.
575 Known.Zero |= RHSKnown.Zero & MaskKnown.One;
576 Known.One |= RHSKnown.One & MaskKnown.One;
577 // assume(~(v & b) = a)
578 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
580 Pred == ICmpInst::ICMP_EQ &&
581 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
582 KnownBits RHSKnown(BitWidth);
583 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
584 KnownBits MaskKnown(BitWidth);
585 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
587 // For those bits in the mask that are known to be one, we can propagate
588 // inverted known bits from the RHS to V.
589 Known.Zero |= RHSKnown.One & MaskKnown.One;
590 Known.One |= RHSKnown.Zero & MaskKnown.One;
592 } else if (match(Arg,
593 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
594 Pred == ICmpInst::ICMP_EQ &&
595 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
596 KnownBits RHSKnown(BitWidth);
597 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
598 KnownBits BKnown(BitWidth);
599 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
601 // For those bits in B that are known to be zero, we can propagate known
602 // bits from the RHS to V.
603 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
604 Known.One |= RHSKnown.One & BKnown.Zero;
605 // assume(~(v | b) = a)
606 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
608 Pred == ICmpInst::ICMP_EQ &&
609 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
610 KnownBits RHSKnown(BitWidth);
611 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
612 KnownBits BKnown(BitWidth);
613 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
615 // For those bits in B that are known to be zero, we can propagate
616 // inverted known bits from the RHS to V.
617 Known.Zero |= RHSKnown.One & BKnown.Zero;
618 Known.One |= RHSKnown.Zero & BKnown.Zero;
620 } else if (match(Arg,
621 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
622 Pred == ICmpInst::ICMP_EQ &&
623 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
624 KnownBits RHSKnown(BitWidth);
625 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
626 KnownBits BKnown(BitWidth);
627 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
629 // For those bits in B that are known to be zero, we can propagate known
630 // bits from the RHS to V. For those bits in B that are known to be one,
631 // we can propagate inverted known bits from the RHS to V.
632 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
633 Known.One |= RHSKnown.One & BKnown.Zero;
634 Known.Zero |= RHSKnown.One & BKnown.One;
635 Known.One |= RHSKnown.Zero & BKnown.One;
636 // assume(~(v ^ b) = a)
637 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
639 Pred == ICmpInst::ICMP_EQ &&
640 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
641 KnownBits RHSKnown(BitWidth);
642 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
643 KnownBits BKnown(BitWidth);
644 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
646 // For those bits in B that are known to be zero, we can propagate
647 // inverted known bits from the RHS to V. For those bits in B that are
648 // known to be one, we can propagate known bits from the RHS to V.
649 Known.Zero |= RHSKnown.One & BKnown.Zero;
650 Known.One |= RHSKnown.Zero & BKnown.Zero;
651 Known.Zero |= RHSKnown.Zero & BKnown.One;
652 Known.One |= RHSKnown.One & BKnown.One;
653 // assume(v << c = a)
654 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
656 Pred == ICmpInst::ICMP_EQ &&
657 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
658 KnownBits RHSKnown(BitWidth);
659 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
660 // For those bits in RHS that are known, we can propagate them to known
661 // bits in V shifted to the right by C.
662 RHSKnown.Zero.lshrInPlace(C->getZExtValue());
663 Known.Zero |= RHSKnown.Zero;
664 RHSKnown.One.lshrInPlace(C->getZExtValue());
665 Known.One |= RHSKnown.One;
666 // assume(~(v << c) = a)
667 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
669 Pred == ICmpInst::ICMP_EQ &&
670 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
671 KnownBits RHSKnown(BitWidth);
672 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
673 // For those bits in RHS that are known, we can propagate them inverted
674 // to known bits in V shifted to the right by C.
675 RHSKnown.One.lshrInPlace(C->getZExtValue());
676 Known.Zero |= RHSKnown.One;
677 RHSKnown.Zero.lshrInPlace(C->getZExtValue());
678 Known.One |= RHSKnown.Zero;
679 // assume(v >> c = a)
680 } else if (match(Arg,
681 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
682 m_AShr(m_V, m_ConstantInt(C))),
684 Pred == ICmpInst::ICMP_EQ &&
685 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
686 KnownBits RHSKnown(BitWidth);
687 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
688 // For those bits in RHS that are known, we can propagate them to known
689 // bits in V shifted to the right by C.
690 Known.Zero |= RHSKnown.Zero << C->getZExtValue();
691 Known.One |= RHSKnown.One << C->getZExtValue();
692 // assume(~(v >> c) = a)
693 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
694 m_LShr(m_V, m_ConstantInt(C)),
695 m_AShr(m_V, m_ConstantInt(C)))),
697 Pred == ICmpInst::ICMP_EQ &&
698 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
699 KnownBits RHSKnown(BitWidth);
700 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
701 // For those bits in RHS that are known, we can propagate them inverted
702 // to known bits in V shifted to the right by C.
703 Known.Zero |= RHSKnown.One << C->getZExtValue();
704 Known.One |= RHSKnown.Zero << C->getZExtValue();
705 // assume(v >=_s c) where c is non-negative
706 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
707 Pred == ICmpInst::ICMP_SGE &&
708 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
709 KnownBits RHSKnown(BitWidth);
710 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
712 if (RHSKnown.Zero.isSignBitSet()) {
713 // We know that the sign bit is zero.
714 Known.Zero.setSignBit();
716 // assume(v >_s c) where c is at least -1.
717 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
718 Pred == ICmpInst::ICMP_SGT &&
719 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
720 KnownBits RHSKnown(BitWidth);
721 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
723 if (RHSKnown.One.isAllOnesValue() || RHSKnown.Zero.isSignBitSet()) {
724 // We know that the sign bit is zero.
725 Known.Zero.setSignBit();
727 // assume(v <=_s c) where c is negative
728 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
729 Pred == ICmpInst::ICMP_SLE &&
730 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
731 KnownBits RHSKnown(BitWidth);
732 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
734 if (RHSKnown.One.isSignBitSet()) {
735 // We know that the sign bit is one.
736 Known.One.setSignBit();
738 // assume(v <_s c) where c is non-positive
739 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
740 Pred == ICmpInst::ICMP_SLT &&
741 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
742 KnownBits RHSKnown(BitWidth);
743 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
745 if (RHSKnown.Zero.isAllOnesValue() || RHSKnown.One.isSignBitSet()) {
746 // We know that the sign bit is one.
747 Known.One.setSignBit();
750 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
751 Pred == ICmpInst::ICMP_ULE &&
752 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
753 KnownBits RHSKnown(BitWidth);
754 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
756 // Whatever high bits in c are zero are known to be zero.
757 Known.Zero.setHighBits(RHSKnown.Zero.countLeadingOnes());
759 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
760 Pred == ICmpInst::ICMP_ULT &&
761 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
762 KnownBits RHSKnown(BitWidth);
763 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
765 // Whatever high bits in c are zero are known to be zero (if c is a power
766 // of 2, then one more).
767 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
768 Known.Zero.setHighBits(RHSKnown.Zero.countLeadingOnes()+1);
770 Known.Zero.setHighBits(RHSKnown.Zero.countLeadingOnes());
774 // If assumptions conflict with each other or previous known bits, then we
775 // have a logical fallacy. It's possible that the assumption is not reachable,
776 // so this isn't a real bug. On the other hand, the program may have undefined
777 // behavior, or we might have a bug in the compiler. We can't assert/crash, so
778 // clear out the known bits, try to warn the user, and hope for the best.
779 if (Known.Zero.intersects(Known.One)) {
780 Known.Zero.clearAllBits();
781 Known.One.clearAllBits();
784 auto *CxtI = const_cast<Instruction *>(Q.CxtI);
785 OptimizationRemarkAnalysis ORA("value-tracking", "BadAssumption", CxtI);
786 Q.ORE->emit(ORA << "Detected conflicting code assumptions. Program may "
787 "have undefined behavior, or compiler may have "
793 // Compute known bits from a shift operator, including those with a
794 // non-constant shift amount. Known is the outputs of this function. Known2 is a
795 // pre-allocated temporary with the/ same bit width as Known. KZF and KOF are
796 // operator-specific functors that, given the known-zero or known-one bits
797 // respectively, and a shift amount, compute the implied known-zero or known-one
798 // bits of the shift operator's result respectively for that shift amount. The
799 // results from calling KZF and KOF are conservatively combined for all
800 // permitted shift amounts.
801 static void computeKnownBitsFromShiftOperator(
802 const Operator *I, KnownBits &Known, KnownBits &Known2,
803 unsigned Depth, const Query &Q,
804 function_ref<APInt(const APInt &, unsigned)> KZF,
805 function_ref<APInt(const APInt &, unsigned)> KOF) {
806 unsigned BitWidth = Known.getBitWidth();
808 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
809 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
811 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
812 Known.Zero = KZF(Known.Zero, ShiftAmt);
813 Known.One = KOF(Known.One, ShiftAmt);
814 // If there is conflict between Known.Zero and Known.One, this must be an
815 // overflowing left shift, so the shift result is undefined. Clear Known
816 // bits so that other code could propagate this undef.
817 if ((Known.Zero & Known.One) != 0) {
818 Known.Zero.clearAllBits();
819 Known.One.clearAllBits();
825 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
827 // If the shift amount could be greater than or equal to the bit-width of the LHS, the
828 // value could be undef, so we don't know anything about it.
829 if ((~Known.Zero).uge(BitWidth)) {
830 Known.Zero.clearAllBits();
831 Known.One.clearAllBits();
835 // Note: We cannot use Known.Zero.getLimitedValue() here, because if
836 // BitWidth > 64 and any upper bits are known, we'll end up returning the
837 // limit value (which implies all bits are known).
838 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
839 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
841 // It would be more-clearly correct to use the two temporaries for this
842 // calculation. Reusing the APInts here to prevent unnecessary allocations.
843 Known.Zero.clearAllBits();
844 Known.One.clearAllBits();
846 // If we know the shifter operand is nonzero, we can sometimes infer more
847 // known bits. However this is expensive to compute, so be lazy about it and
848 // only compute it when absolutely necessary.
849 Optional<bool> ShifterOperandIsNonZero;
851 // Early exit if we can't constrain any well-defined shift amount.
852 if (!(ShiftAmtKZ & (BitWidth - 1)) && !(ShiftAmtKO & (BitWidth - 1))) {
853 ShifterOperandIsNonZero =
854 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
855 if (!*ShifterOperandIsNonZero)
859 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
861 Known.Zero.setAllBits();
862 Known.One.setAllBits();
863 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
864 // Combine the shifted known input bits only for those shift amounts
865 // compatible with its known constraints.
866 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
868 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
870 // If we know the shifter is nonzero, we may be able to infer more known
871 // bits. This check is sunk down as far as possible to avoid the expensive
872 // call to isKnownNonZero if the cheaper checks above fail.
874 if (!ShifterOperandIsNonZero.hasValue())
875 ShifterOperandIsNonZero =
876 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
877 if (*ShifterOperandIsNonZero)
881 Known.Zero &= KZF(Known2.Zero, ShiftAmt);
882 Known.One &= KOF(Known2.One, ShiftAmt);
885 // If there are no compatible shift amounts, then we've proven that the shift
886 // amount must be >= the BitWidth, and the result is undefined. We could
887 // return anything we'd like, but we need to make sure the sets of known bits
888 // stay disjoint (it should be better for some other code to actually
889 // propagate the undef than to pick a value here using known bits).
890 if (Known.Zero.intersects(Known.One)) {
891 Known.Zero.clearAllBits();
892 Known.One.clearAllBits();
896 static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
897 unsigned Depth, const Query &Q) {
898 unsigned BitWidth = Known.getBitWidth();
900 KnownBits Known2(Known);
901 switch (I->getOpcode()) {
903 case Instruction::Load:
904 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
905 computeKnownBitsFromRangeMetadata(*MD, Known.Zero, Known.One);
907 case Instruction::And: {
908 // If either the LHS or the RHS are Zero, the result is zero.
909 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
910 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
912 // Output known-1 bits are only known if set in both the LHS & RHS.
913 Known.One &= Known2.One;
914 // Output known-0 are known to be clear if zero in either the LHS | RHS.
915 Known.Zero |= Known2.Zero;
917 // and(x, add (x, -1)) is a common idiom that always clears the low bit;
918 // here we handle the more general case of adding any odd number by
919 // matching the form add(x, add(x, y)) where y is odd.
920 // TODO: This could be generalized to clearing any bit set in y where the
921 // following bit is known to be unset in y.
923 if (!Known.Zero[0] && !Known.One[0] &&
924 (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)),
926 match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)),
928 Known2.Zero.clearAllBits(); Known2.One.clearAllBits();
929 computeKnownBits(Y, Known2, Depth + 1, Q);
930 if (Known2.One.countTrailingOnes() > 0)
931 Known.Zero.setBit(0);
935 case Instruction::Or: {
936 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
937 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
939 // Output known-0 bits are only known if clear in both the LHS & RHS.
940 Known.Zero &= Known2.Zero;
941 // Output known-1 are known to be set if set in either the LHS | RHS.
942 Known.One |= Known2.One;
945 case Instruction::Xor: {
946 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
947 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
949 // Output known-0 bits are known if clear or set in both the LHS & RHS.
950 APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
951 // Output known-1 are known to be set if set in only one of the LHS, RHS.
952 Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
953 Known.Zero = std::move(KnownZeroOut);
956 case Instruction::Mul: {
957 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
958 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
962 case Instruction::UDiv: {
963 // For the purposes of computing leading zeros we can conservatively
964 // treat a udiv as a logical right shift by the power of 2 known to
965 // be less than the denominator.
966 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
967 unsigned LeadZ = Known2.Zero.countLeadingOnes();
969 Known2.One.clearAllBits();
970 Known2.Zero.clearAllBits();
971 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
972 unsigned RHSUnknownLeadingOnes = Known2.One.countLeadingZeros();
973 if (RHSUnknownLeadingOnes != BitWidth)
974 LeadZ = std::min(BitWidth,
975 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
977 Known.Zero.setHighBits(LeadZ);
980 case Instruction::Select: {
981 const Value *LHS, *RHS;
982 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
983 if (SelectPatternResult::isMinOrMax(SPF)) {
984 computeKnownBits(RHS, Known, Depth + 1, Q);
985 computeKnownBits(LHS, Known2, Depth + 1, Q);
987 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
988 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
991 unsigned MaxHighOnes = 0;
992 unsigned MaxHighZeros = 0;
993 if (SPF == SPF_SMAX) {
994 // If both sides are negative, the result is negative.
995 if (Known.One.isSignBitSet() && Known2.One.isSignBitSet())
996 // We can derive a lower bound on the result by taking the max of the
998 MaxHighOnes = std::max(Known.One.countLeadingOnes(),
999 Known2.One.countLeadingOnes());
1000 // If either side is non-negative, the result is non-negative.
1001 else if (Known.Zero.isSignBitSet() || Known2.Zero.isSignBitSet())
1003 } else if (SPF == SPF_SMIN) {
1004 // If both sides are non-negative, the result is non-negative.
1005 if (Known.Zero.isSignBitSet() && Known2.Zero.isSignBitSet())
1006 // We can derive an upper bound on the result by taking the max of the
1007 // leading zero bits.
1008 MaxHighZeros = std::max(Known.Zero.countLeadingOnes(),
1009 Known2.Zero.countLeadingOnes());
1010 // If either side is negative, the result is negative.
1011 else if (Known.One.isSignBitSet() || Known2.One.isSignBitSet())
1013 } else if (SPF == SPF_UMAX) {
1014 // We can derive a lower bound on the result by taking the max of the
1015 // leading one bits.
1017 std::max(Known.One.countLeadingOnes(), Known2.One.countLeadingOnes());
1018 } else if (SPF == SPF_UMIN) {
1019 // We can derive an upper bound on the result by taking the max of the
1020 // leading zero bits.
1022 std::max(Known.Zero.countLeadingOnes(), Known2.Zero.countLeadingOnes());
1025 // Only known if known in both the LHS and RHS.
1026 Known.One &= Known2.One;
1027 Known.Zero &= Known2.Zero;
1028 if (MaxHighOnes > 0)
1029 Known.One.setHighBits(MaxHighOnes);
1030 if (MaxHighZeros > 0)
1031 Known.Zero.setHighBits(MaxHighZeros);
1034 case Instruction::FPTrunc:
1035 case Instruction::FPExt:
1036 case Instruction::FPToUI:
1037 case Instruction::FPToSI:
1038 case Instruction::SIToFP:
1039 case Instruction::UIToFP:
1040 break; // Can't work with floating point.
1041 case Instruction::PtrToInt:
1042 case Instruction::IntToPtr:
1043 // Fall through and handle them the same as zext/trunc.
1045 case Instruction::ZExt:
1046 case Instruction::Trunc: {
1047 Type *SrcTy = I->getOperand(0)->getType();
1049 unsigned SrcBitWidth;
1050 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1051 // which fall through here.
1052 SrcBitWidth = Q.DL.getTypeSizeInBits(SrcTy->getScalarType());
1054 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1055 Known.Zero = Known.Zero.zextOrTrunc(SrcBitWidth);
1056 Known.One = Known.One.zextOrTrunc(SrcBitWidth);
1057 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1058 Known.Zero = Known.Zero.zextOrTrunc(BitWidth);
1059 Known.One = Known.One.zextOrTrunc(BitWidth);
1060 // Any top bits are known to be zero.
1061 if (BitWidth > SrcBitWidth)
1062 Known.Zero.setBitsFrom(SrcBitWidth);
1065 case Instruction::BitCast: {
1066 Type *SrcTy = I->getOperand(0)->getType();
1067 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
1068 // TODO: For now, not handling conversions like:
1069 // (bitcast i64 %x to <2 x i32>)
1070 !I->getType()->isVectorTy()) {
1071 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1076 case Instruction::SExt: {
1077 // Compute the bits in the result that are not present in the input.
1078 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1080 Known.Zero = Known.Zero.trunc(SrcBitWidth);
1081 Known.One = Known.One.trunc(SrcBitWidth);
1082 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1083 // If the sign bit of the input is known set or clear, then we know the
1084 // top bits of the result.
1085 Known.Zero = Known.Zero.sext(BitWidth);
1086 Known.One = Known.One.sext(BitWidth);
1089 case Instruction::Shl: {
1090 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1091 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1092 auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1093 APInt KZResult = KnownZero << ShiftAmt;
1094 KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1095 // If this shift has "nsw" keyword, then the result is either a poison
1096 // value or has the same sign bit as the first operand.
1097 if (NSW && KnownZero.isSignBitSet())
1098 KZResult.setSignBit();
1102 auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1103 APInt KOResult = KnownOne << ShiftAmt;
1104 if (NSW && KnownOne.isSignBitSet())
1105 KOResult.setSignBit();
1109 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1112 case Instruction::LShr: {
1113 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1114 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1115 APInt KZResult = KnownZero.lshr(ShiftAmt);
1116 // High bits known zero.
1117 KZResult.setHighBits(ShiftAmt);
1121 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1122 return KnownOne.lshr(ShiftAmt);
1125 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1128 case Instruction::AShr: {
1129 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1130 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1131 return KnownZero.ashr(ShiftAmt);
1134 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1135 return KnownOne.ashr(ShiftAmt);
1138 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1141 case Instruction::Sub: {
1142 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1143 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1144 Known, Known2, Depth, Q);
1147 case Instruction::Add: {
1148 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1149 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1150 Known, Known2, Depth, Q);
1153 case Instruction::SRem:
1154 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1155 APInt RA = Rem->getValue().abs();
1156 if (RA.isPowerOf2()) {
1157 APInt LowBits = RA - 1;
1158 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1160 // The low bits of the first operand are unchanged by the srem.
1161 Known.Zero = Known2.Zero & LowBits;
1162 Known.One = Known2.One & LowBits;
1164 // If the first operand is non-negative or has all low bits zero, then
1165 // the upper bits are all zero.
1166 if (Known2.Zero.isSignBitSet() || ((Known2.Zero & LowBits) == LowBits))
1167 Known.Zero |= ~LowBits;
1169 // If the first operand is negative and not all low bits are zero, then
1170 // the upper bits are all one.
1171 if (Known2.One.isSignBitSet() && ((Known2.One & LowBits) != 0))
1172 Known.One |= ~LowBits;
1174 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1179 // The sign bit is the LHS's sign bit, except when the result of the
1180 // remainder is zero.
1181 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1182 // If it's known zero, our sign bit is also zero.
1183 if (Known2.Zero.isSignBitSet())
1184 Known.Zero.setSignBit();
1187 case Instruction::URem: {
1188 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1189 const APInt &RA = Rem->getValue();
1190 if (RA.isPowerOf2()) {
1191 APInt LowBits = (RA - 1);
1192 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1193 Known.Zero |= ~LowBits;
1194 Known.One &= LowBits;
1199 // Since the result is less than or equal to either operand, any leading
1200 // zero bits in either operand must also exist in the result.
1201 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1202 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1204 unsigned Leaders = std::max(Known.Zero.countLeadingOnes(),
1205 Known2.Zero.countLeadingOnes());
1206 Known.One.clearAllBits();
1207 Known.Zero.clearAllBits();
1208 Known.Zero.setHighBits(Leaders);
1212 case Instruction::Alloca: {
1213 const AllocaInst *AI = cast<AllocaInst>(I);
1214 unsigned Align = AI->getAlignment();
1216 Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1219 Known.Zero.setLowBits(countTrailingZeros(Align));
1222 case Instruction::GetElementPtr: {
1223 // Analyze all of the subscripts of this getelementptr instruction
1224 // to determine if we can prove known low zero bits.
1225 KnownBits LocalKnown(BitWidth);
1226 computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1227 unsigned TrailZ = LocalKnown.Zero.countTrailingOnes();
1229 gep_type_iterator GTI = gep_type_begin(I);
1230 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1231 Value *Index = I->getOperand(i);
1232 if (StructType *STy = GTI.getStructTypeOrNull()) {
1233 // Handle struct member offset arithmetic.
1235 // Handle case when index is vector zeroinitializer
1236 Constant *CIndex = cast<Constant>(Index);
1237 if (CIndex->isZeroValue())
1240 if (CIndex->getType()->isVectorTy())
1241 Index = CIndex->getSplatValue();
1243 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1244 const StructLayout *SL = Q.DL.getStructLayout(STy);
1245 uint64_t Offset = SL->getElementOffset(Idx);
1246 TrailZ = std::min<unsigned>(TrailZ,
1247 countTrailingZeros(Offset));
1249 // Handle array index arithmetic.
1250 Type *IndexedTy = GTI.getIndexedType();
1251 if (!IndexedTy->isSized()) {
1255 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1256 uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1257 LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1258 computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1259 TrailZ = std::min(TrailZ,
1260 unsigned(countTrailingZeros(TypeSize) +
1261 LocalKnown.Zero.countTrailingOnes()));
1265 Known.Zero.setLowBits(TrailZ);
1268 case Instruction::PHI: {
1269 const PHINode *P = cast<PHINode>(I);
1270 // Handle the case of a simple two-predecessor recurrence PHI.
1271 // There's a lot more that could theoretically be done here, but
1272 // this is sufficient to catch some interesting cases.
1273 if (P->getNumIncomingValues() == 2) {
1274 for (unsigned i = 0; i != 2; ++i) {
1275 Value *L = P->getIncomingValue(i);
1276 Value *R = P->getIncomingValue(!i);
1277 Operator *LU = dyn_cast<Operator>(L);
1280 unsigned Opcode = LU->getOpcode();
1281 // Check for operations that have the property that if
1282 // both their operands have low zero bits, the result
1283 // will have low zero bits.
1284 if (Opcode == Instruction::Add ||
1285 Opcode == Instruction::Sub ||
1286 Opcode == Instruction::And ||
1287 Opcode == Instruction::Or ||
1288 Opcode == Instruction::Mul) {
1289 Value *LL = LU->getOperand(0);
1290 Value *LR = LU->getOperand(1);
1291 // Find a recurrence.
1298 // Ok, we have a PHI of the form L op= R. Check for low
1300 computeKnownBits(R, Known2, Depth + 1, Q);
1302 // We need to take the minimum number of known bits
1303 KnownBits Known3(Known);
1304 computeKnownBits(L, Known3, Depth + 1, Q);
1306 Known.Zero.setLowBits(std::min(Known2.Zero.countTrailingOnes(),
1307 Known3.Zero.countTrailingOnes()));
1309 if (DontImproveNonNegativePhiBits)
1312 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1313 if (OverflowOp && OverflowOp->hasNoSignedWrap()) {
1314 // If initial value of recurrence is nonnegative, and we are adding
1315 // a nonnegative number with nsw, the result can only be nonnegative
1316 // or poison value regardless of the number of times we execute the
1317 // add in phi recurrence. If initial value is negative and we are
1318 // adding a negative number with nsw, the result can only be
1319 // negative or poison value. Similar arguments apply to sub and mul.
1321 // (add non-negative, non-negative) --> non-negative
1322 // (add negative, negative) --> negative
1323 if (Opcode == Instruction::Add) {
1324 if (Known2.Zero.isSignBitSet() && Known3.Zero.isSignBitSet())
1325 Known.Zero.setSignBit();
1326 else if (Known2.One.isSignBitSet() && Known3.One.isSignBitSet())
1327 Known.One.setSignBit();
1330 // (sub nsw non-negative, negative) --> non-negative
1331 // (sub nsw negative, non-negative) --> negative
1332 else if (Opcode == Instruction::Sub && LL == I) {
1333 if (Known2.Zero.isSignBitSet() && Known3.One.isSignBitSet())
1334 Known.Zero.setSignBit();
1335 else if (Known2.One.isSignBitSet() && Known3.Zero.isSignBitSet())
1336 Known.One.setSignBit();
1339 // (mul nsw non-negative, non-negative) --> non-negative
1340 else if (Opcode == Instruction::Mul && Known2.Zero.isSignBitSet() &&
1341 Known3.Zero.isSignBitSet())
1342 Known.Zero.setSignBit();
1350 // Unreachable blocks may have zero-operand PHI nodes.
1351 if (P->getNumIncomingValues() == 0)
1354 // Otherwise take the unions of the known bit sets of the operands,
1355 // taking conservative care to avoid excessive recursion.
1356 if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1357 // Skip if every incoming value references to ourself.
1358 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1361 Known.Zero.setAllBits();
1362 Known.One.setAllBits();
1363 for (Value *IncValue : P->incoming_values()) {
1364 // Skip direct self references.
1365 if (IncValue == P) continue;
1367 Known2 = KnownBits(BitWidth);
1368 // Recurse, but cap the recursion to one level, because we don't
1369 // want to waste time spinning around in loops.
1370 computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
1371 Known.Zero &= Known2.Zero;
1372 Known.One &= Known2.One;
1373 // If all bits have been ruled out, there's no need to check
1375 if (!Known.Zero && !Known.One)
1381 case Instruction::Call:
1382 case Instruction::Invoke:
1383 // If range metadata is attached to this call, set known bits from that,
1384 // and then intersect with known bits based on other properties of the
1386 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1387 computeKnownBitsFromRangeMetadata(*MD, Known.Zero, Known.One);
1388 if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1389 computeKnownBits(RV, Known2, Depth + 1, Q);
1390 Known.Zero |= Known2.Zero;
1391 Known.One |= Known2.One;
1393 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1394 switch (II->getIntrinsicID()) {
1396 case Intrinsic::bitreverse:
1397 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1398 Known.Zero |= Known2.Zero.reverseBits();
1399 Known.One |= Known2.One.reverseBits();
1401 case Intrinsic::bswap:
1402 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1403 Known.Zero |= Known2.Zero.byteSwap();
1404 Known.One |= Known2.One.byteSwap();
1406 case Intrinsic::ctlz:
1407 case Intrinsic::cttz: {
1408 unsigned LowBits = Log2_32(BitWidth)+1;
1409 // If this call is undefined for 0, the result will be less than 2^n.
1410 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1412 Known.Zero.setBitsFrom(LowBits);
1415 case Intrinsic::ctpop: {
1416 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1417 // We can bound the space the count needs. Also, bits known to be zero
1418 // can't contribute to the population.
1419 unsigned BitsPossiblySet = BitWidth - Known2.Zero.countPopulation();
1420 unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1421 Known.Zero.setBitsFrom(LowBits);
1422 // TODO: we could bound KnownOne using the lower bound on the number
1423 // of bits which might be set provided by popcnt KnownOne2.
1426 case Intrinsic::x86_sse42_crc32_64_64:
1427 Known.Zero.setBitsFrom(32);
1432 case Instruction::ExtractElement:
1433 // Look through extract element. At the moment we keep this simple and skip
1434 // tracking the specific element. But at least we might find information
1435 // valid for all elements of the vector (for example if vector is sign
1436 // extended, shifted, etc).
1437 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1439 case Instruction::ExtractValue:
1440 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1441 const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1442 if (EVI->getNumIndices() != 1) break;
1443 if (EVI->getIndices()[0] == 0) {
1444 switch (II->getIntrinsicID()) {
1446 case Intrinsic::uadd_with_overflow:
1447 case Intrinsic::sadd_with_overflow:
1448 computeKnownBitsAddSub(true, II->getArgOperand(0),
1449 II->getArgOperand(1), false, Known, Known2,
1452 case Intrinsic::usub_with_overflow:
1453 case Intrinsic::ssub_with_overflow:
1454 computeKnownBitsAddSub(false, II->getArgOperand(0),
1455 II->getArgOperand(1), false, Known, Known2,
1458 case Intrinsic::umul_with_overflow:
1459 case Intrinsic::smul_with_overflow:
1460 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1461 Known, Known2, Depth, Q);
1469 /// Determine which bits of V are known to be either zero or one and return
1470 /// them in the Known bit set.
1472 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1473 /// we cannot optimize based on the assumption that it is zero without changing
1474 /// it to be an explicit zero. If we don't change it to zero, other code could
1475 /// optimized based on the contradictory assumption that it is non-zero.
1476 /// Because instcombine aggressively folds operations with undef args anyway,
1477 /// this won't lose us code quality.
1479 /// This function is defined on values with integer type, values with pointer
1480 /// type, and vectors of integers. In the case
1481 /// where V is a vector, known zero, and known one values are the
1482 /// same width as the vector element, and the bit is set only if it is true
1483 /// for all of the elements in the vector.
1484 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1486 assert(V && "No Value?");
1487 assert(Depth <= MaxDepth && "Limit Search Depth");
1488 unsigned BitWidth = Known.getBitWidth();
1490 assert((V->getType()->isIntOrIntVectorTy() ||
1491 V->getType()->getScalarType()->isPointerTy()) &&
1492 "Not integer or pointer type!");
1493 assert((Q.DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
1494 (!V->getType()->isIntOrIntVectorTy() ||
1495 V->getType()->getScalarSizeInBits() == BitWidth) &&
1496 "V and Known should have same BitWidth");
1500 if (match(V, m_APInt(C))) {
1501 // We know all of the bits for a scalar constant or a splat vector constant!
1503 Known.Zero = ~Known.One;
1506 // Null and aggregate-zero are all-zeros.
1507 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1508 Known.One.clearAllBits();
1509 Known.Zero.setAllBits();
1512 // Handle a constant vector by taking the intersection of the known bits of
1514 if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1515 // We know that CDS must be a vector of integers. Take the intersection of
1517 Known.Zero.setAllBits(); Known.One.setAllBits();
1518 APInt Elt(BitWidth, 0);
1519 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1520 Elt = CDS->getElementAsInteger(i);
1527 if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1528 // We know that CV must be a vector of integers. Take the intersection of
1530 Known.Zero.setAllBits(); Known.One.setAllBits();
1531 APInt Elt(BitWidth, 0);
1532 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1533 Constant *Element = CV->getAggregateElement(i);
1534 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1536 Known.Zero.clearAllBits();
1537 Known.One.clearAllBits();
1540 Elt = ElementCI->getValue();
1547 // Start out not knowing anything.
1548 Known.Zero.clearAllBits(); Known.One.clearAllBits();
1550 // We can't imply anything about undefs.
1551 if (isa<UndefValue>(V))
1554 // There's no point in looking through other users of ConstantData for
1555 // assumptions. Confirm that we've handled them all.
1556 assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1558 // Limit search depth.
1559 // All recursive calls that increase depth must come after this.
1560 if (Depth == MaxDepth)
1563 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1564 // the bits of its aliasee.
1565 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1566 if (!GA->isInterposable())
1567 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1571 if (const Operator *I = dyn_cast<Operator>(V))
1572 computeKnownBitsFromOperator(I, Known, Depth, Q);
1574 // Aligned pointers have trailing zeros - refine Known.Zero set
1575 if (V->getType()->isPointerTy()) {
1576 unsigned Align = V->getPointerAlignment(Q.DL);
1578 Known.Zero.setLowBits(countTrailingZeros(Align));
1581 // computeKnownBitsFromAssume strictly refines Known.
1582 // Therefore, we run them after computeKnownBitsFromOperator.
1584 // Check whether a nearby assume intrinsic can determine some known bits.
1585 computeKnownBitsFromAssume(V, Known, Depth, Q);
1587 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1590 /// Determine whether the sign bit is known to be zero or one.
1591 /// Convenience wrapper around computeKnownBits.
1592 void ComputeSignBit(const Value *V, bool &KnownZero, bool &KnownOne,
1593 unsigned Depth, const Query &Q) {
1594 unsigned BitWidth = getBitWidth(V->getType(), Q.DL);
1600 KnownBits Bits(BitWidth);
1601 computeKnownBits(V, Bits, Depth, Q);
1602 KnownOne = Bits.One.isSignBitSet();
1603 KnownZero = Bits.Zero.isSignBitSet();
1606 /// Return true if the given value is known to have exactly one
1607 /// bit set when defined. For vectors return true if every element is known to
1608 /// be a power of two when defined. Supports values with integer or pointer
1609 /// types and vectors of integers.
1610 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1612 if (const Constant *C = dyn_cast<Constant>(V)) {
1613 if (C->isNullValue())
1616 const APInt *ConstIntOrConstSplatInt;
1617 if (match(C, m_APInt(ConstIntOrConstSplatInt)))
1618 return ConstIntOrConstSplatInt->isPowerOf2();
1621 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1622 // it is shifted off the end then the result is undefined.
1623 if (match(V, m_Shl(m_One(), m_Value())))
1626 // (signmask) >>l X is clearly a power of two if the one is not shifted off
1627 // the bottom. If it is shifted off the bottom then the result is undefined.
1628 if (match(V, m_LShr(m_SignMask(), m_Value())))
1631 // The remaining tests are all recursive, so bail out if we hit the limit.
1632 if (Depth++ == MaxDepth)
1635 Value *X = nullptr, *Y = nullptr;
1636 // A shift left or a logical shift right of a power of two is a power of two
1638 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1639 match(V, m_LShr(m_Value(X), m_Value()))))
1640 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1642 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1643 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1645 if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1646 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1647 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1649 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1650 // A power of two and'd with anything is a power of two or zero.
1651 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1652 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1654 // X & (-X) is always a power of two or zero.
1655 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1660 // Adding a power-of-two or zero to the same power-of-two or zero yields
1661 // either the original power-of-two, a larger power-of-two or zero.
1662 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1663 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1664 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1665 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1666 match(X, m_And(m_Value(), m_Specific(Y))))
1667 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1669 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1670 match(Y, m_And(m_Value(), m_Specific(X))))
1671 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1674 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1675 KnownBits LHSBits(BitWidth);
1676 computeKnownBits(X, LHSBits, Depth, Q);
1678 KnownBits RHSBits(BitWidth);
1679 computeKnownBits(Y, RHSBits, Depth, Q);
1680 // If i8 V is a power of two or zero:
1681 // ZeroBits: 1 1 1 0 1 1 1 1
1682 // ~ZeroBits: 0 0 0 1 0 0 0 0
1683 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1684 // If OrZero isn't set, we cannot give back a zero result.
1685 // Make sure either the LHS or RHS has a bit set.
1686 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1691 // An exact divide or right shift can only shift off zero bits, so the result
1692 // is a power of two only if the first operand is a power of two and not
1693 // copying a sign bit (sdiv int_min, 2).
1694 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1695 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1696 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1703 /// \brief Test whether a GEP's result is known to be non-null.
1705 /// Uses properties inherent in a GEP to try to determine whether it is known
1708 /// Currently this routine does not support vector GEPs.
1709 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1711 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1714 // FIXME: Support vector-GEPs.
1715 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1717 // If the base pointer is non-null, we cannot walk to a null address with an
1718 // inbounds GEP in address space zero.
1719 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1722 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1723 // If so, then the GEP cannot produce a null pointer, as doing so would
1724 // inherently violate the inbounds contract within address space zero.
1725 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1726 GTI != GTE; ++GTI) {
1727 // Struct types are easy -- they must always be indexed by a constant.
1728 if (StructType *STy = GTI.getStructTypeOrNull()) {
1729 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1730 unsigned ElementIdx = OpC->getZExtValue();
1731 const StructLayout *SL = Q.DL.getStructLayout(STy);
1732 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1733 if (ElementOffset > 0)
1738 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1739 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1742 // Fast path the constant operand case both for efficiency and so we don't
1743 // increment Depth when just zipping down an all-constant GEP.
1744 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1750 // We post-increment Depth here because while isKnownNonZero increments it
1751 // as well, when we pop back up that increment won't persist. We don't want
1752 // to recurse 10k times just because we have 10k GEP operands. We don't
1753 // bail completely out because we want to handle constant GEPs regardless
1755 if (Depth++ >= MaxDepth)
1758 if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1765 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1766 /// ensure that the value it's attached to is never Value? 'RangeType' is
1767 /// is the type of the value described by the range.
1768 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1769 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1770 assert(NumRanges >= 1);
1771 for (unsigned i = 0; i < NumRanges; ++i) {
1772 ConstantInt *Lower =
1773 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1774 ConstantInt *Upper =
1775 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1776 ConstantRange Range(Lower->getValue(), Upper->getValue());
1777 if (Range.contains(Value))
1783 /// Return true if the given value is known to be non-zero when defined. For
1784 /// vectors, return true if every element is known to be non-zero when
1785 /// defined. For pointers, if the context instruction and dominator tree are
1786 /// specified, perform context-sensitive analysis and return true if the
1787 /// pointer couldn't possibly be null at the specified instruction.
1788 /// Supports values with integer or pointer type and vectors of integers.
1789 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1790 if (auto *C = dyn_cast<Constant>(V)) {
1791 if (C->isNullValue())
1793 if (isa<ConstantInt>(C))
1794 // Must be non-zero due to null test above.
1797 // For constant vectors, check that all elements are undefined or known
1798 // non-zero to determine that the whole vector is known non-zero.
1799 if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1800 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1801 Constant *Elt = C->getAggregateElement(i);
1802 if (!Elt || Elt->isNullValue())
1804 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1813 if (auto *I = dyn_cast<Instruction>(V)) {
1814 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1815 // If the possible ranges don't contain zero, then the value is
1816 // definitely non-zero.
1817 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1818 const APInt ZeroValue(Ty->getBitWidth(), 0);
1819 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1825 // The remaining tests are all recursive, so bail out if we hit the limit.
1826 if (Depth++ >= MaxDepth)
1829 // Check for pointer simplifications.
1830 if (V->getType()->isPointerTy()) {
1831 if (isKnownNonNullAt(V, Q.CxtI, Q.DT))
1833 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1834 if (isGEPKnownNonNull(GEP, Depth, Q))
1838 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
1840 // X | Y != 0 if X != 0 or Y != 0.
1841 Value *X = nullptr, *Y = nullptr;
1842 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1843 return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
1845 // ext X != 0 if X != 0.
1846 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1847 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
1849 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1850 // if the lowest bit is shifted off the end.
1851 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1852 // shl nuw can't remove any non-zero bits.
1853 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1854 if (BO->hasNoUnsignedWrap())
1855 return isKnownNonZero(X, Depth, Q);
1857 KnownBits Known(BitWidth);
1858 computeKnownBits(X, Known, Depth, Q);
1862 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1863 // defined if the sign bit is shifted off the end.
1864 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1865 // shr exact can only shift out zero bits.
1866 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1868 return isKnownNonZero(X, Depth, Q);
1870 bool XKnownNonNegative, XKnownNegative;
1871 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q);
1875 // If the shifter operand is a constant, and all of the bits shifted
1876 // out are known to be zero, and X is known non-zero then at least one
1877 // non-zero bit must remain.
1878 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
1879 KnownBits Known(BitWidth);
1880 computeKnownBits(X, Known, Depth, Q);
1882 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
1883 // Is there a known one in the portion not shifted out?
1884 if (Known.One.countLeadingZeros() < BitWidth - ShiftVal)
1886 // Are all the bits to be shifted out known zero?
1887 if (Known.Zero.countTrailingOnes() >= ShiftVal)
1888 return isKnownNonZero(X, Depth, Q);
1891 // div exact can only produce a zero if the dividend is zero.
1892 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1893 return isKnownNonZero(X, Depth, Q);
1896 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1897 bool XKnownNonNegative, XKnownNegative;
1898 bool YKnownNonNegative, YKnownNegative;
1899 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q);
1900 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Depth, Q);
1902 // If X and Y are both non-negative (as signed values) then their sum is not
1903 // zero unless both X and Y are zero.
1904 if (XKnownNonNegative && YKnownNonNegative)
1905 if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
1908 // If X and Y are both negative (as signed values) then their sum is not
1909 // zero unless both X and Y equal INT_MIN.
1910 if (BitWidth && XKnownNegative && YKnownNegative) {
1911 KnownBits Known(BitWidth);
1912 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1913 // The sign bit of X is set. If some other bit is set then X is not equal
1915 computeKnownBits(X, Known, Depth, Q);
1916 if (Known.One.intersects(Mask))
1918 // The sign bit of Y is set. If some other bit is set then Y is not equal
1920 computeKnownBits(Y, Known, Depth, Q);
1921 if (Known.One.intersects(Mask))
1925 // The sum of a non-negative number and a power of two is not zero.
1926 if (XKnownNonNegative &&
1927 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
1929 if (YKnownNonNegative &&
1930 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
1934 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1935 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1936 // If X and Y are non-zero then so is X * Y as long as the multiplication
1937 // does not overflow.
1938 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1939 isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
1942 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1943 else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
1944 if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
1945 isKnownNonZero(SI->getFalseValue(), Depth, Q))
1949 else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
1950 // Try and detect a recurrence that monotonically increases from a
1951 // starting value, as these are common as induction variables.
1952 if (PN->getNumIncomingValues() == 2) {
1953 Value *Start = PN->getIncomingValue(0);
1954 Value *Induction = PN->getIncomingValue(1);
1955 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
1956 std::swap(Start, Induction);
1957 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
1958 if (!C->isZero() && !C->isNegative()) {
1960 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
1961 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
1967 // Check if all incoming values are non-zero constant.
1968 bool AllNonZeroConstants = all_of(PN->operands(), [](Value *V) {
1969 return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZeroValue();
1971 if (AllNonZeroConstants)
1975 if (!BitWidth) return false;
1976 KnownBits Known(BitWidth);
1977 computeKnownBits(V, Known, Depth, Q);
1978 return Known.One != 0;
1981 /// Return true if V2 == V1 + X, where X is known non-zero.
1982 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
1983 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
1984 if (!BO || BO->getOpcode() != Instruction::Add)
1986 Value *Op = nullptr;
1987 if (V2 == BO->getOperand(0))
1988 Op = BO->getOperand(1);
1989 else if (V2 == BO->getOperand(1))
1990 Op = BO->getOperand(0);
1993 return isKnownNonZero(Op, 0, Q);
1996 /// Return true if it is known that V1 != V2.
1997 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
1998 if (V1->getType()->isVectorTy() || V1 == V2)
2000 if (V1->getType() != V2->getType())
2001 // We can't look through casts yet.
2003 if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
2006 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) {
2007 // Are any known bits in V1 contradictory to known bits in V2? If V1
2008 // has a known zero where V2 has a known one, they must not be equal.
2009 auto BitWidth = Ty->getBitWidth();
2010 KnownBits Known1(BitWidth);
2011 computeKnownBits(V1, Known1, 0, Q);
2012 KnownBits Known2(BitWidth);
2013 computeKnownBits(V2, Known2, 0, Q);
2015 APInt OppositeBits = (Known1.Zero & Known2.One) |
2016 (Known2.Zero & Known1.One);
2017 if (OppositeBits.getBoolValue())
2023 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2024 /// simplify operations downstream. Mask is known to be zero for bits that V
2027 /// This function is defined on values with integer type, values with pointer
2028 /// type, and vectors of integers. In the case
2029 /// where V is a vector, the mask, known zero, and known one values are the
2030 /// same width as the vector element, and the bit is set only if it is true
2031 /// for all of the elements in the vector.
2032 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2034 KnownBits Known(Mask.getBitWidth());
2035 computeKnownBits(V, Known, Depth, Q);
2036 return Mask.isSubsetOf(Known.Zero);
2039 /// For vector constants, loop over the elements and find the constant with the
2040 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2041 /// or if any element was not analyzed; otherwise, return the count for the
2042 /// element with the minimum number of sign bits.
2043 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2045 const auto *CV = dyn_cast<Constant>(V);
2046 if (!CV || !CV->getType()->isVectorTy())
2049 unsigned MinSignBits = TyBits;
2050 unsigned NumElts = CV->getType()->getVectorNumElements();
2051 for (unsigned i = 0; i != NumElts; ++i) {
2052 // If we find a non-ConstantInt, bail out.
2053 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2057 // If the sign bit is 1, flip the bits, so we always count leading zeros.
2058 APInt EltVal = Elt->getValue();
2059 if (EltVal.isNegative())
2061 MinSignBits = std::min(MinSignBits, EltVal.countLeadingZeros());
2067 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2070 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2072 unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2073 assert(Result > 0 && "At least one sign bit needs to be present!");
2077 /// Return the number of times the sign bit of the register is replicated into
2078 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2079 /// (itself), but other cases can give us information. For example, immediately
2080 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2081 /// other, so we return 3. For vectors, return the number of sign bits for the
2082 /// vector element with the mininum number of known sign bits.
2083 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2086 // We return the minimum number of sign bits that are guaranteed to be present
2087 // in V, so for undef we have to conservatively return 1. We don't have the
2088 // same behavior for poison though -- that's a FIXME today.
2090 unsigned TyBits = Q.DL.getTypeSizeInBits(V->getType()->getScalarType());
2092 unsigned FirstAnswer = 1;
2094 // Note that ConstantInt is handled by the general computeKnownBits case
2097 if (Depth == MaxDepth)
2098 return 1; // Limit search depth.
2100 const Operator *U = dyn_cast<Operator>(V);
2101 switch (Operator::getOpcode(V)) {
2103 case Instruction::SExt:
2104 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2105 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2107 case Instruction::SDiv: {
2108 const APInt *Denominator;
2109 // sdiv X, C -> adds log(C) sign bits.
2110 if (match(U->getOperand(1), m_APInt(Denominator))) {
2112 // Ignore non-positive denominator.
2113 if (!Denominator->isStrictlyPositive())
2116 // Calculate the incoming numerator bits.
2117 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2119 // Add floor(log(C)) bits to the numerator bits.
2120 return std::min(TyBits, NumBits + Denominator->logBase2());
2125 case Instruction::SRem: {
2126 const APInt *Denominator;
2127 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2128 // positive constant. This let us put a lower bound on the number of sign
2130 if (match(U->getOperand(1), m_APInt(Denominator))) {
2132 // Ignore non-positive denominator.
2133 if (!Denominator->isStrictlyPositive())
2136 // Calculate the incoming numerator bits. SRem by a positive constant
2137 // can't lower the number of sign bits.
2139 ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2141 // Calculate the leading sign bit constraints by examining the
2142 // denominator. Given that the denominator is positive, there are two
2145 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2146 // (1 << ceilLogBase2(C)).
2148 // 2. the numerator is negative. Then the result range is (-C,0] and
2149 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2151 // Thus a lower bound on the number of sign bits is `TyBits -
2152 // ceilLogBase2(C)`.
2154 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2155 return std::max(NumrBits, ResBits);
2160 case Instruction::AShr: {
2161 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2162 // ashr X, C -> adds C sign bits. Vectors too.
2164 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2165 unsigned ShAmtLimited = ShAmt->getZExtValue();
2166 if (ShAmtLimited >= TyBits)
2167 break; // Bad shift.
2168 Tmp += ShAmtLimited;
2169 if (Tmp > TyBits) Tmp = TyBits;
2173 case Instruction::Shl: {
2175 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2176 // shl destroys sign bits.
2177 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2178 Tmp2 = ShAmt->getZExtValue();
2179 if (Tmp2 >= TyBits || // Bad shift.
2180 Tmp2 >= Tmp) break; // Shifted all sign bits out.
2185 case Instruction::And:
2186 case Instruction::Or:
2187 case Instruction::Xor: // NOT is handled here.
2188 // Logical binary ops preserve the number of sign bits at the worst.
2189 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2191 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2192 FirstAnswer = std::min(Tmp, Tmp2);
2193 // We computed what we know about the sign bits as our first
2194 // answer. Now proceed to the generic code that uses
2195 // computeKnownBits, and pick whichever answer is better.
2199 case Instruction::Select:
2200 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2201 if (Tmp == 1) return 1; // Early out.
2202 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2203 return std::min(Tmp, Tmp2);
2205 case Instruction::Add:
2206 // Add can have at most one carry bit. Thus we know that the output
2207 // is, at worst, one more bit than the inputs.
2208 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2209 if (Tmp == 1) return 1; // Early out.
2211 // Special case decrementing a value (ADD X, -1):
2212 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2213 if (CRHS->isAllOnesValue()) {
2214 KnownBits Known(TyBits);
2215 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2217 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2219 if ((Known.Zero | 1).isAllOnesValue())
2222 // If we are subtracting one from a positive number, there is no carry
2223 // out of the result.
2224 if (Known.Zero.isSignBitSet())
2228 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2229 if (Tmp2 == 1) return 1;
2230 return std::min(Tmp, Tmp2)-1;
2232 case Instruction::Sub:
2233 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2234 if (Tmp2 == 1) return 1;
2237 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2238 if (CLHS->isNullValue()) {
2239 KnownBits Known(TyBits);
2240 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2241 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2243 if ((Known.Zero | 1).isAllOnesValue())
2246 // If the input is known to be positive (the sign bit is known clear),
2247 // the output of the NEG has the same number of sign bits as the input.
2248 if (Known.Zero.isSignBitSet())
2251 // Otherwise, we treat this like a SUB.
2254 // Sub can have at most one carry bit. Thus we know that the output
2255 // is, at worst, one more bit than the inputs.
2256 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2257 if (Tmp == 1) return 1; // Early out.
2258 return std::min(Tmp, Tmp2)-1;
2260 case Instruction::PHI: {
2261 const PHINode *PN = cast<PHINode>(U);
2262 unsigned NumIncomingValues = PN->getNumIncomingValues();
2263 // Don't analyze large in-degree PHIs.
2264 if (NumIncomingValues > 4) break;
2265 // Unreachable blocks may have zero-operand PHI nodes.
2266 if (NumIncomingValues == 0) break;
2268 // Take the minimum of all incoming values. This can't infinitely loop
2269 // because of our depth threshold.
2270 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2271 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2272 if (Tmp == 1) return Tmp;
2274 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2279 case Instruction::Trunc:
2280 // FIXME: it's tricky to do anything useful for this, but it is an important
2281 // case for targets like X86.
2284 case Instruction::ExtractElement:
2285 // Look through extract element. At the moment we keep this simple and skip
2286 // tracking the specific element. But at least we might find information
2287 // valid for all elements of the vector (for example if vector is sign
2288 // extended, shifted, etc).
2289 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2292 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2293 // use this information.
2295 // If we can examine all elements of a vector constant successfully, we're
2296 // done (we can't do any better than that). If not, keep trying.
2297 if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2300 KnownBits Known(TyBits);
2301 computeKnownBits(V, Known, Depth, Q);
2303 // If we know that the sign bit is either zero or one, determine the number of
2304 // identical bits in the top of the input value.
2305 if (Known.Zero.isSignBitSet())
2306 return std::max(FirstAnswer, Known.Zero.countLeadingOnes());
2308 if (Known.One.isSignBitSet())
2309 return std::max(FirstAnswer, Known.One.countLeadingOnes());
2311 // computeKnownBits gave us no extra information about the top bits.
2315 /// This function computes the integer multiple of Base that equals V.
2316 /// If successful, it returns true and returns the multiple in
2317 /// Multiple. If unsuccessful, it returns false. It looks
2318 /// through SExt instructions only if LookThroughSExt is true.
2319 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2320 bool LookThroughSExt, unsigned Depth) {
2321 const unsigned MaxDepth = 6;
2323 assert(V && "No Value?");
2324 assert(Depth <= MaxDepth && "Limit Search Depth");
2325 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2327 Type *T = V->getType();
2329 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2339 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2340 Constant *BaseVal = ConstantInt::get(T, Base);
2341 if (CO && CO == BaseVal) {
2343 Multiple = ConstantInt::get(T, 1);
2347 if (CI && CI->getZExtValue() % Base == 0) {
2348 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2352 if (Depth == MaxDepth) return false; // Limit search depth.
2354 Operator *I = dyn_cast<Operator>(V);
2355 if (!I) return false;
2357 switch (I->getOpcode()) {
2359 case Instruction::SExt:
2360 if (!LookThroughSExt) return false;
2361 // otherwise fall through to ZExt
2362 case Instruction::ZExt:
2363 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2364 LookThroughSExt, Depth+1);
2365 case Instruction::Shl:
2366 case Instruction::Mul: {
2367 Value *Op0 = I->getOperand(0);
2368 Value *Op1 = I->getOperand(1);
2370 if (I->getOpcode() == Instruction::Shl) {
2371 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2372 if (!Op1CI) return false;
2373 // Turn Op0 << Op1 into Op0 * 2^Op1
2374 APInt Op1Int = Op1CI->getValue();
2375 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2376 APInt API(Op1Int.getBitWidth(), 0);
2377 API.setBit(BitToSet);
2378 Op1 = ConstantInt::get(V->getContext(), API);
2381 Value *Mul0 = nullptr;
2382 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2383 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2384 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2385 if (Op1C->getType()->getPrimitiveSizeInBits() <
2386 MulC->getType()->getPrimitiveSizeInBits())
2387 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2388 if (Op1C->getType()->getPrimitiveSizeInBits() >
2389 MulC->getType()->getPrimitiveSizeInBits())
2390 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2392 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2393 Multiple = ConstantExpr::getMul(MulC, Op1C);
2397 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2398 if (Mul0CI->getValue() == 1) {
2399 // V == Base * Op1, so return Op1
2405 Value *Mul1 = nullptr;
2406 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2407 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2408 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2409 if (Op0C->getType()->getPrimitiveSizeInBits() <
2410 MulC->getType()->getPrimitiveSizeInBits())
2411 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2412 if (Op0C->getType()->getPrimitiveSizeInBits() >
2413 MulC->getType()->getPrimitiveSizeInBits())
2414 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2416 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2417 Multiple = ConstantExpr::getMul(MulC, Op0C);
2421 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2422 if (Mul1CI->getValue() == 1) {
2423 // V == Base * Op0, so return Op0
2431 // We could not determine if V is a multiple of Base.
2435 Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
2436 const TargetLibraryInfo *TLI) {
2437 const Function *F = ICS.getCalledFunction();
2439 return Intrinsic::not_intrinsic;
2441 if (F->isIntrinsic())
2442 return F->getIntrinsicID();
2445 return Intrinsic::not_intrinsic;
2448 // We're going to make assumptions on the semantics of the functions, check
2449 // that the target knows that it's available in this environment and it does
2450 // not have local linkage.
2451 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2452 return Intrinsic::not_intrinsic;
2454 if (!ICS.onlyReadsMemory())
2455 return Intrinsic::not_intrinsic;
2457 // Otherwise check if we have a call to a function that can be turned into a
2458 // vector intrinsic.
2465 return Intrinsic::sin;
2469 return Intrinsic::cos;
2473 return Intrinsic::exp;
2477 return Intrinsic::exp2;
2481 return Intrinsic::log;
2483 case LibFunc_log10f:
2484 case LibFunc_log10l:
2485 return Intrinsic::log10;
2489 return Intrinsic::log2;
2493 return Intrinsic::fabs;
2497 return Intrinsic::minnum;
2501 return Intrinsic::maxnum;
2502 case LibFunc_copysign:
2503 case LibFunc_copysignf:
2504 case LibFunc_copysignl:
2505 return Intrinsic::copysign;
2507 case LibFunc_floorf:
2508 case LibFunc_floorl:
2509 return Intrinsic::floor;
2513 return Intrinsic::ceil;
2515 case LibFunc_truncf:
2516 case LibFunc_truncl:
2517 return Intrinsic::trunc;
2521 return Intrinsic::rint;
2522 case LibFunc_nearbyint:
2523 case LibFunc_nearbyintf:
2524 case LibFunc_nearbyintl:
2525 return Intrinsic::nearbyint;
2527 case LibFunc_roundf:
2528 case LibFunc_roundl:
2529 return Intrinsic::round;
2533 return Intrinsic::pow;
2537 if (ICS->hasNoNaNs())
2538 return Intrinsic::sqrt;
2539 return Intrinsic::not_intrinsic;
2542 return Intrinsic::not_intrinsic;
2545 /// Return true if we can prove that the specified FP value is never equal to
2548 /// NOTE: this function will need to be revisited when we support non-default
2551 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
2553 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2554 return !CFP->getValueAPF().isNegZero();
2556 if (Depth == MaxDepth)
2557 return false; // Limit search depth.
2559 const Operator *I = dyn_cast<Operator>(V);
2560 if (!I) return false;
2562 // Check if the nsz fast-math flag is set
2563 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2564 if (FPO->hasNoSignedZeros())
2567 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2568 if (I->getOpcode() == Instruction::FAdd)
2569 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2570 if (CFP->isNullValue())
2573 // sitofp and uitofp turn into +0.0 for zero.
2574 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2577 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2578 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2582 // sqrt(-0.0) = -0.0, no other negative results are possible.
2583 case Intrinsic::sqrt:
2584 return CannotBeNegativeZero(CI->getArgOperand(0), TLI, Depth + 1);
2586 case Intrinsic::fabs:
2594 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2595 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2596 /// bit despite comparing equal.
2597 static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
2598 const TargetLibraryInfo *TLI,
2601 // TODO: This function does not do the right thing when SignBitOnly is true
2602 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2603 // which flips the sign bits of NaNs. See
2604 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2606 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2607 return !CFP->getValueAPF().isNegative() ||
2608 (!SignBitOnly && CFP->getValueAPF().isZero());
2611 if (Depth == MaxDepth)
2612 return false; // Limit search depth.
2614 const Operator *I = dyn_cast<Operator>(V);
2618 switch (I->getOpcode()) {
2621 // Unsigned integers are always nonnegative.
2622 case Instruction::UIToFP:
2624 case Instruction::FMul:
2625 // x*x is always non-negative or a NaN.
2626 if (I->getOperand(0) == I->getOperand(1) &&
2627 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2631 case Instruction::FAdd:
2632 case Instruction::FDiv:
2633 case Instruction::FRem:
2634 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2636 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2638 case Instruction::Select:
2639 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2641 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2643 case Instruction::FPExt:
2644 case Instruction::FPTrunc:
2645 // Widening/narrowing never change sign.
2646 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2648 case Instruction::Call:
2649 const auto *CI = cast<CallInst>(I);
2650 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2654 case Intrinsic::maxnum:
2655 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2657 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2659 case Intrinsic::minnum:
2660 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2662 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2664 case Intrinsic::exp:
2665 case Intrinsic::exp2:
2666 case Intrinsic::fabs:
2669 case Intrinsic::sqrt:
2670 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
2673 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2674 CannotBeNegativeZero(CI->getOperand(0), TLI));
2676 case Intrinsic::powi:
2677 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2678 // powi(x,n) is non-negative if n is even.
2679 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2682 // TODO: This is not correct. Given that exp is an integer, here are the
2683 // ways that pow can return a negative value:
2685 // pow(x, exp) --> negative if exp is odd and x is negative.
2686 // pow(-0, exp) --> -inf if exp is negative odd.
2687 // pow(-0, exp) --> -0 if exp is positive odd.
2688 // pow(-inf, exp) --> -0 if exp is negative odd.
2689 // pow(-inf, exp) --> -inf if exp is positive odd.
2691 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2692 // but we must return false if x == -0. Unfortunately we do not currently
2693 // have a way of expressing this constraint. See details in
2694 // https://llvm.org/bugs/show_bug.cgi?id=31702.
2695 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2698 case Intrinsic::fma:
2699 case Intrinsic::fmuladd:
2700 // x*x+y is non-negative if y is non-negative.
2701 return I->getOperand(0) == I->getOperand(1) &&
2702 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
2703 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2711 bool llvm::CannotBeOrderedLessThanZero(const Value *V,
2712 const TargetLibraryInfo *TLI) {
2713 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
2716 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
2717 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
2720 /// If the specified value can be set by repeating the same byte in memory,
2721 /// return the i8 value that it is represented with. This is
2722 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2723 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2724 /// byte store (e.g. i16 0x1234), return null.
2725 Value *llvm::isBytewiseValue(Value *V) {
2726 // All byte-wide stores are splatable, even of arbitrary variables.
2727 if (V->getType()->isIntegerTy(8)) return V;
2729 // Handle 'null' ConstantArrayZero etc.
2730 if (Constant *C = dyn_cast<Constant>(V))
2731 if (C->isNullValue())
2732 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2734 // Constant float and double values can be handled as integer values if the
2735 // corresponding integer value is "byteable". An important case is 0.0.
2736 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2737 if (CFP->getType()->isFloatTy())
2738 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2739 if (CFP->getType()->isDoubleTy())
2740 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2741 // Don't handle long double formats, which have strange constraints.
2744 // We can handle constant integers that are multiple of 8 bits.
2745 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2746 if (CI->getBitWidth() % 8 == 0) {
2747 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2749 if (!CI->getValue().isSplat(8))
2751 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2755 // A ConstantDataArray/Vector is splatable if all its members are equal and
2757 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2758 Value *Elt = CA->getElementAsConstant(0);
2759 Value *Val = isBytewiseValue(Elt);
2763 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2764 if (CA->getElementAsConstant(I) != Elt)
2770 // Conceptually, we could handle things like:
2771 // %a = zext i8 %X to i16
2772 // %b = shl i16 %a, 8
2773 // %c = or i16 %a, %b
2774 // but until there is an example that actually needs this, it doesn't seem
2775 // worth worrying about.
2780 // This is the recursive version of BuildSubAggregate. It takes a few different
2781 // arguments. Idxs is the index within the nested struct From that we are
2782 // looking at now (which is of type IndexedType). IdxSkip is the number of
2783 // indices from Idxs that should be left out when inserting into the resulting
2784 // struct. To is the result struct built so far, new insertvalue instructions
2786 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2787 SmallVectorImpl<unsigned> &Idxs,
2789 Instruction *InsertBefore) {
2790 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2792 // Save the original To argument so we can modify it
2794 // General case, the type indexed by Idxs is a struct
2795 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2796 // Process each struct element recursively
2799 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2803 // Couldn't find any inserted value for this index? Cleanup
2804 while (PrevTo != OrigTo) {
2805 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2806 PrevTo = Del->getAggregateOperand();
2807 Del->eraseFromParent();
2809 // Stop processing elements
2813 // If we successfully found a value for each of our subaggregates
2817 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2818 // the struct's elements had a value that was inserted directly. In the latter
2819 // case, perhaps we can't determine each of the subelements individually, but
2820 // we might be able to find the complete struct somewhere.
2822 // Find the value that is at that particular spot
2823 Value *V = FindInsertedValue(From, Idxs);
2828 // Insert the value in the new (sub) aggregrate
2829 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2830 "tmp", InsertBefore);
2833 // This helper takes a nested struct and extracts a part of it (which is again a
2834 // struct) into a new value. For example, given the struct:
2835 // { a, { b, { c, d }, e } }
2836 // and the indices "1, 1" this returns
2839 // It does this by inserting an insertvalue for each element in the resulting
2840 // struct, as opposed to just inserting a single struct. This will only work if
2841 // each of the elements of the substruct are known (ie, inserted into From by an
2842 // insertvalue instruction somewhere).
2844 // All inserted insertvalue instructions are inserted before InsertBefore
2845 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2846 Instruction *InsertBefore) {
2847 assert(InsertBefore && "Must have someplace to insert!");
2848 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2850 Value *To = UndefValue::get(IndexedType);
2851 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2852 unsigned IdxSkip = Idxs.size();
2854 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2857 /// Given an aggregrate and an sequence of indices, see if
2858 /// the scalar value indexed is already around as a register, for example if it
2859 /// were inserted directly into the aggregrate.
2861 /// If InsertBefore is not null, this function will duplicate (modified)
2862 /// insertvalues when a part of a nested struct is extracted.
2863 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2864 Instruction *InsertBefore) {
2865 // Nothing to index? Just return V then (this is useful at the end of our
2867 if (idx_range.empty())
2869 // We have indices, so V should have an indexable type.
2870 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2871 "Not looking at a struct or array?");
2872 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2873 "Invalid indices for type?");
2875 if (Constant *C = dyn_cast<Constant>(V)) {
2876 C = C->getAggregateElement(idx_range[0]);
2877 if (!C) return nullptr;
2878 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2881 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2882 // Loop the indices for the insertvalue instruction in parallel with the
2883 // requested indices
2884 const unsigned *req_idx = idx_range.begin();
2885 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2886 i != e; ++i, ++req_idx) {
2887 if (req_idx == idx_range.end()) {
2888 // We can't handle this without inserting insertvalues
2892 // The requested index identifies a part of a nested aggregate. Handle
2893 // this specially. For example,
2894 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2895 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2896 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2897 // This can be changed into
2898 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2899 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2900 // which allows the unused 0,0 element from the nested struct to be
2902 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2906 // This insert value inserts something else than what we are looking for.
2907 // See if the (aggregate) value inserted into has the value we are
2908 // looking for, then.
2910 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2913 // If we end up here, the indices of the insertvalue match with those
2914 // requested (though possibly only partially). Now we recursively look at
2915 // the inserted value, passing any remaining indices.
2916 return FindInsertedValue(I->getInsertedValueOperand(),
2917 makeArrayRef(req_idx, idx_range.end()),
2921 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2922 // If we're extracting a value from an aggregate that was extracted from
2923 // something else, we can extract from that something else directly instead.
2924 // However, we will need to chain I's indices with the requested indices.
2926 // Calculate the number of indices required
2927 unsigned size = I->getNumIndices() + idx_range.size();
2928 // Allocate some space to put the new indices in
2929 SmallVector<unsigned, 5> Idxs;
2931 // Add indices from the extract value instruction
2932 Idxs.append(I->idx_begin(), I->idx_end());
2934 // Add requested indices
2935 Idxs.append(idx_range.begin(), idx_range.end());
2937 assert(Idxs.size() == size
2938 && "Number of indices added not correct?");
2940 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2942 // Otherwise, we don't know (such as, extracting from a function return value
2943 // or load instruction)
2947 /// Analyze the specified pointer to see if it can be expressed as a base
2948 /// pointer plus a constant offset. Return the base and offset to the caller.
2949 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2950 const DataLayout &DL) {
2951 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType());
2952 APInt ByteOffset(BitWidth, 0);
2954 // We walk up the defs but use a visited set to handle unreachable code. In
2955 // that case, we stop after accumulating the cycle once (not that it
2957 SmallPtrSet<Value *, 16> Visited;
2958 while (Visited.insert(Ptr).second) {
2959 if (Ptr->getType()->isVectorTy())
2962 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2963 // If one of the values we have visited is an addrspacecast, then
2964 // the pointer type of this GEP may be different from the type
2965 // of the Ptr parameter which was passed to this function. This
2966 // means when we construct GEPOffset, we need to use the size
2967 // of GEP's pointer type rather than the size of the original
2969 APInt GEPOffset(DL.getPointerTypeSizeInBits(Ptr->getType()), 0);
2970 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
2973 ByteOffset += GEPOffset.getSExtValue();
2975 Ptr = GEP->getPointerOperand();
2976 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2977 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2978 Ptr = cast<Operator>(Ptr)->getOperand(0);
2979 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2980 if (GA->isInterposable())
2982 Ptr = GA->getAliasee();
2987 Offset = ByteOffset.getSExtValue();
2991 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP) {
2992 // Make sure the GEP has exactly three arguments.
2993 if (GEP->getNumOperands() != 3)
2996 // Make sure the index-ee is a pointer to array of i8.
2997 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
2998 if (!AT || !AT->getElementType()->isIntegerTy(8))
3001 // Check to make sure that the first operand of the GEP is an integer and
3002 // has value 0 so that we are sure we're indexing into the initializer.
3003 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
3004 if (!FirstIdx || !FirstIdx->isZero())
3010 /// This function computes the length of a null-terminated C string pointed to
3011 /// by V. If successful, it returns true and returns the string in Str.
3012 /// If unsuccessful, it returns false.
3013 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
3014 uint64_t Offset, bool TrimAtNul) {
3017 // Look through bitcast instructions and geps.
3018 V = V->stripPointerCasts();
3020 // If the value is a GEP instruction or constant expression, treat it as an
3022 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3023 // The GEP operator should be based on a pointer to string constant, and is
3024 // indexing into the string constant.
3025 if (!isGEPBasedOnPointerToString(GEP))
3028 // If the second index isn't a ConstantInt, then this is a variable index
3029 // into the array. If this occurs, we can't say anything meaningful about
3031 uint64_t StartIdx = 0;
3032 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3033 StartIdx = CI->getZExtValue();
3036 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset,
3040 // The GEP instruction, constant or instruction, must reference a global
3041 // variable that is a constant and is initialized. The referenced constant
3042 // initializer is the array that we'll use for optimization.
3043 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3044 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3047 // Handle the all-zeros case.
3048 if (GV->getInitializer()->isNullValue()) {
3049 // This is a degenerate case. The initializer is constant zero so the
3050 // length of the string must be zero.
3055 // This must be a ConstantDataArray.
3056 const auto *Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3057 if (!Array || !Array->isString())
3060 // Get the number of elements in the array.
3061 uint64_t NumElts = Array->getType()->getArrayNumElements();
3063 // Start out with the entire array in the StringRef.
3064 Str = Array->getAsString();
3066 if (Offset > NumElts)
3069 // Skip over 'offset' bytes.
3070 Str = Str.substr(Offset);
3073 // Trim off the \0 and anything after it. If the array is not nul
3074 // terminated, we just return the whole end of string. The client may know
3075 // some other way that the string is length-bound.
3076 Str = Str.substr(0, Str.find('\0'));
3081 // These next two are very similar to the above, but also look through PHI
3083 // TODO: See if we can integrate these two together.
3085 /// If we can compute the length of the string pointed to by
3086 /// the specified pointer, return 'len+1'. If we can't, return 0.
3087 static uint64_t GetStringLengthH(const Value *V,
3088 SmallPtrSetImpl<const PHINode*> &PHIs) {
3089 // Look through noop bitcast instructions.
3090 V = V->stripPointerCasts();
3092 // If this is a PHI node, there are two cases: either we have already seen it
3094 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3095 if (!PHIs.insert(PN).second)
3096 return ~0ULL; // already in the set.
3098 // If it was new, see if all the input strings are the same length.
3099 uint64_t LenSoFar = ~0ULL;
3100 for (Value *IncValue : PN->incoming_values()) {
3101 uint64_t Len = GetStringLengthH(IncValue, PHIs);
3102 if (Len == 0) return 0; // Unknown length -> unknown.
3104 if (Len == ~0ULL) continue;
3106 if (Len != LenSoFar && LenSoFar != ~0ULL)
3107 return 0; // Disagree -> unknown.
3111 // Success, all agree.
3115 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3116 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3117 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
3118 if (Len1 == 0) return 0;
3119 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
3120 if (Len2 == 0) return 0;
3121 if (Len1 == ~0ULL) return Len2;
3122 if (Len2 == ~0ULL) return Len1;
3123 if (Len1 != Len2) return 0;
3127 // Otherwise, see if we can read the string.
3129 if (!getConstantStringInfo(V, StrData))
3132 return StrData.size()+1;
3135 /// If we can compute the length of the string pointed to by
3136 /// the specified pointer, return 'len+1'. If we can't, return 0.
3137 uint64_t llvm::GetStringLength(const Value *V) {
3138 if (!V->getType()->isPointerTy()) return 0;
3140 SmallPtrSet<const PHINode*, 32> PHIs;
3141 uint64_t Len = GetStringLengthH(V, PHIs);
3142 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3143 // an empty string as a length.
3144 return Len == ~0ULL ? 1 : Len;
3147 /// \brief \p PN defines a loop-variant pointer to an object. Check if the
3148 /// previous iteration of the loop was referring to the same object as \p PN.
3149 static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
3150 const LoopInfo *LI) {
3151 // Find the loop-defined value.
3152 Loop *L = LI->getLoopFor(PN->getParent());
3153 if (PN->getNumIncomingValues() != 2)
3156 // Find the value from previous iteration.
3157 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3158 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3159 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3160 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3163 // If a new pointer is loaded in the loop, the pointer references a different
3164 // object in every iteration. E.g.:
3168 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3169 if (!L->isLoopInvariant(Load->getPointerOperand()))
3174 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3175 unsigned MaxLookup) {
3176 if (!V->getType()->isPointerTy())
3178 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3179 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3180 V = GEP->getPointerOperand();
3181 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3182 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3183 V = cast<Operator>(V)->getOperand(0);
3184 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3185 if (GA->isInterposable())
3187 V = GA->getAliasee();
3188 } else if (isa<AllocaInst>(V)) {
3189 // An alloca can't be further simplified.
3192 if (auto CS = CallSite(V))
3193 if (Value *RV = CS.getReturnedArgOperand()) {
3198 // See if InstructionSimplify knows any relevant tricks.
3199 if (Instruction *I = dyn_cast<Instruction>(V))
3200 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3201 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) {
3208 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3213 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3214 const DataLayout &DL, LoopInfo *LI,
3215 unsigned MaxLookup) {
3216 SmallPtrSet<Value *, 4> Visited;
3217 SmallVector<Value *, 4> Worklist;
3218 Worklist.push_back(V);
3220 Value *P = Worklist.pop_back_val();
3221 P = GetUnderlyingObject(P, DL, MaxLookup);
3223 if (!Visited.insert(P).second)
3226 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3227 Worklist.push_back(SI->getTrueValue());
3228 Worklist.push_back(SI->getFalseValue());
3232 if (PHINode *PN = dyn_cast<PHINode>(P)) {
3233 // If this PHI changes the underlying object in every iteration of the
3234 // loop, don't look through it. Consider:
3237 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3241 // Prev is tracking Curr one iteration behind so they refer to different
3242 // underlying objects.
3243 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3244 isSameUnderlyingObjectInLoop(PN, LI))
3245 for (Value *IncValue : PN->incoming_values())
3246 Worklist.push_back(IncValue);
3250 Objects.push_back(P);
3251 } while (!Worklist.empty());
3254 /// Return true if the only users of this pointer are lifetime markers.
3255 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3256 for (const User *U : V->users()) {
3257 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3258 if (!II) return false;
3260 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3261 II->getIntrinsicID() != Intrinsic::lifetime_end)
3267 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3268 const Instruction *CtxI,
3269 const DominatorTree *DT) {
3270 const Operator *Inst = dyn_cast<Operator>(V);
3274 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3275 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3279 switch (Inst->getOpcode()) {
3282 case Instruction::UDiv:
3283 case Instruction::URem: {
3284 // x / y is undefined if y == 0.
3286 if (match(Inst->getOperand(1), m_APInt(V)))
3290 case Instruction::SDiv:
3291 case Instruction::SRem: {
3292 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3293 const APInt *Numerator, *Denominator;
3294 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3296 // We cannot hoist this division if the denominator is 0.
3297 if (*Denominator == 0)
3299 // It's safe to hoist if the denominator is not 0 or -1.
3300 if (*Denominator != -1)
3302 // At this point we know that the denominator is -1. It is safe to hoist as
3303 // long we know that the numerator is not INT_MIN.
3304 if (match(Inst->getOperand(0), m_APInt(Numerator)))
3305 return !Numerator->isMinSignedValue();
3306 // The numerator *might* be MinSignedValue.
3309 case Instruction::Load: {
3310 const LoadInst *LI = cast<LoadInst>(Inst);
3311 if (!LI->isUnordered() ||
3312 // Speculative load may create a race that did not exist in the source.
3313 LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3314 // Speculative load may load data from dirty regions.
3315 LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress))
3317 const DataLayout &DL = LI->getModule()->getDataLayout();
3318 return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
3319 LI->getAlignment(), DL, CtxI, DT);
3321 case Instruction::Call: {
3322 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
3323 switch (II->getIntrinsicID()) {
3324 // These synthetic intrinsics have no side-effects and just mark
3325 // information about their operands.
3326 // FIXME: There are other no-op synthetic instructions that potentially
3327 // should be considered at least *safe* to speculate...
3328 case Intrinsic::dbg_declare:
3329 case Intrinsic::dbg_value:
3332 case Intrinsic::bitreverse:
3333 case Intrinsic::bswap:
3334 case Intrinsic::ctlz:
3335 case Intrinsic::ctpop:
3336 case Intrinsic::cttz:
3337 case Intrinsic::objectsize:
3338 case Intrinsic::sadd_with_overflow:
3339 case Intrinsic::smul_with_overflow:
3340 case Intrinsic::ssub_with_overflow:
3341 case Intrinsic::uadd_with_overflow:
3342 case Intrinsic::umul_with_overflow:
3343 case Intrinsic::usub_with_overflow:
3345 // These intrinsics are defined to have the same behavior as libm
3346 // functions except for setting errno.
3347 case Intrinsic::sqrt:
3348 case Intrinsic::fma:
3349 case Intrinsic::fmuladd:
3351 // These intrinsics are defined to have the same behavior as libm
3352 // functions, and the corresponding libm functions never set errno.
3353 case Intrinsic::trunc:
3354 case Intrinsic::copysign:
3355 case Intrinsic::fabs:
3356 case Intrinsic::minnum:
3357 case Intrinsic::maxnum:
3359 // These intrinsics are defined to have the same behavior as libm
3360 // functions, which never overflow when operating on the IEEE754 types
3361 // that we support, and never set errno otherwise.
3362 case Intrinsic::ceil:
3363 case Intrinsic::floor:
3364 case Intrinsic::nearbyint:
3365 case Intrinsic::rint:
3366 case Intrinsic::round:
3368 // These intrinsics do not correspond to any libm function, and
3369 // do not set errno.
3370 case Intrinsic::powi:
3372 // TODO: are convert_{from,to}_fp16 safe?
3373 // TODO: can we list target-specific intrinsics here?
3377 return false; // The called function could have undefined behavior or
3378 // side-effects, even if marked readnone nounwind.
3380 case Instruction::VAArg:
3381 case Instruction::Alloca:
3382 case Instruction::Invoke:
3383 case Instruction::PHI:
3384 case Instruction::Store:
3385 case Instruction::Ret:
3386 case Instruction::Br:
3387 case Instruction::IndirectBr:
3388 case Instruction::Switch:
3389 case Instruction::Unreachable:
3390 case Instruction::Fence:
3391 case Instruction::AtomicRMW:
3392 case Instruction::AtomicCmpXchg:
3393 case Instruction::LandingPad:
3394 case Instruction::Resume:
3395 case Instruction::CatchSwitch:
3396 case Instruction::CatchPad:
3397 case Instruction::CatchRet:
3398 case Instruction::CleanupPad:
3399 case Instruction::CleanupRet:
3400 return false; // Misc instructions which have effects
3404 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3405 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3408 /// Return true if we know that the specified value is never null.
3409 bool llvm::isKnownNonNull(const Value *V) {
3410 assert(V->getType()->isPointerTy() && "V must be pointer type");
3412 // Alloca never returns null, malloc might.
3413 if (isa<AllocaInst>(V)) return true;
3415 // A byval, inalloca, or nonnull argument is never null.
3416 if (const Argument *A = dyn_cast<Argument>(V))
3417 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
3419 // A global variable in address space 0 is non null unless extern weak
3420 // or an absolute symbol reference. Other address spaces may have null as a
3421 // valid address for a global, so we can't assume anything.
3422 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
3423 return !GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
3424 GV->getType()->getAddressSpace() == 0;
3426 // A Load tagged with nonnull metadata is never null.
3427 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
3428 return LI->getMetadata(LLVMContext::MD_nonnull);
3430 if (auto CS = ImmutableCallSite(V))
3431 if (CS.isReturnNonNull())
3437 static bool isKnownNonNullFromDominatingCondition(const Value *V,
3438 const Instruction *CtxI,
3439 const DominatorTree *DT) {
3440 assert(V->getType()->isPointerTy() && "V must be pointer type");
3441 assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
3442 assert(CtxI && "Context instruction required for analysis");
3443 assert(DT && "Dominator tree required for analysis");
3445 unsigned NumUsesExplored = 0;
3446 for (auto *U : V->users()) {
3447 // Avoid massive lists
3448 if (NumUsesExplored >= DomConditionsMaxUses)
3452 // If the value is used as an argument to a call or invoke, then argument
3453 // attributes may provide an answer about null-ness.
3454 if (auto CS = ImmutableCallSite(U))
3455 if (auto *CalledFunc = CS.getCalledFunction())
3456 for (const Argument &Arg : CalledFunc->args())
3457 if (CS.getArgOperand(Arg.getArgNo()) == V &&
3458 Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
3461 // Consider only compare instructions uniquely controlling a branch
3462 CmpInst::Predicate Pred;
3463 if (!match(const_cast<User *>(U),
3464 m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
3465 (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
3468 for (auto *CmpU : U->users()) {
3469 if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) {
3470 assert(BI->isConditional() && "uses a comparison!");
3472 BasicBlock *NonNullSuccessor =
3473 BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
3474 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
3475 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
3477 } else if (Pred == ICmpInst::ICMP_NE &&
3478 match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) &&
3479 DT->dominates(cast<Instruction>(CmpU), CtxI)) {
3488 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI,
3489 const DominatorTree *DT) {
3490 if (isa<ConstantPointerNull>(V) || isa<UndefValue>(V))
3493 if (isKnownNonNull(V))
3499 return ::isKnownNonNullFromDominatingCondition(V, CtxI, DT);
3502 OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS,
3504 const DataLayout &DL,
3505 AssumptionCache *AC,
3506 const Instruction *CxtI,
3507 const DominatorTree *DT) {
3508 // Multiplying n * m significant bits yields a result of n + m significant
3509 // bits. If the total number of significant bits does not exceed the
3510 // result bit width (minus 1), there is no overflow.
3511 // This means if we have enough leading zero bits in the operands
3512 // we can guarantee that the result does not overflow.
3513 // Ref: "Hacker's Delight" by Henry Warren
3514 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3515 KnownBits LHSKnown(BitWidth);
3516 KnownBits RHSKnown(BitWidth);
3517 computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3518 computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3519 // Note that underestimating the number of zero bits gives a more
3520 // conservative answer.
3521 unsigned ZeroBits = LHSKnown.Zero.countLeadingOnes() +
3522 RHSKnown.Zero.countLeadingOnes();
3523 // First handle the easy case: if we have enough zero bits there's
3524 // definitely no overflow.
3525 if (ZeroBits >= BitWidth)
3526 return OverflowResult::NeverOverflows;
3528 // Get the largest possible values for each operand.
3529 APInt LHSMax = ~LHSKnown.Zero;
3530 APInt RHSMax = ~RHSKnown.Zero;
3532 // We know the multiply operation doesn't overflow if the maximum values for
3533 // each operand will not overflow after we multiply them together.
3535 (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3537 return OverflowResult::NeverOverflows;
3539 // We know it always overflows if multiplying the smallest possible values for
3540 // the operands also results in overflow.
3542 (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3544 return OverflowResult::AlwaysOverflows;
3546 return OverflowResult::MayOverflow;
3549 OverflowResult llvm::computeOverflowForUnsignedAdd(const Value *LHS,
3551 const DataLayout &DL,
3552 AssumptionCache *AC,
3553 const Instruction *CxtI,
3554 const DominatorTree *DT) {
3555 bool LHSKnownNonNegative, LHSKnownNegative;
3556 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3558 if (LHSKnownNonNegative || LHSKnownNegative) {
3559 bool RHSKnownNonNegative, RHSKnownNegative;
3560 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3563 if (LHSKnownNegative && RHSKnownNegative) {
3564 // The sign bit is set in both cases: this MUST overflow.
3565 // Create a simple add instruction, and insert it into the struct.
3566 return OverflowResult::AlwaysOverflows;
3569 if (LHSKnownNonNegative && RHSKnownNonNegative) {
3570 // The sign bit is clear in both cases: this CANNOT overflow.
3571 // Create a simple add instruction, and insert it into the struct.
3572 return OverflowResult::NeverOverflows;
3576 return OverflowResult::MayOverflow;
3579 static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
3581 const AddOperator *Add,
3582 const DataLayout &DL,
3583 AssumptionCache *AC,
3584 const Instruction *CxtI,
3585 const DominatorTree *DT) {
3586 if (Add && Add->hasNoSignedWrap()) {
3587 return OverflowResult::NeverOverflows;
3590 bool LHSKnownNonNegative, LHSKnownNegative;
3591 bool RHSKnownNonNegative, RHSKnownNegative;
3592 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
3594 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
3597 if ((LHSKnownNonNegative && RHSKnownNegative) ||
3598 (LHSKnownNegative && RHSKnownNonNegative)) {
3599 // The sign bits are opposite: this CANNOT overflow.
3600 return OverflowResult::NeverOverflows;
3603 // The remaining code needs Add to be available. Early returns if not so.
3605 return OverflowResult::MayOverflow;
3607 // If the sign of Add is the same as at least one of the operands, this add
3608 // CANNOT overflow. This is particularly useful when the sum is
3609 // @llvm.assume'ed non-negative rather than proved so from analyzing its
3611 bool LHSOrRHSKnownNonNegative =
3612 (LHSKnownNonNegative || RHSKnownNonNegative);
3613 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative);
3614 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3615 bool AddKnownNonNegative, AddKnownNegative;
3616 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL,
3617 /*Depth=*/0, AC, CxtI, DT);
3618 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) ||
3619 (AddKnownNegative && LHSOrRHSKnownNegative)) {
3620 return OverflowResult::NeverOverflows;
3624 return OverflowResult::MayOverflow;
3627 bool llvm::isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
3628 const DominatorTree &DT) {
3630 auto IID = II->getIntrinsicID();
3631 assert((IID == Intrinsic::sadd_with_overflow ||
3632 IID == Intrinsic::uadd_with_overflow ||
3633 IID == Intrinsic::ssub_with_overflow ||
3634 IID == Intrinsic::usub_with_overflow ||
3635 IID == Intrinsic::smul_with_overflow ||
3636 IID == Intrinsic::umul_with_overflow) &&
3637 "Not an overflow intrinsic!");
3640 SmallVector<const BranchInst *, 2> GuardingBranches;
3641 SmallVector<const ExtractValueInst *, 2> Results;
3643 for (const User *U : II->users()) {
3644 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
3645 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
3647 if (EVI->getIndices()[0] == 0)
3648 Results.push_back(EVI);
3650 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
3652 for (const auto *U : EVI->users())
3653 if (const auto *B = dyn_cast<BranchInst>(U)) {
3654 assert(B->isConditional() && "How else is it using an i1?");
3655 GuardingBranches.push_back(B);
3659 // We are using the aggregate directly in a way we don't want to analyze
3660 // here (storing it to a global, say).
3665 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
3666 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
3667 if (!NoWrapEdge.isSingleEdge())
3670 // Check if all users of the add are provably no-wrap.
3671 for (const auto *Result : Results) {
3672 // If the extractvalue itself is not executed on overflow, the we don't
3673 // need to check each use separately, since domination is transitive.
3674 if (DT.dominates(NoWrapEdge, Result->getParent()))
3677 for (auto &RU : Result->uses())
3678 if (!DT.dominates(NoWrapEdge, RU))
3685 return any_of(GuardingBranches, AllUsesGuardedByBranch);
3689 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
3690 const DataLayout &DL,
3691 AssumptionCache *AC,
3692 const Instruction *CxtI,
3693 const DominatorTree *DT) {
3694 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
3695 Add, DL, AC, CxtI, DT);
3698 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
3700 const DataLayout &DL,
3701 AssumptionCache *AC,
3702 const Instruction *CxtI,
3703 const DominatorTree *DT) {
3704 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
3707 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
3708 // A memory operation returns normally if it isn't volatile. A volatile
3709 // operation is allowed to trap.
3711 // An atomic operation isn't guaranteed to return in a reasonable amount of
3712 // time because it's possible for another thread to interfere with it for an
3713 // arbitrary length of time, but programs aren't allowed to rely on that.
3714 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
3715 return !LI->isVolatile();
3716 if (const StoreInst *SI = dyn_cast<StoreInst>(I))
3717 return !SI->isVolatile();
3718 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
3719 return !CXI->isVolatile();
3720 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
3721 return !RMWI->isVolatile();
3722 if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
3723 return !MII->isVolatile();
3725 // If there is no successor, then execution can't transfer to it.
3726 if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
3727 return !CRI->unwindsToCaller();
3728 if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
3729 return !CatchSwitch->unwindsToCaller();
3730 if (isa<ResumeInst>(I))
3732 if (isa<ReturnInst>(I))
3734 if (isa<UnreachableInst>(I))
3737 // Calls can throw, or contain an infinite loop, or kill the process.
3738 if (auto CS = ImmutableCallSite(I)) {
3739 // Call sites that throw have implicit non-local control flow.
3740 if (!CS.doesNotThrow())
3743 // Non-throwing call sites can loop infinitely, call exit/pthread_exit
3744 // etc. and thus not return. However, LLVM already assumes that
3746 // - Thread exiting actions are modeled as writes to memory invisible to
3749 // - Loops that don't have side effects (side effects are volatile/atomic
3750 // stores and IO) always terminate (see http://llvm.org/PR965).
3751 // Furthermore IO itself is also modeled as writes to memory invisible to
3754 // We rely on those assumptions here, and use the memory effects of the call
3755 // target as a proxy for checking that it always returns.
3757 // FIXME: This isn't aggressive enough; a call which only writes to a global
3758 // is guaranteed to return.
3759 return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
3760 match(I, m_Intrinsic<Intrinsic::assume>());
3763 // Other instructions return normally.
3767 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
3769 // The loop header is guaranteed to be executed for every iteration.
3771 // FIXME: Relax this constraint to cover all basic blocks that are
3772 // guaranteed to be executed at every iteration.
3773 if (I->getParent() != L->getHeader()) return false;
3775 for (const Instruction &LI : *L->getHeader()) {
3776 if (&LI == I) return true;
3777 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
3779 llvm_unreachable("Instruction not contained in its own parent basic block.");
3782 bool llvm::propagatesFullPoison(const Instruction *I) {
3783 switch (I->getOpcode()) {
3784 case Instruction::Add:
3785 case Instruction::Sub:
3786 case Instruction::Xor:
3787 case Instruction::Trunc:
3788 case Instruction::BitCast:
3789 case Instruction::AddrSpaceCast:
3790 case Instruction::Mul:
3791 case Instruction::Shl:
3792 case Instruction::GetElementPtr:
3793 // These operations all propagate poison unconditionally. Note that poison
3794 // is not any particular value, so xor or subtraction of poison with
3795 // itself still yields poison, not zero.
3798 case Instruction::AShr:
3799 case Instruction::SExt:
3800 // For these operations, one bit of the input is replicated across
3801 // multiple output bits. A replicated poison bit is still poison.
3804 case Instruction::ICmp:
3805 // Comparing poison with any value yields poison. This is why, for
3806 // instance, x s< (x +nsw 1) can be folded to true.
3814 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
3815 switch (I->getOpcode()) {
3816 case Instruction::Store:
3817 return cast<StoreInst>(I)->getPointerOperand();
3819 case Instruction::Load:
3820 return cast<LoadInst>(I)->getPointerOperand();
3822 case Instruction::AtomicCmpXchg:
3823 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
3825 case Instruction::AtomicRMW:
3826 return cast<AtomicRMWInst>(I)->getPointerOperand();
3828 case Instruction::UDiv:
3829 case Instruction::SDiv:
3830 case Instruction::URem:
3831 case Instruction::SRem:
3832 return I->getOperand(1);
3839 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) {
3840 // We currently only look for uses of poison values within the same basic
3841 // block, as that makes it easier to guarantee that the uses will be
3842 // executed given that PoisonI is executed.
3844 // FIXME: Expand this to consider uses beyond the same basic block. To do
3845 // this, look out for the distinction between post-dominance and strong
3847 const BasicBlock *BB = PoisonI->getParent();
3849 // Set of instructions that we have proved will yield poison if PoisonI
3851 SmallSet<const Value *, 16> YieldsPoison;
3852 SmallSet<const BasicBlock *, 4> Visited;
3853 YieldsPoison.insert(PoisonI);
3854 Visited.insert(PoisonI->getParent());
3856 BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
3859 while (Iter++ < MaxDepth) {
3860 for (auto &I : make_range(Begin, End)) {
3861 if (&I != PoisonI) {
3862 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
3863 if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
3865 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
3869 // Mark poison that propagates from I through uses of I.
3870 if (YieldsPoison.count(&I)) {
3871 for (const User *User : I.users()) {
3872 const Instruction *UserI = cast<Instruction>(User);
3873 if (propagatesFullPoison(UserI))
3874 YieldsPoison.insert(User);
3879 if (auto *NextBB = BB->getSingleSuccessor()) {
3880 if (Visited.insert(NextBB).second) {
3882 Begin = BB->getFirstNonPHI()->getIterator();
3893 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
3897 if (auto *C = dyn_cast<ConstantFP>(V))
3902 static bool isKnownNonZero(const Value *V) {
3903 if (auto *C = dyn_cast<ConstantFP>(V))
3904 return !C->isZero();
3908 /// Match non-obvious integer minimum and maximum sequences.
3909 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
3910 Value *CmpLHS, Value *CmpRHS,
3911 Value *TrueVal, Value *FalseVal,
3912 Value *&LHS, Value *&RHS) {
3913 // Assume success. If there's no match, callers should not use these anyway.
3917 // Recognize variations of:
3918 // CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
3920 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
3923 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
3924 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
3925 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
3926 return {SPF_SMAX, SPNB_NA, false};
3928 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
3929 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
3930 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
3931 return {SPF_SMIN, SPNB_NA, false};
3933 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
3934 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
3935 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
3936 return {SPF_UMAX, SPNB_NA, false};
3938 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
3939 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
3940 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
3941 return {SPF_UMIN, SPNB_NA, false};
3944 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
3945 return {SPF_UNKNOWN, SPNB_NA, false};
3948 // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
3949 // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
3950 if (match(TrueVal, m_Zero()) &&
3951 match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
3952 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
3955 // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
3956 // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
3957 if (match(FalseVal, m_Zero()) &&
3958 match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
3959 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
3961 if (!match(CmpRHS, m_APInt(C1)))
3962 return {SPF_UNKNOWN, SPNB_NA, false};
3964 // An unsigned min/max can be written with a signed compare.
3966 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
3967 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
3968 // Is the sign bit set?
3969 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
3970 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
3971 if (Pred == CmpInst::ICMP_SLT && *C1 == 0 && C2->isMaxSignedValue())
3972 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
3974 // Is the sign bit clear?
3975 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
3976 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
3977 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
3978 C2->isMinSignedValue())
3979 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
3982 // Look through 'not' ops to find disguised signed min/max.
3983 // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
3984 // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
3985 if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
3986 match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
3987 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
3989 // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
3990 // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
3991 if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
3992 match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
3993 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
3995 return {SPF_UNKNOWN, SPNB_NA, false};
3998 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
4000 Value *CmpLHS, Value *CmpRHS,
4001 Value *TrueVal, Value *FalseVal,
4002 Value *&LHS, Value *&RHS) {
4006 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may
4007 // return inconsistent results between implementations.
4008 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4009 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4010 // Therefore we behave conservatively and only proceed if at least one of the
4011 // operands is known to not be zero, or if we don't care about signed zeroes.
4014 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
4015 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
4016 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4017 !isKnownNonZero(CmpRHS))
4018 return {SPF_UNKNOWN, SPNB_NA, false};
4021 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4022 bool Ordered = false;
4024 // When given one NaN and one non-NaN input:
4025 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4026 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4027 // ordered comparison fails), which could be NaN or non-NaN.
4028 // so here we discover exactly what NaN behavior is required/accepted.
4029 if (CmpInst::isFPPredicate(Pred)) {
4030 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4031 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4033 if (LHSSafe && RHSSafe) {
4034 // Both operands are known non-NaN.
4035 NaNBehavior = SPNB_RETURNS_ANY;
4036 } else if (CmpInst::isOrdered(Pred)) {
4037 // An ordered comparison will return false when given a NaN, so it
4041 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
4042 NaNBehavior = SPNB_RETURNS_NAN;
4044 NaNBehavior = SPNB_RETURNS_OTHER;
4046 // Completely unsafe.
4047 return {SPF_UNKNOWN, SPNB_NA, false};
4050 // An unordered comparison will return true when given a NaN, so it
4053 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
4054 NaNBehavior = SPNB_RETURNS_OTHER;
4056 NaNBehavior = SPNB_RETURNS_NAN;
4058 // Completely unsafe.
4059 return {SPF_UNKNOWN, SPNB_NA, false};
4063 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
4064 std::swap(CmpLHS, CmpRHS);
4065 Pred = CmpInst::getSwappedPredicate(Pred);
4066 if (NaNBehavior == SPNB_RETURNS_NAN)
4067 NaNBehavior = SPNB_RETURNS_OTHER;
4068 else if (NaNBehavior == SPNB_RETURNS_OTHER)
4069 NaNBehavior = SPNB_RETURNS_NAN;
4073 // ([if]cmp X, Y) ? X : Y
4074 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
4076 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4077 case ICmpInst::ICMP_UGT:
4078 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4079 case ICmpInst::ICMP_SGT:
4080 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4081 case ICmpInst::ICMP_ULT:
4082 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4083 case ICmpInst::ICMP_SLT:
4084 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4085 case FCmpInst::FCMP_UGT:
4086 case FCmpInst::FCMP_UGE:
4087 case FCmpInst::FCMP_OGT:
4088 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4089 case FCmpInst::FCMP_ULT:
4090 case FCmpInst::FCMP_ULE:
4091 case FCmpInst::FCMP_OLT:
4092 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4097 if (match(CmpRHS, m_APInt(C1))) {
4098 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
4099 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
4101 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
4102 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
4103 if (Pred == ICmpInst::ICMP_SGT && (*C1 == 0 || C1->isAllOnesValue())) {
4104 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
4107 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
4108 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
4109 if (Pred == ICmpInst::ICMP_SLT && (*C1 == 0 || *C1 == 1)) {
4110 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
4115 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4118 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4119 Instruction::CastOps *CastOp) {
4120 auto *Cast1 = dyn_cast<CastInst>(V1);
4124 *CastOp = Cast1->getOpcode();
4125 Type *SrcTy = Cast1->getSrcTy();
4126 if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4127 // If V1 and V2 are both the same cast from the same type, look through V1.
4128 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4129 return Cast2->getOperand(0);
4133 auto *C = dyn_cast<Constant>(V2);
4137 Constant *CastedTo = nullptr;
4139 case Instruction::ZExt:
4140 if (CmpI->isUnsigned())
4141 CastedTo = ConstantExpr::getTrunc(C, SrcTy);
4143 case Instruction::SExt:
4144 if (CmpI->isSigned())
4145 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
4147 case Instruction::Trunc:
4148 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
4150 case Instruction::FPTrunc:
4151 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
4153 case Instruction::FPExt:
4154 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
4156 case Instruction::FPToUI:
4157 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
4159 case Instruction::FPToSI:
4160 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
4162 case Instruction::UIToFP:
4163 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
4165 case Instruction::SIToFP:
4166 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
4175 // Make sure the cast doesn't lose any information.
4176 Constant *CastedBack =
4177 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
4178 if (CastedBack != C)
4184 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
4185 Instruction::CastOps *CastOp) {
4186 SelectInst *SI = dyn_cast<SelectInst>(V);
4187 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4189 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4190 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4192 CmpInst::Predicate Pred = CmpI->getPredicate();
4193 Value *CmpLHS = CmpI->getOperand(0);
4194 Value *CmpRHS = CmpI->getOperand(1);
4195 Value *TrueVal = SI->getTrueValue();
4196 Value *FalseVal = SI->getFalseValue();
4198 if (isa<FPMathOperator>(CmpI))
4199 FMF = CmpI->getFastMathFlags();
4202 if (CmpI->isEquality())
4203 return {SPF_UNKNOWN, SPNB_NA, false};
4205 // Deal with type mismatches.
4206 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4207 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp))
4208 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4209 cast<CastInst>(TrueVal)->getOperand(0), C,
4211 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp))
4212 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4213 C, cast<CastInst>(FalseVal)->getOperand(0),
4216 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4220 /// Return true if "icmp Pred LHS RHS" is always true.
4221 static bool isTruePredicate(CmpInst::Predicate Pred,
4222 const Value *LHS, const Value *RHS,
4223 const DataLayout &DL, unsigned Depth,
4224 AssumptionCache *AC, const Instruction *CxtI,
4225 const DominatorTree *DT) {
4226 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
4227 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
4234 case CmpInst::ICMP_SLE: {
4237 // LHS s<= LHS +_{nsw} C if C >= 0
4238 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
4239 return !C->isNegative();
4243 case CmpInst::ICMP_ULE: {
4246 // LHS u<= LHS +_{nuw} C for any C
4247 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
4250 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
4251 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
4253 const APInt *&CA, const APInt *&CB) {
4254 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
4255 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
4258 // If X & C == 0 then (X | C) == X +_{nuw} C
4259 if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
4260 match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
4261 KnownBits Known(CA->getBitWidth());
4262 computeKnownBits(X, Known, DL, Depth + 1, AC, CxtI, DT);
4264 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
4272 const APInt *CLHS, *CRHS;
4273 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
4274 return CLHS->ule(*CRHS);
4281 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
4282 /// ALHS ARHS" is true. Otherwise, return None.
4283 static Optional<bool>
4284 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
4285 const Value *ARHS, const Value *BLHS,
4286 const Value *BRHS, const DataLayout &DL,
4287 unsigned Depth, AssumptionCache *AC,
4288 const Instruction *CxtI, const DominatorTree *DT) {
4293 case CmpInst::ICMP_SLT:
4294 case CmpInst::ICMP_SLE:
4295 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth, AC, CxtI,
4297 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4301 case CmpInst::ICMP_ULT:
4302 case CmpInst::ICMP_ULE:
4303 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth, AC, CxtI,
4305 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth, AC, CxtI, DT))
4311 /// Return true if the operands of the two compares match. IsSwappedOps is true
4312 /// when the operands match, but are swapped.
4313 static bool isMatchingOps(const Value *ALHS, const Value *ARHS,
4314 const Value *BLHS, const Value *BRHS,
4315 bool &IsSwappedOps) {
4317 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
4318 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
4319 return IsMatchingOps || IsSwappedOps;
4322 /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is
4323 /// true. Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS
4324 /// BRHS" is false. Otherwise, return None if we can't infer anything.
4325 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
4328 CmpInst::Predicate BPred,
4331 bool IsSwappedOps) {
4332 // Canonicalize the operands so they're matching.
4334 std::swap(BLHS, BRHS);
4335 BPred = ICmpInst::getSwappedPredicate(BPred);
4337 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
4339 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
4345 /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is
4346 /// true. Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS
4347 /// C2" is false. Otherwise, return None if we can't infer anything.
4348 static Optional<bool>
4349 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, const Value *ALHS,
4350 const ConstantInt *C1,
4351 CmpInst::Predicate BPred,
4352 const Value *BLHS, const ConstantInt *C2) {
4353 assert(ALHS == BLHS && "LHS operands must match.");
4354 ConstantRange DomCR =
4355 ConstantRange::makeExactICmpRegion(APred, C1->getValue());
4357 ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
4358 ConstantRange Intersection = DomCR.intersectWith(CR);
4359 ConstantRange Difference = DomCR.difference(CR);
4360 if (Intersection.isEmptySet())
4362 if (Difference.isEmptySet())
4367 Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
4368 const DataLayout &DL, bool InvertAPred,
4369 unsigned Depth, AssumptionCache *AC,
4370 const Instruction *CxtI,
4371 const DominatorTree *DT) {
4372 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for example.
4373 if (LHS->getType() != RHS->getType())
4376 Type *OpTy = LHS->getType();
4377 assert(OpTy->getScalarType()->isIntegerTy(1));
4379 // LHS ==> RHS by definition
4380 if (!InvertAPred && LHS == RHS)
4383 if (OpTy->isVectorTy())
4384 // TODO: extending the code below to handle vectors
4386 assert(OpTy->isIntegerTy(1) && "implied by above");
4388 ICmpInst::Predicate APred, BPred;
4392 if (!match(LHS, m_ICmp(APred, m_Value(ALHS), m_Value(ARHS))) ||
4393 !match(RHS, m_ICmp(BPred, m_Value(BLHS), m_Value(BRHS))))
4397 APred = CmpInst::getInversePredicate(APred);
4399 // Can we infer anything when the two compares have matching operands?
4401 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) {
4402 if (Optional<bool> Implication = isImpliedCondMatchingOperands(
4403 APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps))
4405 // No amount of additional analysis will infer the second condition, so
4410 // Can we infer anything when the LHS operands match and the RHS operands are
4411 // constants (not necessarily matching)?
4412 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
4413 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
4414 APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS,
4415 cast<ConstantInt>(BRHS)))
4417 // No amount of additional analysis will infer the second condition, so
4423 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth, AC,