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/AssumptionTracker.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.
68 AssumptionTracker *AT;
69 const Instruction *CxtI;
70 const DominatorTree *DT;
72 Query(AssumptionTracker *AT = nullptr, const Instruction *CxtI = nullptr,
73 const DominatorTree *DT = nullptr)
74 : AT(AT), CxtI(CxtI), DT(DT) {}
76 Query(const Query &Q, const Value *NewExcl)
77 : ExclInvs(Q.ExclInvs), AT(Q.AT), 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 AssumptionTracker *AT, const Instruction *CxtI,
106 const DominatorTree *DT) {
107 ::computeKnownBits(V, KnownZero, KnownOne, TD, Depth,
108 Query(AT, 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 AssumptionTracker *AT, const Instruction *CxtI,
118 const DominatorTree *DT) {
119 ::ComputeSignBit(V, KnownZero, KnownOne, TD, Depth,
120 Query(AT, 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 AssumptionTracker *AT,
128 const Instruction *CxtI,
129 const DominatorTree *DT) {
130 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
131 Query(AT, safeCxtI(V, CxtI), DT));
134 static bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
137 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
138 AssumptionTracker *AT, const Instruction *CxtI,
139 const DominatorTree *DT) {
140 return ::isKnownNonZero(V, TD, Depth, Query(AT, safeCxtI(V, CxtI), DT));
143 static bool MaskedValueIsZero(Value *V, const APInt &Mask,
144 const DataLayout *TD, unsigned Depth,
147 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
148 const DataLayout *TD, unsigned Depth,
149 AssumptionTracker *AT, const Instruction *CxtI,
150 const DominatorTree *DT) {
151 return ::MaskedValueIsZero(V, Mask, TD, Depth,
152 Query(AT, safeCxtI(V, CxtI), DT));
155 static unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
156 unsigned Depth, const Query &Q);
158 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
159 unsigned Depth, AssumptionTracker *AT,
160 const Instruction *CxtI,
161 const DominatorTree *DT) {
162 return ::ComputeNumSignBits(V, TD, Depth, Query(AT, safeCxtI(V, CxtI), DT));
165 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
166 APInt &KnownZero, APInt &KnownOne,
167 APInt &KnownZero2, APInt &KnownOne2,
168 const DataLayout *TD, unsigned Depth,
171 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
172 // We know that the top bits of C-X are clear if X contains less bits
173 // than C (i.e. no wrap-around can happen). For example, 20-X is
174 // positive if we can prove that X is >= 0 and < 16.
175 if (!CLHS->getValue().isNegative()) {
176 unsigned BitWidth = KnownZero.getBitWidth();
177 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
178 // NLZ can't be BitWidth with no sign bit
179 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
180 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
182 // If all of the MaskV bits are known to be zero, then we know the
183 // output top bits are zero, because we now know that the output is
185 if ((KnownZero2 & MaskV) == MaskV) {
186 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
187 // Top bits known zero.
188 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
194 unsigned BitWidth = KnownZero.getBitWidth();
196 // If an initial sequence of bits in the result is not needed, the
197 // corresponding bits in the operands are not needed.
198 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
199 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1, Q);
200 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
202 // Carry in a 1 for a subtract, rather than a 0.
203 APInt CarryIn(BitWidth, 0);
205 // Sum = LHS + ~RHS + 1
206 std::swap(KnownZero2, KnownOne2);
210 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
211 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
213 // Compute known bits of the carry.
214 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
215 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
217 // Compute set of known bits (where all three relevant bits are known).
218 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
219 APInt RHSKnown = KnownZero2 | KnownOne2;
220 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
221 APInt Known = LHSKnown & RHSKnown & CarryKnown;
223 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
224 "known bits of sum differ");
226 // Compute known bits of the result.
227 KnownZero = ~PossibleSumOne & Known;
228 KnownOne = PossibleSumOne & Known;
230 // Are we still trying to solve for the sign bit?
231 if (!Known.isNegative()) {
233 // Adding two non-negative numbers, or subtracting a negative number from
234 // a non-negative one, can't wrap into negative.
235 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
236 KnownZero |= APInt::getSignBit(BitWidth);
237 // Adding two negative numbers, or subtracting a non-negative number from
238 // a negative one, can't wrap into non-negative.
239 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
240 KnownOne |= APInt::getSignBit(BitWidth);
245 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
246 APInt &KnownZero, APInt &KnownOne,
247 APInt &KnownZero2, APInt &KnownOne2,
248 const DataLayout *TD, unsigned Depth,
250 unsigned BitWidth = KnownZero.getBitWidth();
251 computeKnownBits(Op1, KnownZero, KnownOne, TD, Depth+1, Q);
252 computeKnownBits(Op0, KnownZero2, KnownOne2, TD, Depth+1, Q);
254 bool isKnownNegative = false;
255 bool isKnownNonNegative = false;
256 // If the multiplication is known not to overflow, compute the sign bit.
259 // The product of a number with itself is non-negative.
260 isKnownNonNegative = true;
262 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
263 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
264 bool isKnownNegativeOp1 = KnownOne.isNegative();
265 bool isKnownNegativeOp0 = KnownOne2.isNegative();
266 // The product of two numbers with the same sign is non-negative.
267 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
268 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
269 // The product of a negative number and a non-negative number is either
271 if (!isKnownNonNegative)
272 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
273 isKnownNonZero(Op0, TD, Depth, Q)) ||
274 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
275 isKnownNonZero(Op1, TD, Depth, Q));
279 // If low bits are zero in either operand, output low known-0 bits.
280 // Also compute a conserative estimate for high known-0 bits.
281 // More trickiness is possible, but this is sufficient for the
282 // interesting case of alignment computation.
283 KnownOne.clearAllBits();
284 unsigned TrailZ = KnownZero.countTrailingOnes() +
285 KnownZero2.countTrailingOnes();
286 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
287 KnownZero2.countLeadingOnes(),
288 BitWidth) - BitWidth;
290 TrailZ = std::min(TrailZ, BitWidth);
291 LeadZ = std::min(LeadZ, BitWidth);
292 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
293 APInt::getHighBitsSet(BitWidth, LeadZ);
295 // Only make use of no-wrap flags if we failed to compute the sign bit
296 // directly. This matters if the multiplication always overflows, in
297 // which case we prefer to follow the result of the direct computation,
298 // though as the program is invoking undefined behaviour we can choose
299 // whatever we like here.
300 if (isKnownNonNegative && !KnownOne.isNegative())
301 KnownZero.setBit(BitWidth - 1);
302 else if (isKnownNegative && !KnownZero.isNegative())
303 KnownOne.setBit(BitWidth - 1);
306 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
308 unsigned BitWidth = KnownZero.getBitWidth();
309 unsigned NumRanges = Ranges.getNumOperands() / 2;
310 assert(NumRanges >= 1);
312 // Use the high end of the ranges to find leading zeros.
313 unsigned MinLeadingZeros = BitWidth;
314 for (unsigned i = 0; i < NumRanges; ++i) {
315 ConstantInt *Lower = cast<ConstantInt>(Ranges.getOperand(2*i + 0));
316 ConstantInt *Upper = cast<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.AT || !Q.CxtI)
486 unsigned BitWidth = KnownZero.getBitWidth();
488 Function *F = const_cast<Function*>(Q.CxtI->getParent()->getParent());
489 for (auto &CI : Q.AT->assumptions(F)) {
491 if (Q.ExclInvs.count(I))
494 // Warning: This loop can end up being somewhat performance sensetive.
495 // We're running this loop for once for each value queried resulting in a
496 // runtime of ~O(#assumes * #values).
498 assert(isa<IntrinsicInst>(I) &&
499 dyn_cast<IntrinsicInst>(I)->getIntrinsicID() == Intrinsic::assume &&
500 "must be an assume intrinsic");
502 Value *Arg = I->getArgOperand(0);
505 isValidAssumeForContext(I, Q, DL)) {
506 assert(BitWidth == 1 && "assume operand is not i1?");
507 KnownZero.clearAllBits();
508 KnownOne.setAllBits();
513 auto m_V = m_CombineOr(m_Specific(V),
514 m_CombineOr(m_PtrToInt(m_Specific(V)),
515 m_BitCast(m_Specific(V))));
517 CmpInst::Predicate Pred;
520 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
521 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
522 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
523 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
524 KnownZero |= RHSKnownZero;
525 KnownOne |= RHSKnownOne;
527 } else if (match(Arg, m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)),
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 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
533 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
535 // For those bits in the mask that are known to be one, we can propagate
536 // known bits from the RHS to V.
537 KnownZero |= RHSKnownZero & MaskKnownOne;
538 KnownOne |= RHSKnownOne & MaskKnownOne;
539 // assume(~(v & b) = a)
540 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
542 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
543 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
544 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
545 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
546 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
548 // For those bits in the mask that are known to be one, we can propagate
549 // inverted known bits from the RHS to V.
550 KnownZero |= RHSKnownOne & MaskKnownOne;
551 KnownOne |= RHSKnownZero & MaskKnownOne;
553 } else if (match(Arg, m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)),
555 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
556 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
557 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
558 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
559 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
561 // For those bits in B that are known to be zero, we can propagate known
562 // bits from the RHS to V.
563 KnownZero |= RHSKnownZero & BKnownZero;
564 KnownOne |= RHSKnownOne & BKnownZero;
565 // assume(~(v | b) = a)
566 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
568 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
569 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
570 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
571 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
572 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
574 // For those bits in B that are known to be zero, we can propagate
575 // inverted known bits from the RHS to V.
576 KnownZero |= RHSKnownOne & BKnownZero;
577 KnownOne |= RHSKnownZero & BKnownZero;
579 } else if (match(Arg, m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)),
581 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
582 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
583 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
584 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
585 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
587 // For those bits in B that are known to be zero, we can propagate known
588 // bits from the RHS to V. For those bits in B that are known to be one,
589 // we can propagate inverted known bits from the RHS to V.
590 KnownZero |= RHSKnownZero & BKnownZero;
591 KnownOne |= RHSKnownOne & BKnownZero;
592 KnownZero |= RHSKnownOne & BKnownOne;
593 KnownOne |= RHSKnownZero & BKnownOne;
594 // assume(~(v ^ b) = a)
595 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
597 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
598 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
599 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
600 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
601 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
603 // For those bits in B that are known to be zero, we can propagate
604 // inverted known bits from the RHS to V. For those bits in B that are
605 // known to be one, we can propagate known bits from the RHS to V.
606 KnownZero |= RHSKnownOne & BKnownZero;
607 KnownOne |= RHSKnownZero & BKnownZero;
608 KnownZero |= RHSKnownZero & BKnownOne;
609 KnownOne |= RHSKnownOne & BKnownOne;
610 // assume(v << c = a)
611 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
613 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
614 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
615 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
616 // For those bits in RHS that are known, we can propagate them to known
617 // bits in V shifted to the right by C.
618 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
619 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
620 // assume(~(v << c) = a)
621 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
623 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
624 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
625 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
626 // For those bits in RHS that are known, we can propagate them inverted
627 // to known bits in V shifted to the right by C.
628 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
629 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
630 // assume(v >> c = a)
631 } else if (match(Arg,
632 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
636 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
637 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
638 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
639 // For those bits in RHS that are known, we can propagate them to known
640 // bits in V shifted to the right by C.
641 KnownZero |= RHSKnownZero << C->getZExtValue();
642 KnownOne |= RHSKnownOne << C->getZExtValue();
643 // assume(~(v >> c) = a)
644 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
645 m_LShr(m_V, m_ConstantInt(C)),
646 m_AShr(m_V, m_ConstantInt(C)))),
648 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
649 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
650 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
651 // For those bits in RHS that are known, we can propagate them inverted
652 // to known bits in V shifted to the right by C.
653 KnownZero |= RHSKnownOne << C->getZExtValue();
654 KnownOne |= RHSKnownZero << C->getZExtValue();
655 // assume(v >=_s c) where c is non-negative
656 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
657 Pred == ICmpInst::ICMP_SGE &&
658 isValidAssumeForContext(I, Q, DL)) {
659 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
660 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
662 if (RHSKnownZero.isNegative()) {
663 // We know that the sign bit is zero.
664 KnownZero |= APInt::getSignBit(BitWidth);
666 // assume(v >_s c) where c is at least -1.
667 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
668 Pred == ICmpInst::ICMP_SGT &&
669 isValidAssumeForContext(I, Q, DL)) {
670 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
671 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
673 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
674 // We know that the sign bit is zero.
675 KnownZero |= APInt::getSignBit(BitWidth);
677 // assume(v <=_s c) where c is negative
678 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
679 Pred == ICmpInst::ICMP_SLE &&
680 isValidAssumeForContext(I, Q, DL)) {
681 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
682 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
684 if (RHSKnownOne.isNegative()) {
685 // We know that the sign bit is one.
686 KnownOne |= APInt::getSignBit(BitWidth);
688 // assume(v <_s c) where c is non-positive
689 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
690 Pred == ICmpInst::ICMP_SLT &&
691 isValidAssumeForContext(I, Q, DL)) {
692 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
693 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
695 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
696 // We know that the sign bit is one.
697 KnownOne |= APInt::getSignBit(BitWidth);
700 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
701 Pred == ICmpInst::ICMP_ULE &&
702 isValidAssumeForContext(I, Q, DL)) {
703 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
704 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
706 // Whatever high bits in c are zero are known to be zero.
708 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
710 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
711 Pred == ICmpInst::ICMP_ULT &&
712 isValidAssumeForContext(I, Q, DL)) {
713 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
714 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
716 // Whatever high bits in c are zero are known to be zero (if c is a power
717 // of 2, then one more).
718 if (isKnownToBeAPowerOfTwo(A, false, Depth+1, Query(Q, I)))
720 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
723 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
728 /// Determine which bits of V are known to be either zero or one and return
729 /// them in the KnownZero/KnownOne bit sets.
731 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
732 /// we cannot optimize based on the assumption that it is zero without changing
733 /// it to be an explicit zero. If we don't change it to zero, other code could
734 /// optimized based on the contradictory assumption that it is non-zero.
735 /// Because instcombine aggressively folds operations with undef args anyway,
736 /// this won't lose us code quality.
738 /// This function is defined on values with integer type, values with pointer
739 /// type (but only if TD is non-null), and vectors of integers. In the case
740 /// where V is a vector, known zero, and known one values are the
741 /// same width as the vector element, and the bit is set only if it is true
742 /// for all of the elements in the vector.
743 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
744 const DataLayout *TD, unsigned Depth,
746 assert(V && "No Value?");
747 assert(Depth <= MaxDepth && "Limit Search Depth");
748 unsigned BitWidth = KnownZero.getBitWidth();
750 assert((V->getType()->isIntOrIntVectorTy() ||
751 V->getType()->getScalarType()->isPointerTy()) &&
752 "Not integer or pointer type!");
754 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
755 (!V->getType()->isIntOrIntVectorTy() ||
756 V->getType()->getScalarSizeInBits() == BitWidth) &&
757 KnownZero.getBitWidth() == BitWidth &&
758 KnownOne.getBitWidth() == BitWidth &&
759 "V, KnownOne and KnownZero should have same BitWidth");
761 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
762 // We know all of the bits for a constant!
763 KnownOne = CI->getValue();
764 KnownZero = ~KnownOne;
767 // Null and aggregate-zero are all-zeros.
768 if (isa<ConstantPointerNull>(V) ||
769 isa<ConstantAggregateZero>(V)) {
770 KnownOne.clearAllBits();
771 KnownZero = APInt::getAllOnesValue(BitWidth);
774 // Handle a constant vector by taking the intersection of the known bits of
775 // each element. There is no real need to handle ConstantVector here, because
776 // we don't handle undef in any particularly useful way.
777 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
778 // We know that CDS must be a vector of integers. Take the intersection of
780 KnownZero.setAllBits(); KnownOne.setAllBits();
781 APInt Elt(KnownZero.getBitWidth(), 0);
782 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
783 Elt = CDS->getElementAsInteger(i);
790 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
791 // the bits of its aliasee.
792 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
793 if (GA->mayBeOverridden()) {
794 KnownZero.clearAllBits(); KnownOne.clearAllBits();
796 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1, Q);
801 // The address of an aligned GlobalValue has trailing zeros.
802 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
803 unsigned Align = GV->getAlignment();
804 if (Align == 0 && TD) {
805 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
806 Type *ObjectType = GVar->getType()->getElementType();
807 if (ObjectType->isSized()) {
808 // If the object is defined in the current Module, we'll be giving
809 // it the preferred alignment. Otherwise, we have to assume that it
810 // may only have the minimum ABI alignment.
811 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
812 Align = TD->getPreferredAlignment(GVar);
814 Align = TD->getABITypeAlignment(ObjectType);
819 KnownZero = APInt::getLowBitsSet(BitWidth,
820 countTrailingZeros(Align));
822 KnownZero.clearAllBits();
823 KnownOne.clearAllBits();
827 if (Argument *A = dyn_cast<Argument>(V)) {
828 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
830 if (!Align && TD && A->hasStructRetAttr()) {
831 // An sret parameter has at least the ABI alignment of the return type.
832 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
833 if (EltTy->isSized())
834 Align = TD->getABITypeAlignment(EltTy);
838 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
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 if (Depth == MaxDepth)
850 return; // Limit search depth.
852 // Check whether a nearby assume intrinsic can determine some known bits.
853 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
855 Operator *I = dyn_cast<Operator>(V);
858 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
859 switch (I->getOpcode()) {
861 case Instruction::Load:
862 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
863 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
865 case Instruction::And: {
866 // If either the LHS or the RHS are Zero, the result is zero.
867 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
868 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
870 // Output known-1 bits are only known if set in both the LHS & RHS.
871 KnownOne &= KnownOne2;
872 // Output known-0 are known to be clear if zero in either the LHS | RHS.
873 KnownZero |= KnownZero2;
876 case Instruction::Or: {
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-0 bits are only known if clear in both the LHS & RHS.
881 KnownZero &= KnownZero2;
882 // Output known-1 are known to be set if set in either the LHS | RHS.
883 KnownOne |= KnownOne2;
886 case Instruction::Xor: {
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 known if clear or set in both the LHS & RHS.
891 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
892 // Output known-1 are known to be set if set in only one of the LHS, RHS.
893 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
894 KnownZero = KnownZeroOut;
897 case Instruction::Mul: {
898 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
899 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
900 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
904 case Instruction::UDiv: {
905 // For the purposes of computing leading zeros we can conservatively
906 // treat a udiv as a logical right shift by the power of 2 known to
907 // be less than the denominator.
908 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
909 unsigned LeadZ = KnownZero2.countLeadingOnes();
911 KnownOne2.clearAllBits();
912 KnownZero2.clearAllBits();
913 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
914 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
915 if (RHSUnknownLeadingOnes != BitWidth)
916 LeadZ = std::min(BitWidth,
917 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
919 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
922 case Instruction::Select:
923 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1, Q);
924 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
926 // Only known if known in both the LHS and RHS.
927 KnownOne &= KnownOne2;
928 KnownZero &= KnownZero2;
930 case Instruction::FPTrunc:
931 case Instruction::FPExt:
932 case Instruction::FPToUI:
933 case Instruction::FPToSI:
934 case Instruction::SIToFP:
935 case Instruction::UIToFP:
936 break; // Can't work with floating point.
937 case Instruction::PtrToInt:
938 case Instruction::IntToPtr:
939 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
940 // We can't handle these if we don't know the pointer size.
942 // FALL THROUGH and handle them the same as zext/trunc.
943 case Instruction::ZExt:
944 case Instruction::Trunc: {
945 Type *SrcTy = I->getOperand(0)->getType();
947 unsigned SrcBitWidth;
948 // Note that we handle pointer operands here because of inttoptr/ptrtoint
949 // which fall through here.
951 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
953 SrcBitWidth = SrcTy->getScalarSizeInBits();
954 if (!SrcBitWidth) break;
957 assert(SrcBitWidth && "SrcBitWidth can't be zero");
958 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
959 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
960 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
961 KnownZero = KnownZero.zextOrTrunc(BitWidth);
962 KnownOne = KnownOne.zextOrTrunc(BitWidth);
963 // Any top bits are known to be zero.
964 if (BitWidth > SrcBitWidth)
965 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
968 case Instruction::BitCast: {
969 Type *SrcTy = I->getOperand(0)->getType();
970 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
971 // TODO: For now, not handling conversions like:
972 // (bitcast i64 %x to <2 x i32>)
973 !I->getType()->isVectorTy()) {
974 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
979 case Instruction::SExt: {
980 // Compute the bits in the result that are not present in the input.
981 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
983 KnownZero = KnownZero.trunc(SrcBitWidth);
984 KnownOne = KnownOne.trunc(SrcBitWidth);
985 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
986 KnownZero = KnownZero.zext(BitWidth);
987 KnownOne = KnownOne.zext(BitWidth);
989 // If the sign bit of the input is known set or clear, then we know the
990 // top bits of the result.
991 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
992 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
993 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
994 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
997 case Instruction::Shl:
998 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
999 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1000 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1001 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1002 KnownZero <<= ShiftAmt;
1003 KnownOne <<= ShiftAmt;
1004 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1007 case Instruction::LShr:
1008 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1009 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1010 // Compute the new bits that are at the top now.
1011 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1013 // Unsigned shift right.
1014 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1015 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1016 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1017 // high bits known zero.
1018 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1021 case Instruction::AShr:
1022 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1023 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1024 // Compute the new bits that are at the top now.
1025 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1027 // Signed shift right.
1028 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1029 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1030 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1032 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1033 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1034 KnownZero |= HighBits;
1035 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1036 KnownOne |= HighBits;
1039 case Instruction::Sub: {
1040 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1041 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1042 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1046 case Instruction::Add: {
1047 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1048 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1049 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1053 case Instruction::SRem:
1054 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1055 APInt RA = Rem->getValue().abs();
1056 if (RA.isPowerOf2()) {
1057 APInt LowBits = RA - 1;
1058 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD,
1061 // The low bits of the first operand are unchanged by the srem.
1062 KnownZero = KnownZero2 & LowBits;
1063 KnownOne = KnownOne2 & LowBits;
1065 // If the first operand is non-negative or has all low bits zero, then
1066 // the upper bits are all zero.
1067 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1068 KnownZero |= ~LowBits;
1070 // If the first operand is negative and not all low bits are zero, then
1071 // the upper bits are all one.
1072 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1073 KnownOne |= ~LowBits;
1075 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1079 // The sign bit is the LHS's sign bit, except when the result of the
1080 // remainder is zero.
1081 if (KnownZero.isNonNegative()) {
1082 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1083 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
1085 // If it's known zero, our sign bit is also zero.
1086 if (LHSKnownZero.isNegative())
1087 KnownZero.setBit(BitWidth - 1);
1091 case Instruction::URem: {
1092 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1093 APInt RA = Rem->getValue();
1094 if (RA.isPowerOf2()) {
1095 APInt LowBits = (RA - 1);
1096 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
1098 KnownZero |= ~LowBits;
1099 KnownOne &= LowBits;
1104 // Since the result is less than or equal to either operand, any leading
1105 // zero bits in either operand must also exist in the result.
1106 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1107 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
1109 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1110 KnownZero2.countLeadingOnes());
1111 KnownOne.clearAllBits();
1112 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1116 case Instruction::Alloca: {
1117 AllocaInst *AI = cast<AllocaInst>(V);
1118 unsigned Align = AI->getAlignment();
1119 if (Align == 0 && TD)
1120 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1123 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1126 case Instruction::GetElementPtr: {
1127 // Analyze all of the subscripts of this getelementptr instruction
1128 // to determine if we can prove known low zero bits.
1129 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1130 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
1132 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1134 gep_type_iterator GTI = gep_type_begin(I);
1135 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1136 Value *Index = I->getOperand(i);
1137 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1138 // Handle struct member offset arithmetic.
1144 // Handle case when index is vector zeroinitializer
1145 Constant *CIndex = cast<Constant>(Index);
1146 if (CIndex->isZeroValue())
1149 if (CIndex->getType()->isVectorTy())
1150 Index = CIndex->getSplatValue();
1152 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1153 const StructLayout *SL = TD->getStructLayout(STy);
1154 uint64_t Offset = SL->getElementOffset(Idx);
1155 TrailZ = std::min<unsigned>(TrailZ,
1156 countTrailingZeros(Offset));
1158 // Handle array index arithmetic.
1159 Type *IndexedTy = GTI.getIndexedType();
1160 if (!IndexedTy->isSized()) {
1164 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1165 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
1166 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1167 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1, Q);
1168 TrailZ = std::min(TrailZ,
1169 unsigned(countTrailingZeros(TypeSize) +
1170 LocalKnownZero.countTrailingOnes()));
1174 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1177 case Instruction::PHI: {
1178 PHINode *P = cast<PHINode>(I);
1179 // Handle the case of a simple two-predecessor recurrence PHI.
1180 // There's a lot more that could theoretically be done here, but
1181 // this is sufficient to catch some interesting cases.
1182 if (P->getNumIncomingValues() == 2) {
1183 for (unsigned i = 0; i != 2; ++i) {
1184 Value *L = P->getIncomingValue(i);
1185 Value *R = P->getIncomingValue(!i);
1186 Operator *LU = dyn_cast<Operator>(L);
1189 unsigned Opcode = LU->getOpcode();
1190 // Check for operations that have the property that if
1191 // both their operands have low zero bits, the result
1192 // will have low zero bits.
1193 if (Opcode == Instruction::Add ||
1194 Opcode == Instruction::Sub ||
1195 Opcode == Instruction::And ||
1196 Opcode == Instruction::Or ||
1197 Opcode == Instruction::Mul) {
1198 Value *LL = LU->getOperand(0);
1199 Value *LR = LU->getOperand(1);
1200 // Find a recurrence.
1207 // Ok, we have a PHI of the form L op= R. Check for low
1209 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1, Q);
1211 // We need to take the minimum number of known bits
1212 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1213 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1, Q);
1215 KnownZero = APInt::getLowBitsSet(BitWidth,
1216 std::min(KnownZero2.countTrailingOnes(),
1217 KnownZero3.countTrailingOnes()));
1223 // Unreachable blocks may have zero-operand PHI nodes.
1224 if (P->getNumIncomingValues() == 0)
1227 // Otherwise take the unions of the known bit sets of the operands,
1228 // taking conservative care to avoid excessive recursion.
1229 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1230 // Skip if every incoming value references to ourself.
1231 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1234 KnownZero = APInt::getAllOnesValue(BitWidth);
1235 KnownOne = APInt::getAllOnesValue(BitWidth);
1236 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1237 // Skip direct self references.
1238 if (P->getIncomingValue(i) == P) continue;
1240 KnownZero2 = APInt(BitWidth, 0);
1241 KnownOne2 = APInt(BitWidth, 0);
1242 // Recurse, but cap the recursion to one level, because we don't
1243 // want to waste time spinning around in loops.
1244 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
1246 KnownZero &= KnownZero2;
1247 KnownOne &= KnownOne2;
1248 // If all bits have been ruled out, there's no need to check
1250 if (!KnownZero && !KnownOne)
1256 case Instruction::Call:
1257 case Instruction::Invoke:
1258 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1259 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1260 // If a range metadata is attached to this IntrinsicInst, intersect the
1261 // explicit range specified by the metadata and the implicit range of
1263 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1264 switch (II->getIntrinsicID()) {
1266 case Intrinsic::ctlz:
1267 case Intrinsic::cttz: {
1268 unsigned LowBits = Log2_32(BitWidth)+1;
1269 // If this call is undefined for 0, the result will be less than 2^n.
1270 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1272 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1275 case Intrinsic::ctpop: {
1276 unsigned LowBits = Log2_32(BitWidth)+1;
1277 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1280 case Intrinsic::x86_sse42_crc32_64_64:
1281 KnownZero |= APInt::getHighBitsSet(64, 32);
1286 case Instruction::ExtractValue:
1287 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1288 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1289 if (EVI->getNumIndices() != 1) break;
1290 if (EVI->getIndices()[0] == 0) {
1291 switch (II->getIntrinsicID()) {
1293 case Intrinsic::uadd_with_overflow:
1294 case Intrinsic::sadd_with_overflow:
1295 computeKnownBitsAddSub(true, II->getArgOperand(0),
1296 II->getArgOperand(1), false, KnownZero,
1297 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1299 case Intrinsic::usub_with_overflow:
1300 case Intrinsic::ssub_with_overflow:
1301 computeKnownBitsAddSub(false, II->getArgOperand(0),
1302 II->getArgOperand(1), false, KnownZero,
1303 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1305 case Intrinsic::umul_with_overflow:
1306 case Intrinsic::smul_with_overflow:
1307 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
1308 false, KnownZero, KnownOne,
1309 KnownZero2, KnownOne2, TD, Depth, Q);
1316 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1319 /// Determine whether the sign bit is known to be zero or one.
1320 /// Convenience wrapper around computeKnownBits.
1321 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1322 const DataLayout *TD, unsigned Depth,
1324 unsigned BitWidth = getBitWidth(V->getType(), TD);
1330 APInt ZeroBits(BitWidth, 0);
1331 APInt OneBits(BitWidth, 0);
1332 computeKnownBits(V, ZeroBits, OneBits, TD, Depth, Q);
1333 KnownOne = OneBits[BitWidth - 1];
1334 KnownZero = ZeroBits[BitWidth - 1];
1337 /// Return true if the given value is known to have exactly one
1338 /// bit set when defined. For vectors return true if every element is known to
1339 /// be a power of two when defined. Supports values with integer or pointer
1340 /// types and vectors of integers.
1341 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1343 if (Constant *C = dyn_cast<Constant>(V)) {
1344 if (C->isNullValue())
1346 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1347 return CI->getValue().isPowerOf2();
1348 // TODO: Handle vector constants.
1351 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1352 // it is shifted off the end then the result is undefined.
1353 if (match(V, m_Shl(m_One(), m_Value())))
1356 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1357 // bottom. If it is shifted off the bottom then the result is undefined.
1358 if (match(V, m_LShr(m_SignBit(), m_Value())))
1361 // The remaining tests are all recursive, so bail out if we hit the limit.
1362 if (Depth++ == MaxDepth)
1365 Value *X = nullptr, *Y = nullptr;
1366 // A shift of a power of two is a power of two or zero.
1367 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1368 match(V, m_Shr(m_Value(X), m_Value()))))
1369 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q);
1371 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1372 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1374 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1376 isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1377 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1379 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1380 // A power of two and'd with anything is a power of two or zero.
1381 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q) ||
1382 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth, Q))
1384 // X & (-X) is always a power of two or zero.
1385 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1390 // Adding a power-of-two or zero to the same power-of-two or zero yields
1391 // either the original power-of-two, a larger power-of-two or zero.
1392 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1393 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1394 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1395 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1396 match(X, m_And(m_Value(), m_Specific(Y))))
1397 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1399 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1400 match(Y, m_And(m_Value(), m_Specific(X))))
1401 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1404 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1405 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1406 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth, Q);
1408 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1409 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth, Q);
1410 // If i8 V is a power of two or zero:
1411 // ZeroBits: 1 1 1 0 1 1 1 1
1412 // ~ZeroBits: 0 0 0 1 0 0 0 0
1413 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1414 // If OrZero isn't set, we cannot give back a zero result.
1415 // Make sure either the LHS or RHS has a bit set.
1416 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1421 // An exact divide or right shift can only shift off zero bits, so the result
1422 // is a power of two only if the first operand is a power of two and not
1423 // copying a sign bit (sdiv int_min, 2).
1424 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1425 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1426 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1433 /// \brief Test whether a GEP's result is known to be non-null.
1435 /// Uses properties inherent in a GEP to try to determine whether it is known
1438 /// Currently this routine does not support vector GEPs.
1439 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
1440 unsigned Depth, const Query &Q) {
1441 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1444 // FIXME: Support vector-GEPs.
1445 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1447 // If the base pointer is non-null, we cannot walk to a null address with an
1448 // inbounds GEP in address space zero.
1449 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1452 // Past this, if we don't have DataLayout, we can't do much.
1456 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1457 // If so, then the GEP cannot produce a null pointer, as doing so would
1458 // inherently violate the inbounds contract within address space zero.
1459 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1460 GTI != GTE; ++GTI) {
1461 // Struct types are easy -- they must always be indexed by a constant.
1462 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1463 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1464 unsigned ElementIdx = OpC->getZExtValue();
1465 const StructLayout *SL = DL->getStructLayout(STy);
1466 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1467 if (ElementOffset > 0)
1472 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1473 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
1476 // Fast path the constant operand case both for efficiency and so we don't
1477 // increment Depth when just zipping down an all-constant GEP.
1478 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1484 // We post-increment Depth here because while isKnownNonZero increments it
1485 // as well, when we pop back up that increment won't persist. We don't want
1486 // to recurse 10k times just because we have 10k GEP operands. We don't
1487 // bail completely out because we want to handle constant GEPs regardless
1489 if (Depth++ >= MaxDepth)
1492 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1499 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1500 /// ensure that the value it's attached to is never Value? 'RangeType' is
1501 /// is the type of the value described by the range.
1502 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1503 const APInt& Value) {
1504 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1505 assert(NumRanges >= 1);
1506 for (unsigned i = 0; i < NumRanges; ++i) {
1507 ConstantInt *Lower = cast<ConstantInt>(Ranges->getOperand(2*i + 0));
1508 ConstantInt *Upper = cast<ConstantInt>(Ranges->getOperand(2*i + 1));
1509 ConstantRange Range(Lower->getValue(), Upper->getValue());
1510 if (Range.contains(Value))
1516 /// Return true if the given value is known to be non-zero when defined.
1517 /// For vectors return true if every element is known to be non-zero when
1518 /// defined. Supports values with integer or pointer type and vectors of
1520 bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
1522 if (Constant *C = dyn_cast<Constant>(V)) {
1523 if (C->isNullValue())
1525 if (isa<ConstantInt>(C))
1526 // Must be non-zero due to null test above.
1528 // TODO: Handle vectors
1532 if (Instruction* I = dyn_cast<Instruction>(V)) {
1533 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1534 // If the possible ranges don't contain zero, then the value is
1535 // definitely non-zero.
1536 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1537 const APInt ZeroValue(Ty->getBitWidth(), 0);
1538 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1544 // The remaining tests are all recursive, so bail out if we hit the limit.
1545 if (Depth++ >= MaxDepth)
1548 // Check for pointer simplifications.
1549 if (V->getType()->isPointerTy()) {
1550 if (isKnownNonNull(V))
1552 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1553 if (isGEPKnownNonNull(GEP, TD, Depth, Q))
1557 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1559 // X | Y != 0 if X != 0 or Y != 0.
1560 Value *X = nullptr, *Y = nullptr;
1561 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1562 return isKnownNonZero(X, TD, Depth, Q) ||
1563 isKnownNonZero(Y, TD, Depth, Q);
1565 // ext X != 0 if X != 0.
1566 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1567 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth, Q);
1569 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1570 // if the lowest bit is shifted off the end.
1571 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1572 // shl nuw can't remove any non-zero bits.
1573 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1574 if (BO->hasNoUnsignedWrap())
1575 return isKnownNonZero(X, TD, Depth, Q);
1577 APInt KnownZero(BitWidth, 0);
1578 APInt KnownOne(BitWidth, 0);
1579 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1583 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1584 // defined if the sign bit is shifted off the end.
1585 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1586 // shr exact can only shift out zero bits.
1587 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1589 return isKnownNonZero(X, TD, Depth, Q);
1591 bool XKnownNonNegative, XKnownNegative;
1592 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1596 // div exact can only produce a zero if the dividend is zero.
1597 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1598 return isKnownNonZero(X, TD, Depth, Q);
1601 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1602 bool XKnownNonNegative, XKnownNegative;
1603 bool YKnownNonNegative, YKnownNegative;
1604 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1605 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth, Q);
1607 // If X and Y are both non-negative (as signed values) then their sum is not
1608 // zero unless both X and Y are zero.
1609 if (XKnownNonNegative && YKnownNonNegative)
1610 if (isKnownNonZero(X, TD, Depth, Q) ||
1611 isKnownNonZero(Y, TD, Depth, Q))
1614 // If X and Y are both negative (as signed values) then their sum is not
1615 // zero unless both X and Y equal INT_MIN.
1616 if (BitWidth && XKnownNegative && YKnownNegative) {
1617 APInt KnownZero(BitWidth, 0);
1618 APInt KnownOne(BitWidth, 0);
1619 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1620 // The sign bit of X is set. If some other bit is set then X is not equal
1622 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1623 if ((KnownOne & Mask) != 0)
1625 // The sign bit of Y is set. If some other bit is set then Y is not equal
1627 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth, Q);
1628 if ((KnownOne & Mask) != 0)
1632 // The sum of a non-negative number and a power of two is not zero.
1633 if (XKnownNonNegative &&
1634 isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth, Q))
1636 if (YKnownNonNegative &&
1637 isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth, Q))
1641 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1642 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1643 // If X and Y are non-zero then so is X * Y as long as the multiplication
1644 // does not overflow.
1645 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1646 isKnownNonZero(X, TD, Depth, Q) &&
1647 isKnownNonZero(Y, TD, Depth, Q))
1650 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1651 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1652 if (isKnownNonZero(SI->getTrueValue(), TD, Depth, Q) &&
1653 isKnownNonZero(SI->getFalseValue(), TD, Depth, Q))
1657 if (!BitWidth) return false;
1658 APInt KnownZero(BitWidth, 0);
1659 APInt KnownOne(BitWidth, 0);
1660 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1661 return KnownOne != 0;
1664 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1665 /// simplify operations downstream. Mask is known to be zero for bits that V
1668 /// This function is defined on values with integer type, values with pointer
1669 /// type (but only if TD is non-null), and vectors of integers. In the case
1670 /// where V is a vector, the mask, known zero, and known one values are the
1671 /// same width as the vector element, and the bit is set only if it is true
1672 /// for all of the elements in the vector.
1673 bool MaskedValueIsZero(Value *V, const APInt &Mask,
1674 const DataLayout *TD, unsigned Depth,
1676 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1677 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1678 return (KnownZero & Mask) == Mask;
1683 /// Return the number of times the sign bit of the register is replicated into
1684 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1685 /// (itself), but other cases can give us information. For example, immediately
1686 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1687 /// other, so we return 3.
1689 /// 'Op' must have a scalar integer type.
1691 unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
1692 unsigned Depth, const Query &Q) {
1693 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1694 "ComputeNumSignBits requires a DataLayout object to operate "
1695 "on non-integer values!");
1696 Type *Ty = V->getType();
1697 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1698 Ty->getScalarSizeInBits();
1700 unsigned FirstAnswer = 1;
1702 // Note that ConstantInt is handled by the general computeKnownBits case
1706 return 1; // Limit search depth.
1708 Operator *U = dyn_cast<Operator>(V);
1709 switch (Operator::getOpcode(V)) {
1711 case Instruction::SExt:
1712 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1713 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q) + Tmp;
1715 case Instruction::AShr: {
1716 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1717 // ashr X, C -> adds C sign bits. Vectors too.
1719 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1720 Tmp += ShAmt->getZExtValue();
1721 if (Tmp > TyBits) Tmp = TyBits;
1725 case Instruction::Shl: {
1727 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1728 // shl destroys sign bits.
1729 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1730 Tmp2 = ShAmt->getZExtValue();
1731 if (Tmp2 >= TyBits || // Bad shift.
1732 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1737 case Instruction::And:
1738 case Instruction::Or:
1739 case Instruction::Xor: // NOT is handled here.
1740 // Logical binary ops preserve the number of sign bits at the worst.
1741 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1743 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1744 FirstAnswer = std::min(Tmp, Tmp2);
1745 // We computed what we know about the sign bits as our first
1746 // answer. Now proceed to the generic code that uses
1747 // computeKnownBits, and pick whichever answer is better.
1751 case Instruction::Select:
1752 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1753 if (Tmp == 1) return 1; // Early out.
1754 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1, Q);
1755 return std::min(Tmp, Tmp2);
1757 case Instruction::Add:
1758 // Add can have at most one carry bit. Thus we know that the output
1759 // is, at worst, one more bit than the inputs.
1760 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1761 if (Tmp == 1) return 1; // Early out.
1763 // Special case decrementing a value (ADD X, -1):
1764 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1765 if (CRHS->isAllOnesValue()) {
1766 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1767 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1769 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1771 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1774 // If we are subtracting one from a positive number, there is no carry
1775 // out of the result.
1776 if (KnownZero.isNegative())
1780 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1781 if (Tmp2 == 1) return 1;
1782 return std::min(Tmp, Tmp2)-1;
1784 case Instruction::Sub:
1785 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1786 if (Tmp2 == 1) return 1;
1789 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1790 if (CLHS->isNullValue()) {
1791 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1792 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
1793 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1795 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1798 // If the input is known to be positive (the sign bit is known clear),
1799 // the output of the NEG has the same number of sign bits as the input.
1800 if (KnownZero.isNegative())
1803 // Otherwise, we treat this like a SUB.
1806 // Sub can have at most one carry bit. Thus we know that the output
1807 // is, at worst, one more bit than the inputs.
1808 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1809 if (Tmp == 1) return 1; // Early out.
1810 return std::min(Tmp, Tmp2)-1;
1812 case Instruction::PHI: {
1813 PHINode *PN = cast<PHINode>(U);
1814 // Don't analyze large in-degree PHIs.
1815 if (PN->getNumIncomingValues() > 4) break;
1817 // Take the minimum of all incoming values. This can't infinitely loop
1818 // because of our depth threshold.
1819 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1, Q);
1820 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1821 if (Tmp == 1) return Tmp;
1823 ComputeNumSignBits(PN->getIncomingValue(i), TD,
1829 case Instruction::Trunc:
1830 // FIXME: it's tricky to do anything useful for this, but it is an important
1831 // case for targets like X86.
1835 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1836 // use this information.
1837 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1839 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1841 if (KnownZero.isNegative()) { // sign bit is 0
1843 } else if (KnownOne.isNegative()) { // sign bit is 1;
1850 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1851 // the number of identical bits in the top of the input value.
1853 Mask <<= Mask.getBitWidth()-TyBits;
1854 // Return # leading zeros. We use 'min' here in case Val was zero before
1855 // shifting. We don't want to return '64' as for an i32 "0".
1856 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1859 /// This function computes the integer multiple of Base that equals V.
1860 /// If successful, it returns true and returns the multiple in
1861 /// Multiple. If unsuccessful, it returns false. It looks
1862 /// through SExt instructions only if LookThroughSExt is true.
1863 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1864 bool LookThroughSExt, unsigned Depth) {
1865 const unsigned MaxDepth = 6;
1867 assert(V && "No Value?");
1868 assert(Depth <= MaxDepth && "Limit Search Depth");
1869 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1871 Type *T = V->getType();
1873 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1883 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1884 Constant *BaseVal = ConstantInt::get(T, Base);
1885 if (CO && CO == BaseVal) {
1887 Multiple = ConstantInt::get(T, 1);
1891 if (CI && CI->getZExtValue() % Base == 0) {
1892 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1896 if (Depth == MaxDepth) return false; // Limit search depth.
1898 Operator *I = dyn_cast<Operator>(V);
1899 if (!I) return false;
1901 switch (I->getOpcode()) {
1903 case Instruction::SExt:
1904 if (!LookThroughSExt) return false;
1905 // otherwise fall through to ZExt
1906 case Instruction::ZExt:
1907 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1908 LookThroughSExt, Depth+1);
1909 case Instruction::Shl:
1910 case Instruction::Mul: {
1911 Value *Op0 = I->getOperand(0);
1912 Value *Op1 = I->getOperand(1);
1914 if (I->getOpcode() == Instruction::Shl) {
1915 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1916 if (!Op1CI) return false;
1917 // Turn Op0 << Op1 into Op0 * 2^Op1
1918 APInt Op1Int = Op1CI->getValue();
1919 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1920 APInt API(Op1Int.getBitWidth(), 0);
1921 API.setBit(BitToSet);
1922 Op1 = ConstantInt::get(V->getContext(), API);
1925 Value *Mul0 = nullptr;
1926 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1927 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1928 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1929 if (Op1C->getType()->getPrimitiveSizeInBits() <
1930 MulC->getType()->getPrimitiveSizeInBits())
1931 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1932 if (Op1C->getType()->getPrimitiveSizeInBits() >
1933 MulC->getType()->getPrimitiveSizeInBits())
1934 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1936 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1937 Multiple = ConstantExpr::getMul(MulC, Op1C);
1941 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1942 if (Mul0CI->getValue() == 1) {
1943 // V == Base * Op1, so return Op1
1949 Value *Mul1 = nullptr;
1950 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1951 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1952 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1953 if (Op0C->getType()->getPrimitiveSizeInBits() <
1954 MulC->getType()->getPrimitiveSizeInBits())
1955 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1956 if (Op0C->getType()->getPrimitiveSizeInBits() >
1957 MulC->getType()->getPrimitiveSizeInBits())
1958 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1960 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1961 Multiple = ConstantExpr::getMul(MulC, Op0C);
1965 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1966 if (Mul1CI->getValue() == 1) {
1967 // V == Base * Op0, so return Op0
1975 // We could not determine if V is a multiple of Base.
1979 /// Return true if we can prove that the specified FP value is never equal to
1982 /// NOTE: this function will need to be revisited when we support non-default
1985 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1986 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1987 return !CFP->getValueAPF().isNegZero();
1990 return 1; // Limit search depth.
1992 const Operator *I = dyn_cast<Operator>(V);
1993 if (!I) return false;
1995 // Check if the nsz fast-math flag is set
1996 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1997 if (FPO->hasNoSignedZeros())
2000 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2001 if (I->getOpcode() == Instruction::FAdd)
2002 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2003 if (CFP->isNullValue())
2006 // sitofp and uitofp turn into +0.0 for zero.
2007 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2010 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2011 // sqrt(-0.0) = -0.0, no other negative results are possible.
2012 if (II->getIntrinsicID() == Intrinsic::sqrt)
2013 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2015 if (const CallInst *CI = dyn_cast<CallInst>(I))
2016 if (const Function *F = CI->getCalledFunction()) {
2017 if (F->isDeclaration()) {
2019 if (F->getName() == "abs") return true;
2020 // fabs[lf](x) != -0.0
2021 if (F->getName() == "fabs") return true;
2022 if (F->getName() == "fabsf") return true;
2023 if (F->getName() == "fabsl") return true;
2024 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2025 F->getName() == "sqrtl")
2026 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2033 /// If the specified value can be set by repeating the same byte in memory,
2034 /// return the i8 value that it is represented with. This is
2035 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2036 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2037 /// byte store (e.g. i16 0x1234), return null.
2038 Value *llvm::isBytewiseValue(Value *V) {
2039 // All byte-wide stores are splatable, even of arbitrary variables.
2040 if (V->getType()->isIntegerTy(8)) return V;
2042 // Handle 'null' ConstantArrayZero etc.
2043 if (Constant *C = dyn_cast<Constant>(V))
2044 if (C->isNullValue())
2045 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2047 // Constant float and double values can be handled as integer values if the
2048 // corresponding integer value is "byteable". An important case is 0.0.
2049 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2050 if (CFP->getType()->isFloatTy())
2051 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2052 if (CFP->getType()->isDoubleTy())
2053 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2054 // Don't handle long double formats, which have strange constraints.
2057 // We can handle constant integers that are power of two in size and a
2058 // multiple of 8 bits.
2059 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2060 unsigned Width = CI->getBitWidth();
2061 if (isPowerOf2_32(Width) && Width > 8) {
2062 // We can handle this value if the recursive binary decomposition is the
2063 // same at all levels.
2064 APInt Val = CI->getValue();
2066 while (Val.getBitWidth() != 8) {
2067 unsigned NextWidth = Val.getBitWidth()/2;
2068 Val2 = Val.lshr(NextWidth);
2069 Val2 = Val2.trunc(Val.getBitWidth()/2);
2070 Val = Val.trunc(Val.getBitWidth()/2);
2072 // If the top/bottom halves aren't the same, reject it.
2076 return ConstantInt::get(V->getContext(), Val);
2080 // A ConstantDataArray/Vector is splatable if all its members are equal and
2082 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2083 Value *Elt = CA->getElementAsConstant(0);
2084 Value *Val = isBytewiseValue(Elt);
2088 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2089 if (CA->getElementAsConstant(I) != Elt)
2095 // Conceptually, we could handle things like:
2096 // %a = zext i8 %X to i16
2097 // %b = shl i16 %a, 8
2098 // %c = or i16 %a, %b
2099 // but until there is an example that actually needs this, it doesn't seem
2100 // worth worrying about.
2105 // This is the recursive version of BuildSubAggregate. It takes a few different
2106 // arguments. Idxs is the index within the nested struct From that we are
2107 // looking at now (which is of type IndexedType). IdxSkip is the number of
2108 // indices from Idxs that should be left out when inserting into the resulting
2109 // struct. To is the result struct built so far, new insertvalue instructions
2111 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2112 SmallVectorImpl<unsigned> &Idxs,
2114 Instruction *InsertBefore) {
2115 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2117 // Save the original To argument so we can modify it
2119 // General case, the type indexed by Idxs is a struct
2120 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2121 // Process each struct element recursively
2124 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2128 // Couldn't find any inserted value for this index? Cleanup
2129 while (PrevTo != OrigTo) {
2130 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2131 PrevTo = Del->getAggregateOperand();
2132 Del->eraseFromParent();
2134 // Stop processing elements
2138 // If we successfully found a value for each of our subaggregates
2142 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2143 // the struct's elements had a value that was inserted directly. In the latter
2144 // case, perhaps we can't determine each of the subelements individually, but
2145 // we might be able to find the complete struct somewhere.
2147 // Find the value that is at that particular spot
2148 Value *V = FindInsertedValue(From, Idxs);
2153 // Insert the value in the new (sub) aggregrate
2154 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2155 "tmp", InsertBefore);
2158 // This helper takes a nested struct and extracts a part of it (which is again a
2159 // struct) into a new value. For example, given the struct:
2160 // { a, { b, { c, d }, e } }
2161 // and the indices "1, 1" this returns
2164 // It does this by inserting an insertvalue for each element in the resulting
2165 // struct, as opposed to just inserting a single struct. This will only work if
2166 // each of the elements of the substruct are known (ie, inserted into From by an
2167 // insertvalue instruction somewhere).
2169 // All inserted insertvalue instructions are inserted before InsertBefore
2170 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2171 Instruction *InsertBefore) {
2172 assert(InsertBefore && "Must have someplace to insert!");
2173 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2175 Value *To = UndefValue::get(IndexedType);
2176 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2177 unsigned IdxSkip = Idxs.size();
2179 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2182 /// Given an aggregrate and an sequence of indices, see if
2183 /// the scalar value indexed is already around as a register, for example if it
2184 /// were inserted directly into the aggregrate.
2186 /// If InsertBefore is not null, this function will duplicate (modified)
2187 /// insertvalues when a part of a nested struct is extracted.
2188 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2189 Instruction *InsertBefore) {
2190 // Nothing to index? Just return V then (this is useful at the end of our
2192 if (idx_range.empty())
2194 // We have indices, so V should have an indexable type.
2195 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2196 "Not looking at a struct or array?");
2197 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2198 "Invalid indices for type?");
2200 if (Constant *C = dyn_cast<Constant>(V)) {
2201 C = C->getAggregateElement(idx_range[0]);
2202 if (!C) return nullptr;
2203 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2206 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2207 // Loop the indices for the insertvalue instruction in parallel with the
2208 // requested indices
2209 const unsigned *req_idx = idx_range.begin();
2210 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2211 i != e; ++i, ++req_idx) {
2212 if (req_idx == idx_range.end()) {
2213 // We can't handle this without inserting insertvalues
2217 // The requested index identifies a part of a nested aggregate. Handle
2218 // this specially. For example,
2219 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2220 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2221 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2222 // This can be changed into
2223 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2224 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2225 // which allows the unused 0,0 element from the nested struct to be
2227 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2231 // This insert value inserts something else than what we are looking for.
2232 // See if the (aggregrate) value inserted into has the value we are
2233 // looking for, then.
2235 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2238 // If we end up here, the indices of the insertvalue match with those
2239 // requested (though possibly only partially). Now we recursively look at
2240 // the inserted value, passing any remaining indices.
2241 return FindInsertedValue(I->getInsertedValueOperand(),
2242 makeArrayRef(req_idx, idx_range.end()),
2246 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2247 // If we're extracting a value from an aggregrate that was extracted from
2248 // something else, we can extract from that something else directly instead.
2249 // However, we will need to chain I's indices with the requested indices.
2251 // Calculate the number of indices required
2252 unsigned size = I->getNumIndices() + idx_range.size();
2253 // Allocate some space to put the new indices in
2254 SmallVector<unsigned, 5> Idxs;
2256 // Add indices from the extract value instruction
2257 Idxs.append(I->idx_begin(), I->idx_end());
2259 // Add requested indices
2260 Idxs.append(idx_range.begin(), idx_range.end());
2262 assert(Idxs.size() == size
2263 && "Number of indices added not correct?");
2265 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2267 // Otherwise, we don't know (such as, extracting from a function return value
2268 // or load instruction)
2272 /// Analyze the specified pointer to see if it can be expressed as a base
2273 /// pointer plus a constant offset. Return the base and offset to the caller.
2274 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2275 const DataLayout *DL) {
2276 // Without DataLayout, conservatively assume 64-bit offsets, which is
2277 // the widest we support.
2278 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
2279 APInt ByteOffset(BitWidth, 0);
2281 if (Ptr->getType()->isVectorTy())
2284 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2286 APInt GEPOffset(BitWidth, 0);
2287 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
2290 ByteOffset += GEPOffset;
2293 Ptr = GEP->getPointerOperand();
2294 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2295 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2296 Ptr = cast<Operator>(Ptr)->getOperand(0);
2297 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2298 if (GA->mayBeOverridden())
2300 Ptr = GA->getAliasee();
2305 Offset = ByteOffset.getSExtValue();
2310 /// This function computes the length of a null-terminated C string pointed to
2311 /// by V. If successful, it returns true and returns the string in Str.
2312 /// If unsuccessful, it returns false.
2313 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2314 uint64_t Offset, bool TrimAtNul) {
2317 // Look through bitcast instructions and geps.
2318 V = V->stripPointerCasts();
2320 // If the value is a GEP instructionor constant expression, treat it as an
2322 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2323 // Make sure the GEP has exactly three arguments.
2324 if (GEP->getNumOperands() != 3)
2327 // Make sure the index-ee is a pointer to array of i8.
2328 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2329 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2330 if (!AT || !AT->getElementType()->isIntegerTy(8))
2333 // Check to make sure that the first operand of the GEP is an integer and
2334 // has value 0 so that we are sure we're indexing into the initializer.
2335 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2336 if (!FirstIdx || !FirstIdx->isZero())
2339 // If the second index isn't a ConstantInt, then this is a variable index
2340 // into the array. If this occurs, we can't say anything meaningful about
2342 uint64_t StartIdx = 0;
2343 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2344 StartIdx = CI->getZExtValue();
2347 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
2350 // The GEP instruction, constant or instruction, must reference a global
2351 // variable that is a constant and is initialized. The referenced constant
2352 // initializer is the array that we'll use for optimization.
2353 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2354 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2357 // Handle the all-zeros case
2358 if (GV->getInitializer()->isNullValue()) {
2359 // This is a degenerate case. The initializer is constant zero so the
2360 // length of the string must be zero.
2365 // Must be a Constant Array
2366 const ConstantDataArray *Array =
2367 dyn_cast<ConstantDataArray>(GV->getInitializer());
2368 if (!Array || !Array->isString())
2371 // Get the number of elements in the array
2372 uint64_t NumElts = Array->getType()->getArrayNumElements();
2374 // Start out with the entire array in the StringRef.
2375 Str = Array->getAsString();
2377 if (Offset > NumElts)
2380 // Skip over 'offset' bytes.
2381 Str = Str.substr(Offset);
2384 // Trim off the \0 and anything after it. If the array is not nul
2385 // terminated, we just return the whole end of string. The client may know
2386 // some other way that the string is length-bound.
2387 Str = Str.substr(0, Str.find('\0'));
2392 // These next two are very similar to the above, but also look through PHI
2394 // TODO: See if we can integrate these two together.
2396 /// If we can compute the length of the string pointed to by
2397 /// the specified pointer, return 'len+1'. If we can't, return 0.
2398 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2399 // Look through noop bitcast instructions.
2400 V = V->stripPointerCasts();
2402 // If this is a PHI node, there are two cases: either we have already seen it
2404 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2405 if (!PHIs.insert(PN).second)
2406 return ~0ULL; // already in the set.
2408 // If it was new, see if all the input strings are the same length.
2409 uint64_t LenSoFar = ~0ULL;
2410 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2411 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2412 if (Len == 0) return 0; // Unknown length -> unknown.
2414 if (Len == ~0ULL) continue;
2416 if (Len != LenSoFar && LenSoFar != ~0ULL)
2417 return 0; // Disagree -> unknown.
2421 // Success, all agree.
2425 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2426 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2427 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2428 if (Len1 == 0) return 0;
2429 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2430 if (Len2 == 0) return 0;
2431 if (Len1 == ~0ULL) return Len2;
2432 if (Len2 == ~0ULL) return Len1;
2433 if (Len1 != Len2) return 0;
2437 // Otherwise, see if we can read the string.
2439 if (!getConstantStringInfo(V, StrData))
2442 return StrData.size()+1;
2445 /// If we can compute the length of the string pointed to by
2446 /// the specified pointer, return 'len+1'. If we can't, return 0.
2447 uint64_t llvm::GetStringLength(Value *V) {
2448 if (!V->getType()->isPointerTy()) return 0;
2450 SmallPtrSet<PHINode*, 32> PHIs;
2451 uint64_t Len = GetStringLengthH(V, PHIs);
2452 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2453 // an empty string as a length.
2454 return Len == ~0ULL ? 1 : Len;
2458 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
2459 if (!V->getType()->isPointerTy())
2461 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2462 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2463 V = GEP->getPointerOperand();
2464 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2465 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2466 V = cast<Operator>(V)->getOperand(0);
2467 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2468 if (GA->mayBeOverridden())
2470 V = GA->getAliasee();
2472 // See if InstructionSimplify knows any relevant tricks.
2473 if (Instruction *I = dyn_cast<Instruction>(V))
2474 // TODO: Acquire a DominatorTree and AssumptionTracker and use them.
2475 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
2482 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2488 llvm::GetUnderlyingObjects(Value *V,
2489 SmallVectorImpl<Value *> &Objects,
2490 const DataLayout *TD,
2491 unsigned MaxLookup) {
2492 SmallPtrSet<Value *, 4> Visited;
2493 SmallVector<Value *, 4> Worklist;
2494 Worklist.push_back(V);
2496 Value *P = Worklist.pop_back_val();
2497 P = GetUnderlyingObject(P, TD, MaxLookup);
2499 if (!Visited.insert(P).second)
2502 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2503 Worklist.push_back(SI->getTrueValue());
2504 Worklist.push_back(SI->getFalseValue());
2508 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2509 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2510 Worklist.push_back(PN->getIncomingValue(i));
2514 Objects.push_back(P);
2515 } while (!Worklist.empty());
2518 /// Return true if the only users of this pointer are lifetime markers.
2519 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2520 for (const User *U : V->users()) {
2521 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2522 if (!II) return false;
2524 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2525 II->getIntrinsicID() != Intrinsic::lifetime_end)
2531 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
2532 const DataLayout *TD) {
2533 const Operator *Inst = dyn_cast<Operator>(V);
2537 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2538 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2542 switch (Inst->getOpcode()) {
2545 case Instruction::UDiv:
2546 case Instruction::URem: {
2547 // x / y is undefined if y == 0.
2549 if (match(Inst->getOperand(1), m_APInt(V)))
2553 case Instruction::SDiv:
2554 case Instruction::SRem: {
2555 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
2557 if (match(Inst->getOperand(1), m_APInt(Y))) {
2560 // The numerator can't be MinSignedValue if the denominator is -1.
2561 if (match(Inst->getOperand(0), m_APInt(X)))
2562 return !Y->isMinSignedValue();
2563 // The numerator *might* be MinSignedValue.
2566 // The denominator is not 0 or -1, it's safe to proceed.
2572 case Instruction::Load: {
2573 const LoadInst *LI = cast<LoadInst>(Inst);
2574 if (!LI->isUnordered() ||
2575 // Speculative load may create a race that did not exist in the source.
2576 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2578 return LI->getPointerOperand()->isDereferenceablePointer(TD);
2580 case Instruction::Call: {
2581 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2582 switch (II->getIntrinsicID()) {
2583 // These synthetic intrinsics have no side-effects and just mark
2584 // information about their operands.
2585 // FIXME: There are other no-op synthetic instructions that potentially
2586 // should be considered at least *safe* to speculate...
2587 case Intrinsic::dbg_declare:
2588 case Intrinsic::dbg_value:
2591 case Intrinsic::bswap:
2592 case Intrinsic::ctlz:
2593 case Intrinsic::ctpop:
2594 case Intrinsic::cttz:
2595 case Intrinsic::objectsize:
2596 case Intrinsic::sadd_with_overflow:
2597 case Intrinsic::smul_with_overflow:
2598 case Intrinsic::ssub_with_overflow:
2599 case Intrinsic::uadd_with_overflow:
2600 case Intrinsic::umul_with_overflow:
2601 case Intrinsic::usub_with_overflow:
2603 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2604 // errno like libm sqrt would.
2605 case Intrinsic::sqrt:
2606 case Intrinsic::fma:
2607 case Intrinsic::fmuladd:
2608 case Intrinsic::fabs:
2609 case Intrinsic::minnum:
2610 case Intrinsic::maxnum:
2612 // TODO: some fp intrinsics are marked as having the same error handling
2613 // as libm. They're safe to speculate when they won't error.
2614 // TODO: are convert_{from,to}_fp16 safe?
2615 // TODO: can we list target-specific intrinsics here?
2619 return false; // The called function could have undefined behavior or
2620 // side-effects, even if marked readnone nounwind.
2622 case Instruction::VAArg:
2623 case Instruction::Alloca:
2624 case Instruction::Invoke:
2625 case Instruction::PHI:
2626 case Instruction::Store:
2627 case Instruction::Ret:
2628 case Instruction::Br:
2629 case Instruction::IndirectBr:
2630 case Instruction::Switch:
2631 case Instruction::Unreachable:
2632 case Instruction::Fence:
2633 case Instruction::LandingPad:
2634 case Instruction::AtomicRMW:
2635 case Instruction::AtomicCmpXchg:
2636 case Instruction::Resume:
2637 return false; // Misc instructions which have effects
2641 /// Return true if we know that the specified value is never null.
2642 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2643 // Alloca never returns null, malloc might.
2644 if (isa<AllocaInst>(V)) return true;
2646 // A byval, inalloca, or nonnull argument is never null.
2647 if (const Argument *A = dyn_cast<Argument>(V))
2648 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2650 // Global values are not null unless extern weak.
2651 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2652 return !GV->hasExternalWeakLinkage();
2654 // A Load tagged w/nonnull metadata is never null.
2655 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2656 return LI->getMetadata(LLVMContext::MD_nonnull);
2658 if (ImmutableCallSite CS = V)
2659 if (CS.isReturnNonNull())
2662 // operator new never returns null.
2663 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))