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/Analysis/AssumptionCache.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/MemoryBuiltins.h"
20 #include "llvm/IR/CallSite.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/MathExtras.h"
38 using namespace llvm::PatternMatch;
40 const unsigned MaxDepth = 6;
42 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
43 /// 0). For vector types, returns the element type's bitwidth.
44 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
45 if (unsigned BitWidth = Ty->getScalarSizeInBits())
48 return TD ? TD->getPointerTypeSizeInBits(Ty) : 0;
51 // Many of these functions have internal versions that take an assumption
52 // exclusion set. This is because of the potential for mutual recursion to
53 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
54 // classic case of this is assume(x = y), which will attempt to determine
55 // bits in x from bits in y, which will attempt to determine bits in y from
56 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
57 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
58 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
59 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
62 // Simplifying using an assume can only be done in a particular control-flow
63 // context (the context instruction provides that context). If an assume and
64 // the context instruction are not in the same block then the DT helps in
65 // figuring out if we can use it.
69 const Instruction *CxtI;
70 const DominatorTree *DT;
72 Query(AssumptionCache *AC = nullptr, const Instruction *CxtI = nullptr,
73 const DominatorTree *DT = nullptr)
74 : AC(AC), CxtI(CxtI), DT(DT) {}
76 Query(const Query &Q, const Value *NewExcl)
77 : ExclInvs(Q.ExclInvs), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT) {
78 ExclInvs.insert(NewExcl);
81 } // end anonymous namespace
83 // Given the provided Value and, potentially, a context instruction, return
84 // the preferred context instruction (if any).
85 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
86 // If we've been provided with a context instruction, then use that (provided
87 // it has been inserted).
88 if (CxtI && CxtI->getParent())
91 // If the value is really an already-inserted instruction, then use that.
92 CxtI = dyn_cast<Instruction>(V);
93 if (CxtI && CxtI->getParent())
99 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
100 const DataLayout *TD, unsigned Depth,
103 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
104 const DataLayout *TD, unsigned Depth,
105 AssumptionCache *AC, const Instruction *CxtI,
106 const DominatorTree *DT) {
107 ::computeKnownBits(V, KnownZero, KnownOne, TD, Depth,
108 Query(AC, safeCxtI(V, CxtI), DT));
111 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
112 const DataLayout *TD, unsigned Depth,
115 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
116 const DataLayout *TD, unsigned Depth,
117 AssumptionCache *AC, const Instruction *CxtI,
118 const DominatorTree *DT) {
119 ::ComputeSignBit(V, KnownZero, KnownOne, TD, Depth,
120 Query(AC, safeCxtI(V, CxtI), DT));
123 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
126 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
127 AssumptionCache *AC, const Instruction *CxtI,
128 const DominatorTree *DT) {
129 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
130 Query(AC, safeCxtI(V, CxtI), DT));
133 static bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
136 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
137 AssumptionCache *AC, const Instruction *CxtI,
138 const DominatorTree *DT) {
139 return ::isKnownNonZero(V, TD, Depth, Query(AC, safeCxtI(V, CxtI), DT));
142 static bool MaskedValueIsZero(Value *V, const APInt &Mask,
143 const DataLayout *TD, unsigned Depth,
146 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout *TD,
147 unsigned Depth, AssumptionCache *AC,
148 const Instruction *CxtI, const DominatorTree *DT) {
149 return ::MaskedValueIsZero(V, Mask, TD, Depth,
150 Query(AC, safeCxtI(V, CxtI), DT));
153 static unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
154 unsigned Depth, const Query &Q);
156 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
157 unsigned Depth, AssumptionCache *AC,
158 const Instruction *CxtI,
159 const DominatorTree *DT) {
160 return ::ComputeNumSignBits(V, TD, Depth, Query(AC, safeCxtI(V, CxtI), DT));
163 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
164 APInt &KnownZero, APInt &KnownOne,
165 APInt &KnownZero2, APInt &KnownOne2,
166 const DataLayout *TD, unsigned Depth,
169 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
170 // We know that the top bits of C-X are clear if X contains less bits
171 // than C (i.e. no wrap-around can happen). For example, 20-X is
172 // positive if we can prove that X is >= 0 and < 16.
173 if (!CLHS->getValue().isNegative()) {
174 unsigned BitWidth = KnownZero.getBitWidth();
175 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
176 // NLZ can't be BitWidth with no sign bit
177 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
178 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
180 // If all of the MaskV bits are known to be zero, then we know the
181 // output top bits are zero, because we now know that the output is
183 if ((KnownZero2 & MaskV) == MaskV) {
184 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
185 // Top bits known zero.
186 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
192 unsigned BitWidth = KnownZero.getBitWidth();
194 // If an initial sequence of bits in the result is not needed, the
195 // corresponding bits in the operands are not needed.
196 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
197 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1, Q);
198 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
200 // Carry in a 1 for a subtract, rather than a 0.
201 APInt CarryIn(BitWidth, 0);
203 // Sum = LHS + ~RHS + 1
204 std::swap(KnownZero2, KnownOne2);
208 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
209 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
211 // Compute known bits of the carry.
212 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
213 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
215 // Compute set of known bits (where all three relevant bits are known).
216 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
217 APInt RHSKnown = KnownZero2 | KnownOne2;
218 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
219 APInt Known = LHSKnown & RHSKnown & CarryKnown;
221 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
222 "known bits of sum differ");
224 // Compute known bits of the result.
225 KnownZero = ~PossibleSumOne & Known;
226 KnownOne = PossibleSumOne & Known;
228 // Are we still trying to solve for the sign bit?
229 if (!Known.isNegative()) {
231 // Adding two non-negative numbers, or subtracting a negative number from
232 // a non-negative one, can't wrap into negative.
233 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
234 KnownZero |= APInt::getSignBit(BitWidth);
235 // Adding two negative numbers, or subtracting a non-negative number from
236 // a negative one, can't wrap into non-negative.
237 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
238 KnownOne |= APInt::getSignBit(BitWidth);
243 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
244 APInt &KnownZero, APInt &KnownOne,
245 APInt &KnownZero2, APInt &KnownOne2,
246 const DataLayout *TD, unsigned Depth,
248 unsigned BitWidth = KnownZero.getBitWidth();
249 computeKnownBits(Op1, KnownZero, KnownOne, TD, Depth+1, Q);
250 computeKnownBits(Op0, KnownZero2, KnownOne2, TD, Depth+1, Q);
252 bool isKnownNegative = false;
253 bool isKnownNonNegative = false;
254 // If the multiplication is known not to overflow, compute the sign bit.
257 // The product of a number with itself is non-negative.
258 isKnownNonNegative = true;
260 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
261 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
262 bool isKnownNegativeOp1 = KnownOne.isNegative();
263 bool isKnownNegativeOp0 = KnownOne2.isNegative();
264 // The product of two numbers with the same sign is non-negative.
265 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
266 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
267 // The product of a negative number and a non-negative number is either
269 if (!isKnownNonNegative)
270 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
271 isKnownNonZero(Op0, TD, Depth, Q)) ||
272 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
273 isKnownNonZero(Op1, TD, Depth, Q));
277 // If low bits are zero in either operand, output low known-0 bits.
278 // Also compute a conserative estimate for high known-0 bits.
279 // More trickiness is possible, but this is sufficient for the
280 // interesting case of alignment computation.
281 KnownOne.clearAllBits();
282 unsigned TrailZ = KnownZero.countTrailingOnes() +
283 KnownZero2.countTrailingOnes();
284 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
285 KnownZero2.countLeadingOnes(),
286 BitWidth) - BitWidth;
288 TrailZ = std::min(TrailZ, BitWidth);
289 LeadZ = std::min(LeadZ, BitWidth);
290 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
291 APInt::getHighBitsSet(BitWidth, LeadZ);
293 // Only make use of no-wrap flags if we failed to compute the sign bit
294 // directly. This matters if the multiplication always overflows, in
295 // which case we prefer to follow the result of the direct computation,
296 // though as the program is invoking undefined behaviour we can choose
297 // whatever we like here.
298 if (isKnownNonNegative && !KnownOne.isNegative())
299 KnownZero.setBit(BitWidth - 1);
300 else if (isKnownNegative && !KnownZero.isNegative())
301 KnownOne.setBit(BitWidth - 1);
304 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
306 unsigned BitWidth = KnownZero.getBitWidth();
307 unsigned NumRanges = Ranges.getNumOperands() / 2;
308 assert(NumRanges >= 1);
310 // Use the high end of the ranges to find leading zeros.
311 unsigned MinLeadingZeros = BitWidth;
312 for (unsigned i = 0; i < NumRanges; ++i) {
314 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
316 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
317 ConstantRange Range(Lower->getValue(), Upper->getValue());
318 if (Range.isWrappedSet())
319 MinLeadingZeros = 0; // -1 has no zeros
320 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
321 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
324 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
327 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
328 SmallVector<const Value *, 16> WorkSet(1, I);
329 SmallPtrSet<const Value *, 32> Visited;
330 SmallPtrSet<const Value *, 16> EphValues;
332 while (!WorkSet.empty()) {
333 const Value *V = WorkSet.pop_back_val();
334 if (!Visited.insert(V).second)
337 // If all uses of this value are ephemeral, then so is this value.
338 bool FoundNEUse = false;
339 for (const User *I : V->users())
340 if (!EphValues.count(I)) {
350 if (const User *U = dyn_cast<User>(V))
351 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
353 if (isSafeToSpeculativelyExecute(*J))
354 WorkSet.push_back(*J);
362 // Is this an intrinsic that cannot be speculated but also cannot trap?
363 static bool isAssumeLikeIntrinsic(const Instruction *I) {
364 if (const CallInst *CI = dyn_cast<CallInst>(I))
365 if (Function *F = CI->getCalledFunction())
366 switch (F->getIntrinsicID()) {
368 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
369 case Intrinsic::assume:
370 case Intrinsic::dbg_declare:
371 case Intrinsic::dbg_value:
372 case Intrinsic::invariant_start:
373 case Intrinsic::invariant_end:
374 case Intrinsic::lifetime_start:
375 case Intrinsic::lifetime_end:
376 case Intrinsic::objectsize:
377 case Intrinsic::ptr_annotation:
378 case Intrinsic::var_annotation:
385 static bool isValidAssumeForContext(Value *V, const Query &Q,
386 const DataLayout *DL) {
387 Instruction *Inv = cast<Instruction>(V);
389 // There are two restrictions on the use of an assume:
390 // 1. The assume must dominate the context (or the control flow must
391 // reach the assume whenever it reaches the context).
392 // 2. The context must not be in the assume's set of ephemeral values
393 // (otherwise we will use the assume to prove that the condition
394 // feeding the assume is trivially true, thus causing the removal of
398 if (Q.DT->dominates(Inv, Q.CxtI)) {
400 } else if (Inv->getParent() == Q.CxtI->getParent()) {
401 // The context comes first, but they're both in the same block. Make sure
402 // there is nothing in between that might interrupt the control flow.
403 for (BasicBlock::const_iterator I =
404 std::next(BasicBlock::const_iterator(Q.CxtI)),
405 IE(Inv); I != IE; ++I)
406 if (!isSafeToSpeculativelyExecute(I, DL) &&
407 !isAssumeLikeIntrinsic(I))
410 return !isEphemeralValueOf(Inv, Q.CxtI);
416 // When we don't have a DT, we do a limited search...
417 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
419 } else if (Inv->getParent() == Q.CxtI->getParent()) {
420 // Search forward from the assume until we reach the context (or the end
421 // of the block); the common case is that the assume will come first.
422 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
423 IE = Inv->getParent()->end(); I != IE; ++I)
427 // The context must come first...
428 for (BasicBlock::const_iterator I =
429 std::next(BasicBlock::const_iterator(Q.CxtI)),
430 IE(Inv); I != IE; ++I)
431 if (!isSafeToSpeculativelyExecute(I, DL) &&
432 !isAssumeLikeIntrinsic(I))
435 return !isEphemeralValueOf(Inv, Q.CxtI);
441 bool llvm::isValidAssumeForContext(const Instruction *I,
442 const Instruction *CxtI,
443 const DataLayout *DL,
444 const DominatorTree *DT) {
445 return ::isValidAssumeForContext(const_cast<Instruction*>(I),
446 Query(nullptr, CxtI, DT), DL);
449 template<typename LHS, typename RHS>
450 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
451 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
452 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
453 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
456 template<typename LHS, typename RHS>
457 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
458 BinaryOp_match<RHS, LHS, Instruction::And>>
459 m_c_And(const LHS &L, const RHS &R) {
460 return m_CombineOr(m_And(L, R), m_And(R, L));
463 template<typename LHS, typename RHS>
464 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
465 BinaryOp_match<RHS, LHS, Instruction::Or>>
466 m_c_Or(const LHS &L, const RHS &R) {
467 return m_CombineOr(m_Or(L, R), m_Or(R, L));
470 template<typename LHS, typename RHS>
471 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
472 BinaryOp_match<RHS, LHS, Instruction::Xor>>
473 m_c_Xor(const LHS &L, const RHS &R) {
474 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
477 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
479 const DataLayout *DL,
480 unsigned Depth, const Query &Q) {
481 // Use of assumptions is context-sensitive. If we don't have a context, we
483 if (!Q.AC || !Q.CxtI)
486 unsigned BitWidth = KnownZero.getBitWidth();
488 for (auto &AssumeVH : Q.AC->assumptions()) {
491 CallInst *I = cast<CallInst>(AssumeVH);
492 assert((I->getParent()->getParent() ==
493 const_cast<Function*>(Q.CxtI->getParent()->getParent())) &&
494 "Got assumption for the wrong function!");
495 if (Q.ExclInvs.count(I))
498 // Warning: This loop can end up being somewhat performance sensetive.
499 // We're running this loop for once for each value queried resulting in a
500 // runtime of ~O(#assumes * #values).
502 assert(isa<IntrinsicInst>(I) &&
503 dyn_cast<IntrinsicInst>(I)->getIntrinsicID() == Intrinsic::assume &&
504 "must be an assume intrinsic");
506 Value *Arg = I->getArgOperand(0);
509 isValidAssumeForContext(I, Q, DL)) {
510 assert(BitWidth == 1 && "assume operand is not i1?");
511 KnownZero.clearAllBits();
512 KnownOne.setAllBits();
516 // The remaining tests are all recursive, so bail out if we hit the limit.
517 if (Depth == MaxDepth)
521 auto m_V = m_CombineOr(m_Specific(V),
522 m_CombineOr(m_PtrToInt(m_Specific(V)),
523 m_BitCast(m_Specific(V))));
525 CmpInst::Predicate Pred;
528 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
529 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
530 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
531 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
532 KnownZero |= RHSKnownZero;
533 KnownOne |= RHSKnownOne;
535 } else if (match(Arg, m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)),
537 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
538 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
539 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
540 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
541 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
543 // For those bits in the mask that are known to be one, we can propagate
544 // known bits from the RHS to V.
545 KnownZero |= RHSKnownZero & MaskKnownOne;
546 KnownOne |= RHSKnownOne & MaskKnownOne;
547 // assume(~(v & b) = a)
548 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
550 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
551 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
552 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
553 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
554 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
556 // For those bits in the mask that are known to be one, we can propagate
557 // inverted known bits from the RHS to V.
558 KnownZero |= RHSKnownOne & MaskKnownOne;
559 KnownOne |= RHSKnownZero & MaskKnownOne;
561 } else if (match(Arg, m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)),
563 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
564 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
565 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
566 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
567 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
569 // For those bits in B that are known to be zero, we can propagate known
570 // bits from the RHS to V.
571 KnownZero |= RHSKnownZero & BKnownZero;
572 KnownOne |= RHSKnownOne & BKnownZero;
573 // assume(~(v | b) = a)
574 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
576 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
577 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
578 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
579 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
580 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
582 // For those bits in B that are known to be zero, we can propagate
583 // inverted known bits from the RHS to V.
584 KnownZero |= RHSKnownOne & BKnownZero;
585 KnownOne |= RHSKnownZero & BKnownZero;
587 } else if (match(Arg, m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)),
589 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
590 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
591 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
592 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
593 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
595 // For those bits in B that are known to be zero, we can propagate known
596 // bits from the RHS to V. For those bits in B that are known to be one,
597 // we can propagate inverted known bits from the RHS to V.
598 KnownZero |= RHSKnownZero & BKnownZero;
599 KnownOne |= RHSKnownOne & BKnownZero;
600 KnownZero |= RHSKnownOne & BKnownOne;
601 KnownOne |= RHSKnownZero & BKnownOne;
602 // assume(~(v ^ b) = a)
603 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
605 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
606 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
607 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
608 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
609 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
611 // For those bits in B that are known to be zero, we can propagate
612 // inverted known bits from the RHS to V. For those bits in B that are
613 // known to be one, we can propagate known bits from the RHS to V.
614 KnownZero |= RHSKnownOne & BKnownZero;
615 KnownOne |= RHSKnownZero & BKnownZero;
616 KnownZero |= RHSKnownZero & BKnownOne;
617 KnownOne |= RHSKnownOne & BKnownOne;
618 // assume(v << c = a)
619 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
621 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
622 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
623 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
624 // For those bits in RHS that are known, we can propagate them to known
625 // bits in V shifted to the right by C.
626 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
627 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
628 // assume(~(v << c) = a)
629 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
631 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
632 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
633 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
634 // For those bits in RHS that are known, we can propagate them inverted
635 // to known bits in V shifted to the right by C.
636 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
637 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
638 // assume(v >> c = a)
639 } else if (match(Arg,
640 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
644 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
645 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
646 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
647 // For those bits in RHS that are known, we can propagate them to known
648 // bits in V shifted to the right by C.
649 KnownZero |= RHSKnownZero << C->getZExtValue();
650 KnownOne |= RHSKnownOne << C->getZExtValue();
651 // assume(~(v >> c) = a)
652 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
653 m_LShr(m_V, m_ConstantInt(C)),
654 m_AShr(m_V, m_ConstantInt(C)))),
656 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
657 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
658 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
659 // For those bits in RHS that are known, we can propagate them inverted
660 // to known bits in V shifted to the right by C.
661 KnownZero |= RHSKnownOne << C->getZExtValue();
662 KnownOne |= RHSKnownZero << C->getZExtValue();
663 // assume(v >=_s c) where c is non-negative
664 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
665 Pred == ICmpInst::ICMP_SGE &&
666 isValidAssumeForContext(I, Q, DL)) {
667 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
668 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
670 if (RHSKnownZero.isNegative()) {
671 // We know that the sign bit is zero.
672 KnownZero |= APInt::getSignBit(BitWidth);
674 // assume(v >_s c) where c is at least -1.
675 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
676 Pred == ICmpInst::ICMP_SGT &&
677 isValidAssumeForContext(I, Q, DL)) {
678 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
679 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
681 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
682 // We know that the sign bit is zero.
683 KnownZero |= APInt::getSignBit(BitWidth);
685 // assume(v <=_s c) where c is negative
686 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
687 Pred == ICmpInst::ICMP_SLE &&
688 isValidAssumeForContext(I, Q, DL)) {
689 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
690 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
692 if (RHSKnownOne.isNegative()) {
693 // We know that the sign bit is one.
694 KnownOne |= APInt::getSignBit(BitWidth);
696 // assume(v <_s c) where c is non-positive
697 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
698 Pred == ICmpInst::ICMP_SLT &&
699 isValidAssumeForContext(I, Q, DL)) {
700 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
701 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
703 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
704 // We know that the sign bit is one.
705 KnownOne |= APInt::getSignBit(BitWidth);
708 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
709 Pred == ICmpInst::ICMP_ULE &&
710 isValidAssumeForContext(I, Q, DL)) {
711 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
712 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
714 // Whatever high bits in c are zero are known to be zero.
716 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
718 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
719 Pred == ICmpInst::ICMP_ULT &&
720 isValidAssumeForContext(I, Q, DL)) {
721 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
722 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
724 // Whatever high bits in c are zero are known to be zero (if c is a power
725 // of 2, then one more).
726 if (isKnownToBeAPowerOfTwo(A, false, Depth+1, Query(Q, I)))
728 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
731 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
736 /// Determine which bits of V are known to be either zero or one and return
737 /// them in the KnownZero/KnownOne bit sets.
739 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
740 /// we cannot optimize based on the assumption that it is zero without changing
741 /// it to be an explicit zero. If we don't change it to zero, other code could
742 /// optimized based on the contradictory assumption that it is non-zero.
743 /// Because instcombine aggressively folds operations with undef args anyway,
744 /// this won't lose us code quality.
746 /// This function is defined on values with integer type, values with pointer
747 /// type (but only if TD is non-null), and vectors of integers. In the case
748 /// where V is a vector, known zero, and known one values are the
749 /// same width as the vector element, and the bit is set only if it is true
750 /// for all of the elements in the vector.
751 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
752 const DataLayout *TD, unsigned Depth,
754 assert(V && "No Value?");
755 assert(Depth <= MaxDepth && "Limit Search Depth");
756 unsigned BitWidth = KnownZero.getBitWidth();
758 assert((V->getType()->isIntOrIntVectorTy() ||
759 V->getType()->getScalarType()->isPointerTy()) &&
760 "Not integer or pointer type!");
762 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
763 (!V->getType()->isIntOrIntVectorTy() ||
764 V->getType()->getScalarSizeInBits() == BitWidth) &&
765 KnownZero.getBitWidth() == BitWidth &&
766 KnownOne.getBitWidth() == BitWidth &&
767 "V, KnownOne and KnownZero should have same BitWidth");
769 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
770 // We know all of the bits for a constant!
771 KnownOne = CI->getValue();
772 KnownZero = ~KnownOne;
775 // Null and aggregate-zero are all-zeros.
776 if (isa<ConstantPointerNull>(V) ||
777 isa<ConstantAggregateZero>(V)) {
778 KnownOne.clearAllBits();
779 KnownZero = APInt::getAllOnesValue(BitWidth);
782 // Handle a constant vector by taking the intersection of the known bits of
783 // each element. There is no real need to handle ConstantVector here, because
784 // we don't handle undef in any particularly useful way.
785 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
786 // We know that CDS must be a vector of integers. Take the intersection of
788 KnownZero.setAllBits(); KnownOne.setAllBits();
789 APInt Elt(KnownZero.getBitWidth(), 0);
790 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
791 Elt = CDS->getElementAsInteger(i);
798 // The address of an aligned GlobalValue has trailing zeros.
799 if (auto *GO = dyn_cast<GlobalObject>(V)) {
800 unsigned Align = GO->getAlignment();
801 if (Align == 0 && TD) {
802 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
803 Type *ObjectType = GVar->getType()->getElementType();
804 if (ObjectType->isSized()) {
805 // If the object is defined in the current Module, we'll be giving
806 // it the preferred alignment. Otherwise, we have to assume that it
807 // may only have the minimum ABI alignment.
808 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
809 Align = TD->getPreferredAlignment(GVar);
811 Align = TD->getABITypeAlignment(ObjectType);
816 KnownZero = APInt::getLowBitsSet(BitWidth,
817 countTrailingZeros(Align));
819 KnownZero.clearAllBits();
820 KnownOne.clearAllBits();
824 if (Argument *A = dyn_cast<Argument>(V)) {
825 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
827 if (!Align && TD && A->hasStructRetAttr()) {
828 // An sret parameter has at least the ABI alignment of the return type.
829 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
830 if (EltTy->isSized())
831 Align = TD->getABITypeAlignment(EltTy);
835 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
837 KnownZero.clearAllBits();
838 KnownOne.clearAllBits();
840 // Don't give up yet... there might be an assumption that provides more
842 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
846 // Start out not knowing anything.
847 KnownZero.clearAllBits(); KnownOne.clearAllBits();
849 // Limit search depth.
850 // All recursive calls that increase depth must come after this.
851 if (Depth == MaxDepth)
854 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
855 // the bits of its aliasee.
856 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
857 if (!GA->mayBeOverridden())
858 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth + 1, Q);
862 // Check whether a nearby assume intrinsic can determine some known bits.
863 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
865 Operator *I = dyn_cast<Operator>(V);
868 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
869 switch (I->getOpcode()) {
871 case Instruction::Load:
872 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
873 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
875 case Instruction::And: {
876 // If either the LHS or the RHS are Zero, the result is zero.
877 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
878 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
880 // Output known-1 bits are only known if set in both the LHS & RHS.
881 KnownOne &= KnownOne2;
882 // Output known-0 are known to be clear if zero in either the LHS | RHS.
883 KnownZero |= KnownZero2;
886 case Instruction::Or: {
887 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
888 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
890 // Output known-0 bits are only known if clear in both the LHS & RHS.
891 KnownZero &= KnownZero2;
892 // Output known-1 are known to be set if set in either the LHS | RHS.
893 KnownOne |= KnownOne2;
896 case Instruction::Xor: {
897 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
898 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
900 // Output known-0 bits are known if clear or set in both the LHS & RHS.
901 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
902 // Output known-1 are known to be set if set in only one of the LHS, RHS.
903 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
904 KnownZero = KnownZeroOut;
907 case Instruction::Mul: {
908 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
909 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
910 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
914 case Instruction::UDiv: {
915 // For the purposes of computing leading zeros we can conservatively
916 // treat a udiv as a logical right shift by the power of 2 known to
917 // be less than the denominator.
918 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
919 unsigned LeadZ = KnownZero2.countLeadingOnes();
921 KnownOne2.clearAllBits();
922 KnownZero2.clearAllBits();
923 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
924 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
925 if (RHSUnknownLeadingOnes != BitWidth)
926 LeadZ = std::min(BitWidth,
927 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
929 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
932 case Instruction::Select:
933 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1, Q);
934 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
936 // Only known if known in both the LHS and RHS.
937 KnownOne &= KnownOne2;
938 KnownZero &= KnownZero2;
940 case Instruction::FPTrunc:
941 case Instruction::FPExt:
942 case Instruction::FPToUI:
943 case Instruction::FPToSI:
944 case Instruction::SIToFP:
945 case Instruction::UIToFP:
946 break; // Can't work with floating point.
947 case Instruction::PtrToInt:
948 case Instruction::IntToPtr:
949 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
950 // We can't handle these if we don't know the pointer size.
952 // FALL THROUGH and handle them the same as zext/trunc.
953 case Instruction::ZExt:
954 case Instruction::Trunc: {
955 Type *SrcTy = I->getOperand(0)->getType();
957 unsigned SrcBitWidth;
958 // Note that we handle pointer operands here because of inttoptr/ptrtoint
959 // which fall through here.
961 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
963 SrcBitWidth = SrcTy->getScalarSizeInBits();
964 if (!SrcBitWidth) break;
967 assert(SrcBitWidth && "SrcBitWidth can't be zero");
968 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
969 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
970 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
971 KnownZero = KnownZero.zextOrTrunc(BitWidth);
972 KnownOne = KnownOne.zextOrTrunc(BitWidth);
973 // Any top bits are known to be zero.
974 if (BitWidth > SrcBitWidth)
975 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
978 case Instruction::BitCast: {
979 Type *SrcTy = I->getOperand(0)->getType();
980 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
981 // TODO: For now, not handling conversions like:
982 // (bitcast i64 %x to <2 x i32>)
983 !I->getType()->isVectorTy()) {
984 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
989 case Instruction::SExt: {
990 // Compute the bits in the result that are not present in the input.
991 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
993 KnownZero = KnownZero.trunc(SrcBitWidth);
994 KnownOne = KnownOne.trunc(SrcBitWidth);
995 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
996 KnownZero = KnownZero.zext(BitWidth);
997 KnownOne = KnownOne.zext(BitWidth);
999 // If the sign bit of the input is known set or clear, then we know the
1000 // top bits of the result.
1001 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1002 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1003 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1004 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1007 case Instruction::Shl:
1008 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1009 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1010 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1011 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1012 KnownZero <<= ShiftAmt;
1013 KnownOne <<= ShiftAmt;
1014 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1017 case Instruction::LShr:
1018 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1019 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1020 // Compute the new bits that are at the top now.
1021 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1023 // Unsigned shift right.
1024 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1025 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1026 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1027 // high bits known zero.
1028 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1031 case Instruction::AShr:
1032 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1033 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1034 // Compute the new bits that are at the top now.
1035 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1037 // Signed shift right.
1038 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1039 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1040 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1042 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1043 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1044 KnownZero |= HighBits;
1045 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1046 KnownOne |= HighBits;
1049 case Instruction::Sub: {
1050 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1051 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1052 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1056 case Instruction::Add: {
1057 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1058 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1059 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1063 case Instruction::SRem:
1064 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1065 APInt RA = Rem->getValue().abs();
1066 if (RA.isPowerOf2()) {
1067 APInt LowBits = RA - 1;
1068 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD,
1071 // The low bits of the first operand are unchanged by the srem.
1072 KnownZero = KnownZero2 & LowBits;
1073 KnownOne = KnownOne2 & LowBits;
1075 // If the first operand is non-negative or has all low bits zero, then
1076 // the upper bits are all zero.
1077 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1078 KnownZero |= ~LowBits;
1080 // If the first operand is negative and not all low bits are zero, then
1081 // the upper bits are all one.
1082 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1083 KnownOne |= ~LowBits;
1085 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1089 // The sign bit is the LHS's sign bit, except when the result of the
1090 // remainder is zero.
1091 if (KnownZero.isNonNegative()) {
1092 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1093 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
1095 // If it's known zero, our sign bit is also zero.
1096 if (LHSKnownZero.isNegative())
1097 KnownZero.setBit(BitWidth - 1);
1101 case Instruction::URem: {
1102 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1103 APInt RA = Rem->getValue();
1104 if (RA.isPowerOf2()) {
1105 APInt LowBits = (RA - 1);
1106 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
1108 KnownZero |= ~LowBits;
1109 KnownOne &= LowBits;
1114 // Since the result is less than or equal to either operand, any leading
1115 // zero bits in either operand must also exist in the result.
1116 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1117 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
1119 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1120 KnownZero2.countLeadingOnes());
1121 KnownOne.clearAllBits();
1122 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1126 case Instruction::Alloca: {
1127 AllocaInst *AI = cast<AllocaInst>(V);
1128 unsigned Align = AI->getAlignment();
1129 if (Align == 0 && TD)
1130 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1133 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1136 case Instruction::GetElementPtr: {
1137 // Analyze all of the subscripts of this getelementptr instruction
1138 // to determine if we can prove known low zero bits.
1139 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1140 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
1142 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1144 gep_type_iterator GTI = gep_type_begin(I);
1145 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1146 Value *Index = I->getOperand(i);
1147 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1148 // Handle struct member offset arithmetic.
1154 // Handle case when index is vector zeroinitializer
1155 Constant *CIndex = cast<Constant>(Index);
1156 if (CIndex->isZeroValue())
1159 if (CIndex->getType()->isVectorTy())
1160 Index = CIndex->getSplatValue();
1162 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1163 const StructLayout *SL = TD->getStructLayout(STy);
1164 uint64_t Offset = SL->getElementOffset(Idx);
1165 TrailZ = std::min<unsigned>(TrailZ,
1166 countTrailingZeros(Offset));
1168 // Handle array index arithmetic.
1169 Type *IndexedTy = GTI.getIndexedType();
1170 if (!IndexedTy->isSized()) {
1174 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1175 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
1176 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1177 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1, Q);
1178 TrailZ = std::min(TrailZ,
1179 unsigned(countTrailingZeros(TypeSize) +
1180 LocalKnownZero.countTrailingOnes()));
1184 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1187 case Instruction::PHI: {
1188 PHINode *P = cast<PHINode>(I);
1189 // Handle the case of a simple two-predecessor recurrence PHI.
1190 // There's a lot more that could theoretically be done here, but
1191 // this is sufficient to catch some interesting cases.
1192 if (P->getNumIncomingValues() == 2) {
1193 for (unsigned i = 0; i != 2; ++i) {
1194 Value *L = P->getIncomingValue(i);
1195 Value *R = P->getIncomingValue(!i);
1196 Operator *LU = dyn_cast<Operator>(L);
1199 unsigned Opcode = LU->getOpcode();
1200 // Check for operations that have the property that if
1201 // both their operands have low zero bits, the result
1202 // will have low zero bits.
1203 if (Opcode == Instruction::Add ||
1204 Opcode == Instruction::Sub ||
1205 Opcode == Instruction::And ||
1206 Opcode == Instruction::Or ||
1207 Opcode == Instruction::Mul) {
1208 Value *LL = LU->getOperand(0);
1209 Value *LR = LU->getOperand(1);
1210 // Find a recurrence.
1217 // Ok, we have a PHI of the form L op= R. Check for low
1219 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1, Q);
1221 // We need to take the minimum number of known bits
1222 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1223 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1, Q);
1225 KnownZero = APInt::getLowBitsSet(BitWidth,
1226 std::min(KnownZero2.countTrailingOnes(),
1227 KnownZero3.countTrailingOnes()));
1233 // Unreachable blocks may have zero-operand PHI nodes.
1234 if (P->getNumIncomingValues() == 0)
1237 // Otherwise take the unions of the known bit sets of the operands,
1238 // taking conservative care to avoid excessive recursion.
1239 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1240 // Skip if every incoming value references to ourself.
1241 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1244 KnownZero = APInt::getAllOnesValue(BitWidth);
1245 KnownOne = APInt::getAllOnesValue(BitWidth);
1246 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1247 // Skip direct self references.
1248 if (P->getIncomingValue(i) == P) continue;
1250 KnownZero2 = APInt(BitWidth, 0);
1251 KnownOne2 = APInt(BitWidth, 0);
1252 // Recurse, but cap the recursion to one level, because we don't
1253 // want to waste time spinning around in loops.
1254 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
1256 KnownZero &= KnownZero2;
1257 KnownOne &= KnownOne2;
1258 // If all bits have been ruled out, there's no need to check
1260 if (!KnownZero && !KnownOne)
1266 case Instruction::Call:
1267 case Instruction::Invoke:
1268 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1269 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1270 // If a range metadata is attached to this IntrinsicInst, intersect the
1271 // explicit range specified by the metadata and the implicit range of
1273 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1274 switch (II->getIntrinsicID()) {
1276 case Intrinsic::ctlz:
1277 case Intrinsic::cttz: {
1278 unsigned LowBits = Log2_32(BitWidth)+1;
1279 // If this call is undefined for 0, the result will be less than 2^n.
1280 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1282 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1285 case Intrinsic::ctpop: {
1286 unsigned LowBits = Log2_32(BitWidth)+1;
1287 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1290 case Intrinsic::x86_sse42_crc32_64_64:
1291 KnownZero |= APInt::getHighBitsSet(64, 32);
1296 case Instruction::ExtractValue:
1297 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1298 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1299 if (EVI->getNumIndices() != 1) break;
1300 if (EVI->getIndices()[0] == 0) {
1301 switch (II->getIntrinsicID()) {
1303 case Intrinsic::uadd_with_overflow:
1304 case Intrinsic::sadd_with_overflow:
1305 computeKnownBitsAddSub(true, II->getArgOperand(0),
1306 II->getArgOperand(1), false, KnownZero,
1307 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1309 case Intrinsic::usub_with_overflow:
1310 case Intrinsic::ssub_with_overflow:
1311 computeKnownBitsAddSub(false, II->getArgOperand(0),
1312 II->getArgOperand(1), false, KnownZero,
1313 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1315 case Intrinsic::umul_with_overflow:
1316 case Intrinsic::smul_with_overflow:
1317 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
1318 false, KnownZero, KnownOne,
1319 KnownZero2, KnownOne2, TD, Depth, Q);
1326 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1329 /// Determine whether the sign bit is known to be zero or one.
1330 /// Convenience wrapper around computeKnownBits.
1331 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1332 const DataLayout *TD, unsigned Depth,
1334 unsigned BitWidth = getBitWidth(V->getType(), TD);
1340 APInt ZeroBits(BitWidth, 0);
1341 APInt OneBits(BitWidth, 0);
1342 computeKnownBits(V, ZeroBits, OneBits, TD, Depth, Q);
1343 KnownOne = OneBits[BitWidth - 1];
1344 KnownZero = ZeroBits[BitWidth - 1];
1347 /// Return true if the given value is known to have exactly one
1348 /// bit set when defined. For vectors return true if every element is known to
1349 /// be a power of two when defined. Supports values with integer or pointer
1350 /// types and vectors of integers.
1351 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1353 if (Constant *C = dyn_cast<Constant>(V)) {
1354 if (C->isNullValue())
1356 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1357 return CI->getValue().isPowerOf2();
1358 // TODO: Handle vector constants.
1361 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1362 // it is shifted off the end then the result is undefined.
1363 if (match(V, m_Shl(m_One(), m_Value())))
1366 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1367 // bottom. If it is shifted off the bottom then the result is undefined.
1368 if (match(V, m_LShr(m_SignBit(), m_Value())))
1371 // The remaining tests are all recursive, so bail out if we hit the limit.
1372 if (Depth++ == MaxDepth)
1375 Value *X = nullptr, *Y = nullptr;
1376 // A shift of a power of two is a power of two or zero.
1377 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1378 match(V, m_Shr(m_Value(X), m_Value()))))
1379 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q);
1381 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1382 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1384 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1386 isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1387 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1389 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1390 // A power of two and'd with anything is a power of two or zero.
1391 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q) ||
1392 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth, Q))
1394 // X & (-X) is always a power of two or zero.
1395 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1400 // Adding a power-of-two or zero to the same power-of-two or zero yields
1401 // either the original power-of-two, a larger power-of-two or zero.
1402 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1403 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1404 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1405 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1406 match(X, m_And(m_Value(), m_Specific(Y))))
1407 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1409 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1410 match(Y, m_And(m_Value(), m_Specific(X))))
1411 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1414 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1415 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1416 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth, Q);
1418 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1419 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth, Q);
1420 // If i8 V is a power of two or zero:
1421 // ZeroBits: 1 1 1 0 1 1 1 1
1422 // ~ZeroBits: 0 0 0 1 0 0 0 0
1423 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1424 // If OrZero isn't set, we cannot give back a zero result.
1425 // Make sure either the LHS or RHS has a bit set.
1426 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1431 // An exact divide or right shift can only shift off zero bits, so the result
1432 // is a power of two only if the first operand is a power of two and not
1433 // copying a sign bit (sdiv int_min, 2).
1434 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1435 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1436 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1443 /// \brief Test whether a GEP's result is known to be non-null.
1445 /// Uses properties inherent in a GEP to try to determine whether it is known
1448 /// Currently this routine does not support vector GEPs.
1449 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
1450 unsigned Depth, const Query &Q) {
1451 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1454 // FIXME: Support vector-GEPs.
1455 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1457 // If the base pointer is non-null, we cannot walk to a null address with an
1458 // inbounds GEP in address space zero.
1459 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1462 // Past this, if we don't have DataLayout, we can't do much.
1466 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1467 // If so, then the GEP cannot produce a null pointer, as doing so would
1468 // inherently violate the inbounds contract within address space zero.
1469 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1470 GTI != GTE; ++GTI) {
1471 // Struct types are easy -- they must always be indexed by a constant.
1472 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1473 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1474 unsigned ElementIdx = OpC->getZExtValue();
1475 const StructLayout *SL = DL->getStructLayout(STy);
1476 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1477 if (ElementOffset > 0)
1482 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1483 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
1486 // Fast path the constant operand case both for efficiency and so we don't
1487 // increment Depth when just zipping down an all-constant GEP.
1488 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1494 // We post-increment Depth here because while isKnownNonZero increments it
1495 // as well, when we pop back up that increment won't persist. We don't want
1496 // to recurse 10k times just because we have 10k GEP operands. We don't
1497 // bail completely out because we want to handle constant GEPs regardless
1499 if (Depth++ >= MaxDepth)
1502 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1509 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1510 /// ensure that the value it's attached to is never Value? 'RangeType' is
1511 /// is the type of the value described by the range.
1512 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1513 const APInt& Value) {
1514 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1515 assert(NumRanges >= 1);
1516 for (unsigned i = 0; i < NumRanges; ++i) {
1517 ConstantInt *Lower =
1518 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1519 ConstantInt *Upper =
1520 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1521 ConstantRange Range(Lower->getValue(), Upper->getValue());
1522 if (Range.contains(Value))
1528 /// Return true if the given value is known to be non-zero when defined.
1529 /// For vectors return true if every element is known to be non-zero when
1530 /// defined. Supports values with integer or pointer type and vectors of
1532 bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
1534 if (Constant *C = dyn_cast<Constant>(V)) {
1535 if (C->isNullValue())
1537 if (isa<ConstantInt>(C))
1538 // Must be non-zero due to null test above.
1540 // TODO: Handle vectors
1544 if (Instruction* I = dyn_cast<Instruction>(V)) {
1545 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1546 // If the possible ranges don't contain zero, then the value is
1547 // definitely non-zero.
1548 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1549 const APInt ZeroValue(Ty->getBitWidth(), 0);
1550 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1556 // The remaining tests are all recursive, so bail out if we hit the limit.
1557 if (Depth++ >= MaxDepth)
1560 // Check for pointer simplifications.
1561 if (V->getType()->isPointerTy()) {
1562 if (isKnownNonNull(V))
1564 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1565 if (isGEPKnownNonNull(GEP, TD, Depth, Q))
1569 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1571 // X | Y != 0 if X != 0 or Y != 0.
1572 Value *X = nullptr, *Y = nullptr;
1573 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1574 return isKnownNonZero(X, TD, Depth, Q) ||
1575 isKnownNonZero(Y, TD, Depth, Q);
1577 // ext X != 0 if X != 0.
1578 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1579 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth, Q);
1581 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1582 // if the lowest bit is shifted off the end.
1583 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1584 // shl nuw can't remove any non-zero bits.
1585 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1586 if (BO->hasNoUnsignedWrap())
1587 return isKnownNonZero(X, TD, Depth, Q);
1589 APInt KnownZero(BitWidth, 0);
1590 APInt KnownOne(BitWidth, 0);
1591 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1595 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1596 // defined if the sign bit is shifted off the end.
1597 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1598 // shr exact can only shift out zero bits.
1599 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1601 return isKnownNonZero(X, TD, Depth, Q);
1603 bool XKnownNonNegative, XKnownNegative;
1604 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1608 // div exact can only produce a zero if the dividend is zero.
1609 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1610 return isKnownNonZero(X, TD, Depth, Q);
1613 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1614 bool XKnownNonNegative, XKnownNegative;
1615 bool YKnownNonNegative, YKnownNegative;
1616 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1617 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth, Q);
1619 // If X and Y are both non-negative (as signed values) then their sum is not
1620 // zero unless both X and Y are zero.
1621 if (XKnownNonNegative && YKnownNonNegative)
1622 if (isKnownNonZero(X, TD, Depth, Q) ||
1623 isKnownNonZero(Y, TD, Depth, Q))
1626 // If X and Y are both negative (as signed values) then their sum is not
1627 // zero unless both X and Y equal INT_MIN.
1628 if (BitWidth && XKnownNegative && YKnownNegative) {
1629 APInt KnownZero(BitWidth, 0);
1630 APInt KnownOne(BitWidth, 0);
1631 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1632 // The sign bit of X is set. If some other bit is set then X is not equal
1634 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1635 if ((KnownOne & Mask) != 0)
1637 // The sign bit of Y is set. If some other bit is set then Y is not equal
1639 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth, Q);
1640 if ((KnownOne & Mask) != 0)
1644 // The sum of a non-negative number and a power of two is not zero.
1645 if (XKnownNonNegative &&
1646 isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth, Q))
1648 if (YKnownNonNegative &&
1649 isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth, Q))
1653 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1654 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1655 // If X and Y are non-zero then so is X * Y as long as the multiplication
1656 // does not overflow.
1657 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1658 isKnownNonZero(X, TD, Depth, Q) &&
1659 isKnownNonZero(Y, TD, Depth, Q))
1662 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1663 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1664 if (isKnownNonZero(SI->getTrueValue(), TD, Depth, Q) &&
1665 isKnownNonZero(SI->getFalseValue(), TD, Depth, Q))
1669 if (!BitWidth) return false;
1670 APInt KnownZero(BitWidth, 0);
1671 APInt KnownOne(BitWidth, 0);
1672 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1673 return KnownOne != 0;
1676 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1677 /// simplify operations downstream. Mask is known to be zero for bits that V
1680 /// This function is defined on values with integer type, values with pointer
1681 /// type (but only if TD is non-null), and vectors of integers. In the case
1682 /// where V is a vector, the mask, known zero, and known one values are the
1683 /// same width as the vector element, and the bit is set only if it is true
1684 /// for all of the elements in the vector.
1685 bool MaskedValueIsZero(Value *V, const APInt &Mask,
1686 const DataLayout *TD, unsigned Depth,
1688 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1689 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1690 return (KnownZero & Mask) == Mask;
1695 /// Return the number of times the sign bit of the register is replicated into
1696 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1697 /// (itself), but other cases can give us information. For example, immediately
1698 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1699 /// other, so we return 3.
1701 /// 'Op' must have a scalar integer type.
1703 unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
1704 unsigned Depth, const Query &Q) {
1705 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1706 "ComputeNumSignBits requires a DataLayout object to operate "
1707 "on non-integer values!");
1708 Type *Ty = V->getType();
1709 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1710 Ty->getScalarSizeInBits();
1712 unsigned FirstAnswer = 1;
1714 // Note that ConstantInt is handled by the general computeKnownBits case
1718 return 1; // Limit search depth.
1720 Operator *U = dyn_cast<Operator>(V);
1721 switch (Operator::getOpcode(V)) {
1723 case Instruction::SExt:
1724 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1725 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q) + Tmp;
1727 case Instruction::AShr: {
1728 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1729 // ashr X, C -> adds C sign bits. Vectors too.
1731 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1732 Tmp += ShAmt->getZExtValue();
1733 if (Tmp > TyBits) Tmp = TyBits;
1737 case Instruction::Shl: {
1739 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1740 // shl destroys sign bits.
1741 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1742 Tmp2 = ShAmt->getZExtValue();
1743 if (Tmp2 >= TyBits || // Bad shift.
1744 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1749 case Instruction::And:
1750 case Instruction::Or:
1751 case Instruction::Xor: // NOT is handled here.
1752 // Logical binary ops preserve the number of sign bits at the worst.
1753 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1755 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1756 FirstAnswer = std::min(Tmp, Tmp2);
1757 // We computed what we know about the sign bits as our first
1758 // answer. Now proceed to the generic code that uses
1759 // computeKnownBits, and pick whichever answer is better.
1763 case Instruction::Select:
1764 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1765 if (Tmp == 1) return 1; // Early out.
1766 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1, Q);
1767 return std::min(Tmp, Tmp2);
1769 case Instruction::Add:
1770 // Add can have at most one carry bit. Thus we know that the output
1771 // is, at worst, one more bit than the inputs.
1772 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1773 if (Tmp == 1) return 1; // Early out.
1775 // Special case decrementing a value (ADD X, -1):
1776 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
1777 if (CRHS->isAllOnesValue()) {
1778 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1779 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1781 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1783 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1786 // If we are subtracting one from a positive number, there is no carry
1787 // out of the result.
1788 if (KnownZero.isNegative())
1792 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1793 if (Tmp2 == 1) return 1;
1794 return std::min(Tmp, Tmp2)-1;
1796 case Instruction::Sub:
1797 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1798 if (Tmp2 == 1) return 1;
1801 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
1802 if (CLHS->isNullValue()) {
1803 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1804 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
1805 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1807 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1810 // If the input is known to be positive (the sign bit is known clear),
1811 // the output of the NEG has the same number of sign bits as the input.
1812 if (KnownZero.isNegative())
1815 // Otherwise, we treat this like a SUB.
1818 // Sub can have at most one carry bit. Thus we know that the output
1819 // is, at worst, one more bit than the inputs.
1820 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1821 if (Tmp == 1) return 1; // Early out.
1822 return std::min(Tmp, Tmp2)-1;
1824 case Instruction::PHI: {
1825 PHINode *PN = cast<PHINode>(U);
1826 unsigned NumIncomingValues = PN->getNumIncomingValues();
1827 // Don't analyze large in-degree PHIs.
1828 if (NumIncomingValues > 4) break;
1829 // Unreachable blocks may have zero-operand PHI nodes.
1830 if (NumIncomingValues == 0) break;
1832 // Take the minimum of all incoming values. This can't infinitely loop
1833 // because of our depth threshold.
1834 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1, Q);
1835 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
1836 if (Tmp == 1) return Tmp;
1838 ComputeNumSignBits(PN->getIncomingValue(i), TD,
1844 case Instruction::Trunc:
1845 // FIXME: it's tricky to do anything useful for this, but it is an important
1846 // case for targets like X86.
1850 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1851 // use this information.
1852 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1854 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1856 if (KnownZero.isNegative()) { // sign bit is 0
1858 } else if (KnownOne.isNegative()) { // sign bit is 1;
1865 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1866 // the number of identical bits in the top of the input value.
1868 Mask <<= Mask.getBitWidth()-TyBits;
1869 // Return # leading zeros. We use 'min' here in case Val was zero before
1870 // shifting. We don't want to return '64' as for an i32 "0".
1871 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1874 /// This function computes the integer multiple of Base that equals V.
1875 /// If successful, it returns true and returns the multiple in
1876 /// Multiple. If unsuccessful, it returns false. It looks
1877 /// through SExt instructions only if LookThroughSExt is true.
1878 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1879 bool LookThroughSExt, unsigned Depth) {
1880 const unsigned MaxDepth = 6;
1882 assert(V && "No Value?");
1883 assert(Depth <= MaxDepth && "Limit Search Depth");
1884 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1886 Type *T = V->getType();
1888 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1898 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1899 Constant *BaseVal = ConstantInt::get(T, Base);
1900 if (CO && CO == BaseVal) {
1902 Multiple = ConstantInt::get(T, 1);
1906 if (CI && CI->getZExtValue() % Base == 0) {
1907 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1911 if (Depth == MaxDepth) return false; // Limit search depth.
1913 Operator *I = dyn_cast<Operator>(V);
1914 if (!I) return false;
1916 switch (I->getOpcode()) {
1918 case Instruction::SExt:
1919 if (!LookThroughSExt) return false;
1920 // otherwise fall through to ZExt
1921 case Instruction::ZExt:
1922 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1923 LookThroughSExt, Depth+1);
1924 case Instruction::Shl:
1925 case Instruction::Mul: {
1926 Value *Op0 = I->getOperand(0);
1927 Value *Op1 = I->getOperand(1);
1929 if (I->getOpcode() == Instruction::Shl) {
1930 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1931 if (!Op1CI) return false;
1932 // Turn Op0 << Op1 into Op0 * 2^Op1
1933 APInt Op1Int = Op1CI->getValue();
1934 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1935 APInt API(Op1Int.getBitWidth(), 0);
1936 API.setBit(BitToSet);
1937 Op1 = ConstantInt::get(V->getContext(), API);
1940 Value *Mul0 = nullptr;
1941 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1942 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1943 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1944 if (Op1C->getType()->getPrimitiveSizeInBits() <
1945 MulC->getType()->getPrimitiveSizeInBits())
1946 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1947 if (Op1C->getType()->getPrimitiveSizeInBits() >
1948 MulC->getType()->getPrimitiveSizeInBits())
1949 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1951 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1952 Multiple = ConstantExpr::getMul(MulC, Op1C);
1956 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1957 if (Mul0CI->getValue() == 1) {
1958 // V == Base * Op1, so return Op1
1964 Value *Mul1 = nullptr;
1965 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1966 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1967 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1968 if (Op0C->getType()->getPrimitiveSizeInBits() <
1969 MulC->getType()->getPrimitiveSizeInBits())
1970 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1971 if (Op0C->getType()->getPrimitiveSizeInBits() >
1972 MulC->getType()->getPrimitiveSizeInBits())
1973 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1975 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1976 Multiple = ConstantExpr::getMul(MulC, Op0C);
1980 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1981 if (Mul1CI->getValue() == 1) {
1982 // V == Base * Op0, so return Op0
1990 // We could not determine if V is a multiple of Base.
1994 /// Return true if we can prove that the specified FP value is never equal to
1997 /// NOTE: this function will need to be revisited when we support non-default
2000 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2001 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2002 return !CFP->getValueAPF().isNegZero();
2005 return 1; // Limit search depth.
2007 const Operator *I = dyn_cast<Operator>(V);
2008 if (!I) return false;
2010 // Check if the nsz fast-math flag is set
2011 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2012 if (FPO->hasNoSignedZeros())
2015 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2016 if (I->getOpcode() == Instruction::FAdd)
2017 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2018 if (CFP->isNullValue())
2021 // sitofp and uitofp turn into +0.0 for zero.
2022 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2025 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2026 // sqrt(-0.0) = -0.0, no other negative results are possible.
2027 if (II->getIntrinsicID() == Intrinsic::sqrt)
2028 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2030 if (const CallInst *CI = dyn_cast<CallInst>(I))
2031 if (const Function *F = CI->getCalledFunction()) {
2032 if (F->isDeclaration()) {
2034 if (F->getName() == "abs") return true;
2035 // fabs[lf](x) != -0.0
2036 if (F->getName() == "fabs") return true;
2037 if (F->getName() == "fabsf") return true;
2038 if (F->getName() == "fabsl") return true;
2039 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2040 F->getName() == "sqrtl")
2041 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2048 /// If the specified value can be set by repeating the same byte in memory,
2049 /// return the i8 value that it is represented with. This is
2050 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2051 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2052 /// byte store (e.g. i16 0x1234), return null.
2053 Value *llvm::isBytewiseValue(Value *V) {
2054 // All byte-wide stores are splatable, even of arbitrary variables.
2055 if (V->getType()->isIntegerTy(8)) return V;
2057 // Handle 'null' ConstantArrayZero etc.
2058 if (Constant *C = dyn_cast<Constant>(V))
2059 if (C->isNullValue())
2060 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2062 // Constant float and double values can be handled as integer values if the
2063 // corresponding integer value is "byteable". An important case is 0.0.
2064 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2065 if (CFP->getType()->isFloatTy())
2066 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2067 if (CFP->getType()->isDoubleTy())
2068 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2069 // Don't handle long double formats, which have strange constraints.
2072 // We can handle constant integers that are power of two in size and a
2073 // multiple of 8 bits.
2074 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2075 unsigned Width = CI->getBitWidth();
2076 if (isPowerOf2_32(Width) && Width > 8) {
2077 // We can handle this value if the recursive binary decomposition is the
2078 // same at all levels.
2079 APInt Val = CI->getValue();
2081 while (Val.getBitWidth() != 8) {
2082 unsigned NextWidth = Val.getBitWidth()/2;
2083 Val2 = Val.lshr(NextWidth);
2084 Val2 = Val2.trunc(Val.getBitWidth()/2);
2085 Val = Val.trunc(Val.getBitWidth()/2);
2087 // If the top/bottom halves aren't the same, reject it.
2091 return ConstantInt::get(V->getContext(), Val);
2095 // A ConstantDataArray/Vector is splatable if all its members are equal and
2097 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2098 Value *Elt = CA->getElementAsConstant(0);
2099 Value *Val = isBytewiseValue(Elt);
2103 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2104 if (CA->getElementAsConstant(I) != Elt)
2110 // Conceptually, we could handle things like:
2111 // %a = zext i8 %X to i16
2112 // %b = shl i16 %a, 8
2113 // %c = or i16 %a, %b
2114 // but until there is an example that actually needs this, it doesn't seem
2115 // worth worrying about.
2120 // This is the recursive version of BuildSubAggregate. It takes a few different
2121 // arguments. Idxs is the index within the nested struct From that we are
2122 // looking at now (which is of type IndexedType). IdxSkip is the number of
2123 // indices from Idxs that should be left out when inserting into the resulting
2124 // struct. To is the result struct built so far, new insertvalue instructions
2126 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2127 SmallVectorImpl<unsigned> &Idxs,
2129 Instruction *InsertBefore) {
2130 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2132 // Save the original To argument so we can modify it
2134 // General case, the type indexed by Idxs is a struct
2135 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2136 // Process each struct element recursively
2139 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2143 // Couldn't find any inserted value for this index? Cleanup
2144 while (PrevTo != OrigTo) {
2145 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2146 PrevTo = Del->getAggregateOperand();
2147 Del->eraseFromParent();
2149 // Stop processing elements
2153 // If we successfully found a value for each of our subaggregates
2157 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2158 // the struct's elements had a value that was inserted directly. In the latter
2159 // case, perhaps we can't determine each of the subelements individually, but
2160 // we might be able to find the complete struct somewhere.
2162 // Find the value that is at that particular spot
2163 Value *V = FindInsertedValue(From, Idxs);
2168 // Insert the value in the new (sub) aggregrate
2169 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2170 "tmp", InsertBefore);
2173 // This helper takes a nested struct and extracts a part of it (which is again a
2174 // struct) into a new value. For example, given the struct:
2175 // { a, { b, { c, d }, e } }
2176 // and the indices "1, 1" this returns
2179 // It does this by inserting an insertvalue for each element in the resulting
2180 // struct, as opposed to just inserting a single struct. This will only work if
2181 // each of the elements of the substruct are known (ie, inserted into From by an
2182 // insertvalue instruction somewhere).
2184 // All inserted insertvalue instructions are inserted before InsertBefore
2185 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2186 Instruction *InsertBefore) {
2187 assert(InsertBefore && "Must have someplace to insert!");
2188 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2190 Value *To = UndefValue::get(IndexedType);
2191 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2192 unsigned IdxSkip = Idxs.size();
2194 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2197 /// Given an aggregrate and an sequence of indices, see if
2198 /// the scalar value indexed is already around as a register, for example if it
2199 /// were inserted directly into the aggregrate.
2201 /// If InsertBefore is not null, this function will duplicate (modified)
2202 /// insertvalues when a part of a nested struct is extracted.
2203 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2204 Instruction *InsertBefore) {
2205 // Nothing to index? Just return V then (this is useful at the end of our
2207 if (idx_range.empty())
2209 // We have indices, so V should have an indexable type.
2210 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2211 "Not looking at a struct or array?");
2212 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2213 "Invalid indices for type?");
2215 if (Constant *C = dyn_cast<Constant>(V)) {
2216 C = C->getAggregateElement(idx_range[0]);
2217 if (!C) return nullptr;
2218 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2221 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2222 // Loop the indices for the insertvalue instruction in parallel with the
2223 // requested indices
2224 const unsigned *req_idx = idx_range.begin();
2225 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2226 i != e; ++i, ++req_idx) {
2227 if (req_idx == idx_range.end()) {
2228 // We can't handle this without inserting insertvalues
2232 // The requested index identifies a part of a nested aggregate. Handle
2233 // this specially. For example,
2234 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2235 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2236 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2237 // This can be changed into
2238 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2239 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2240 // which allows the unused 0,0 element from the nested struct to be
2242 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2246 // This insert value inserts something else than what we are looking for.
2247 // See if the (aggregrate) value inserted into has the value we are
2248 // looking for, then.
2250 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2253 // If we end up here, the indices of the insertvalue match with those
2254 // requested (though possibly only partially). Now we recursively look at
2255 // the inserted value, passing any remaining indices.
2256 return FindInsertedValue(I->getInsertedValueOperand(),
2257 makeArrayRef(req_idx, idx_range.end()),
2261 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2262 // If we're extracting a value from an aggregrate that was extracted from
2263 // something else, we can extract from that something else directly instead.
2264 // However, we will need to chain I's indices with the requested indices.
2266 // Calculate the number of indices required
2267 unsigned size = I->getNumIndices() + idx_range.size();
2268 // Allocate some space to put the new indices in
2269 SmallVector<unsigned, 5> Idxs;
2271 // Add indices from the extract value instruction
2272 Idxs.append(I->idx_begin(), I->idx_end());
2274 // Add requested indices
2275 Idxs.append(idx_range.begin(), idx_range.end());
2277 assert(Idxs.size() == size
2278 && "Number of indices added not correct?");
2280 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2282 // Otherwise, we don't know (such as, extracting from a function return value
2283 // or load instruction)
2287 /// Analyze the specified pointer to see if it can be expressed as a base
2288 /// pointer plus a constant offset. Return the base and offset to the caller.
2289 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2290 const DataLayout *DL) {
2291 // Without DataLayout, conservatively assume 64-bit offsets, which is
2292 // the widest we support.
2293 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
2294 APInt ByteOffset(BitWidth, 0);
2296 if (Ptr->getType()->isVectorTy())
2299 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2301 APInt GEPOffset(BitWidth, 0);
2302 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
2305 ByteOffset += GEPOffset;
2308 Ptr = GEP->getPointerOperand();
2309 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2310 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2311 Ptr = cast<Operator>(Ptr)->getOperand(0);
2312 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2313 if (GA->mayBeOverridden())
2315 Ptr = GA->getAliasee();
2320 Offset = ByteOffset.getSExtValue();
2325 /// This function computes the length of a null-terminated C string pointed to
2326 /// by V. If successful, it returns true and returns the string in Str.
2327 /// If unsuccessful, it returns false.
2328 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2329 uint64_t Offset, bool TrimAtNul) {
2332 // Look through bitcast instructions and geps.
2333 V = V->stripPointerCasts();
2335 // If the value is a GEP instructionor constant expression, treat it as an
2337 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2338 // Make sure the GEP has exactly three arguments.
2339 if (GEP->getNumOperands() != 3)
2342 // Make sure the index-ee is a pointer to array of i8.
2343 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2344 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2345 if (!AT || !AT->getElementType()->isIntegerTy(8))
2348 // Check to make sure that the first operand of the GEP is an integer and
2349 // has value 0 so that we are sure we're indexing into the initializer.
2350 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2351 if (!FirstIdx || !FirstIdx->isZero())
2354 // If the second index isn't a ConstantInt, then this is a variable index
2355 // into the array. If this occurs, we can't say anything meaningful about
2357 uint64_t StartIdx = 0;
2358 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2359 StartIdx = CI->getZExtValue();
2362 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
2365 // The GEP instruction, constant or instruction, must reference a global
2366 // variable that is a constant and is initialized. The referenced constant
2367 // initializer is the array that we'll use for optimization.
2368 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2369 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2372 // Handle the all-zeros case
2373 if (GV->getInitializer()->isNullValue()) {
2374 // This is a degenerate case. The initializer is constant zero so the
2375 // length of the string must be zero.
2380 // Must be a Constant Array
2381 const ConstantDataArray *Array =
2382 dyn_cast<ConstantDataArray>(GV->getInitializer());
2383 if (!Array || !Array->isString())
2386 // Get the number of elements in the array
2387 uint64_t NumElts = Array->getType()->getArrayNumElements();
2389 // Start out with the entire array in the StringRef.
2390 Str = Array->getAsString();
2392 if (Offset > NumElts)
2395 // Skip over 'offset' bytes.
2396 Str = Str.substr(Offset);
2399 // Trim off the \0 and anything after it. If the array is not nul
2400 // terminated, we just return the whole end of string. The client may know
2401 // some other way that the string is length-bound.
2402 Str = Str.substr(0, Str.find('\0'));
2407 // These next two are very similar to the above, but also look through PHI
2409 // TODO: See if we can integrate these two together.
2411 /// If we can compute the length of the string pointed to by
2412 /// the specified pointer, return 'len+1'. If we can't, return 0.
2413 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2414 // Look through noop bitcast instructions.
2415 V = V->stripPointerCasts();
2417 // If this is a PHI node, there are two cases: either we have already seen it
2419 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2420 if (!PHIs.insert(PN).second)
2421 return ~0ULL; // already in the set.
2423 // If it was new, see if all the input strings are the same length.
2424 uint64_t LenSoFar = ~0ULL;
2425 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2426 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2427 if (Len == 0) return 0; // Unknown length -> unknown.
2429 if (Len == ~0ULL) continue;
2431 if (Len != LenSoFar && LenSoFar != ~0ULL)
2432 return 0; // Disagree -> unknown.
2436 // Success, all agree.
2440 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2441 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2442 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2443 if (Len1 == 0) return 0;
2444 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2445 if (Len2 == 0) return 0;
2446 if (Len1 == ~0ULL) return Len2;
2447 if (Len2 == ~0ULL) return Len1;
2448 if (Len1 != Len2) return 0;
2452 // Otherwise, see if we can read the string.
2454 if (!getConstantStringInfo(V, StrData))
2457 return StrData.size()+1;
2460 /// If we can compute the length of the string pointed to by
2461 /// the specified pointer, return 'len+1'. If we can't, return 0.
2462 uint64_t llvm::GetStringLength(Value *V) {
2463 if (!V->getType()->isPointerTy()) return 0;
2465 SmallPtrSet<PHINode*, 32> PHIs;
2466 uint64_t Len = GetStringLengthH(V, PHIs);
2467 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2468 // an empty string as a length.
2469 return Len == ~0ULL ? 1 : Len;
2473 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
2474 if (!V->getType()->isPointerTy())
2476 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2477 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2478 V = GEP->getPointerOperand();
2479 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2480 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2481 V = cast<Operator>(V)->getOperand(0);
2482 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2483 if (GA->mayBeOverridden())
2485 V = GA->getAliasee();
2487 // See if InstructionSimplify knows any relevant tricks.
2488 if (Instruction *I = dyn_cast<Instruction>(V))
2489 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2490 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
2497 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2503 llvm::GetUnderlyingObjects(Value *V,
2504 SmallVectorImpl<Value *> &Objects,
2505 const DataLayout *TD,
2506 unsigned MaxLookup) {
2507 SmallPtrSet<Value *, 4> Visited;
2508 SmallVector<Value *, 4> Worklist;
2509 Worklist.push_back(V);
2511 Value *P = Worklist.pop_back_val();
2512 P = GetUnderlyingObject(P, TD, MaxLookup);
2514 if (!Visited.insert(P).second)
2517 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2518 Worklist.push_back(SI->getTrueValue());
2519 Worklist.push_back(SI->getFalseValue());
2523 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2524 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2525 Worklist.push_back(PN->getIncomingValue(i));
2529 Objects.push_back(P);
2530 } while (!Worklist.empty());
2533 /// Return true if the only users of this pointer are lifetime markers.
2534 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2535 for (const User *U : V->users()) {
2536 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2537 if (!II) return false;
2539 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2540 II->getIntrinsicID() != Intrinsic::lifetime_end)
2546 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
2547 const DataLayout *TD) {
2548 const Operator *Inst = dyn_cast<Operator>(V);
2552 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2553 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2557 switch (Inst->getOpcode()) {
2560 case Instruction::UDiv:
2561 case Instruction::URem: {
2562 // x / y is undefined if y == 0.
2564 if (match(Inst->getOperand(1), m_APInt(V)))
2568 case Instruction::SDiv:
2569 case Instruction::SRem: {
2570 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
2572 if (match(Inst->getOperand(1), m_APInt(Y))) {
2575 // The numerator can't be MinSignedValue if the denominator is -1.
2576 if (match(Inst->getOperand(0), m_APInt(X)))
2577 return !Y->isMinSignedValue();
2578 // The numerator *might* be MinSignedValue.
2581 // The denominator is not 0 or -1, it's safe to proceed.
2587 case Instruction::Load: {
2588 const LoadInst *LI = cast<LoadInst>(Inst);
2589 if (!LI->isUnordered() ||
2590 // Speculative load may create a race that did not exist in the source.
2591 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2593 return LI->getPointerOperand()->isDereferenceablePointer(TD);
2595 case Instruction::Call: {
2596 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2597 switch (II->getIntrinsicID()) {
2598 // These synthetic intrinsics have no side-effects and just mark
2599 // information about their operands.
2600 // FIXME: There are other no-op synthetic instructions that potentially
2601 // should be considered at least *safe* to speculate...
2602 case Intrinsic::dbg_declare:
2603 case Intrinsic::dbg_value:
2606 case Intrinsic::bswap:
2607 case Intrinsic::ctlz:
2608 case Intrinsic::ctpop:
2609 case Intrinsic::cttz:
2610 case Intrinsic::objectsize:
2611 case Intrinsic::sadd_with_overflow:
2612 case Intrinsic::smul_with_overflow:
2613 case Intrinsic::ssub_with_overflow:
2614 case Intrinsic::uadd_with_overflow:
2615 case Intrinsic::umul_with_overflow:
2616 case Intrinsic::usub_with_overflow:
2618 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2619 // errno like libm sqrt would.
2620 case Intrinsic::sqrt:
2621 case Intrinsic::fma:
2622 case Intrinsic::fmuladd:
2623 case Intrinsic::fabs:
2624 case Intrinsic::minnum:
2625 case Intrinsic::maxnum:
2627 // TODO: some fp intrinsics are marked as having the same error handling
2628 // as libm. They're safe to speculate when they won't error.
2629 // TODO: are convert_{from,to}_fp16 safe?
2630 // TODO: can we list target-specific intrinsics here?
2634 return false; // The called function could have undefined behavior or
2635 // side-effects, even if marked readnone nounwind.
2637 case Instruction::VAArg:
2638 case Instruction::Alloca:
2639 case Instruction::Invoke:
2640 case Instruction::PHI:
2641 case Instruction::Store:
2642 case Instruction::Ret:
2643 case Instruction::Br:
2644 case Instruction::IndirectBr:
2645 case Instruction::Switch:
2646 case Instruction::Unreachable:
2647 case Instruction::Fence:
2648 case Instruction::LandingPad:
2649 case Instruction::AtomicRMW:
2650 case Instruction::AtomicCmpXchg:
2651 case Instruction::Resume:
2652 return false; // Misc instructions which have effects
2656 /// Return true if we know that the specified value is never null.
2657 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2658 // Alloca never returns null, malloc might.
2659 if (isa<AllocaInst>(V)) return true;
2661 // A byval, inalloca, or nonnull argument is never null.
2662 if (const Argument *A = dyn_cast<Argument>(V))
2663 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2665 // Global values are not null unless extern weak.
2666 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2667 return !GV->hasExternalWeakLinkage();
2669 // A Load tagged w/nonnull metadata is never null.
2670 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2671 return LI->getMetadata(LLVMContext::MD_nonnull);
2673 if (ImmutableCallSite CS = V)
2674 if (CS.isReturnNonNull())
2677 // operator new never returns null.
2678 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
2684 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
2685 const DataLayout *DL,
2686 AssumptionCache *AC,
2687 const Instruction *CxtI,
2688 const DominatorTree *DT) {
2689 // Multiplying n * m significant bits yields a result of n + m significant
2690 // bits. If the total number of significant bits does not exceed the
2691 // result bit width (minus 1), there is no overflow.
2692 // This means if we have enough leading zero bits in the operands
2693 // we can guarantee that the result does not overflow.
2694 // Ref: "Hacker's Delight" by Henry Warren
2695 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
2696 APInt LHSKnownZero(BitWidth, 0);
2697 APInt LHSKnownOne(BitWidth, 0);
2698 APInt RHSKnownZero(BitWidth, 0);
2699 APInt RHSKnownOne(BitWidth, 0);
2700 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
2702 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
2704 // Note that underestimating the number of zero bits gives a more
2705 // conservative answer.
2706 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
2707 RHSKnownZero.countLeadingOnes();
2708 // First handle the easy case: if we have enough zero bits there's
2709 // definitely no overflow.
2710 if (ZeroBits >= BitWidth)
2711 return OverflowResult::NeverOverflows;
2713 // Get the largest possible values for each operand.
2714 APInt LHSMax = ~LHSKnownZero;
2715 APInt RHSMax = ~RHSKnownZero;
2717 // We know the multiply operation doesn't overflow if the maximum values for
2718 // each operand will not overflow after we multiply them together.
2720 LHSMax.umul_ov(RHSMax, MaxOverflow);
2722 return OverflowResult::NeverOverflows;
2724 // We know it always overflows if multiplying the smallest possible values for
2725 // the operands also results in overflow.
2727 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
2729 return OverflowResult::AlwaysOverflows;
2731 return OverflowResult::MayOverflow;