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 /// getBitWidth - Returns the bitwidth of the given scalar or pointer type (if
43 /// unknown returns 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, returned
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))
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 if (match(I, m_Intrinsic<Intrinsic::assume>(m_Specific(V))) &&
495 isValidAssumeForContext(I, Q, DL)) {
496 assert(BitWidth == 1 && "assume operand is not i1?");
497 KnownZero.clearAllBits();
498 KnownOne.setAllBits();
503 auto m_V = m_CombineOr(m_Specific(V),
504 m_CombineOr(m_PtrToInt(m_Specific(V)),
505 m_BitCast(m_Specific(V))));
507 CmpInst::Predicate Pred;
510 if (match(I, m_Intrinsic<Intrinsic::assume>(
511 m_c_ICmp(Pred, m_V, m_Value(A)))) &&
512 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
513 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
514 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
515 KnownZero |= RHSKnownZero;
516 KnownOne |= RHSKnownOne;
518 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
519 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A)))) &&
520 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
521 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
522 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
523 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
524 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
526 // For those bits in the mask that are known to be one, we can propagate
527 // known bits from the RHS to V.
528 KnownZero |= RHSKnownZero & MaskKnownOne;
529 KnownOne |= RHSKnownOne & MaskKnownOne;
530 // assume(~(v & b) = a)
531 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
532 m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
534 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
535 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
536 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
537 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
538 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
540 // For those bits in the mask that are known to be one, we can propagate
541 // inverted known bits from the RHS to V.
542 KnownZero |= RHSKnownOne & MaskKnownOne;
543 KnownOne |= RHSKnownZero & MaskKnownOne;
545 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
546 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A)))) &&
547 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
548 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
549 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
550 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
551 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
553 // For those bits in B that are known to be zero, we can propagate known
554 // bits from the RHS to V.
555 KnownZero |= RHSKnownZero & BKnownZero;
556 KnownOne |= RHSKnownOne & BKnownZero;
557 // assume(~(v | b) = a)
558 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
559 m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
561 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
562 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
563 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
564 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
565 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
567 // For those bits in B that are known to be zero, we can propagate
568 // inverted known bits from the RHS to V.
569 KnownZero |= RHSKnownOne & BKnownZero;
570 KnownOne |= RHSKnownZero & BKnownZero;
572 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
573 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A)))) &&
574 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
575 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
576 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
577 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
578 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
580 // For those bits in B that are known to be zero, we can propagate known
581 // bits from the RHS to V. For those bits in B that are known to be one,
582 // we can propagate inverted known bits from the RHS to V.
583 KnownZero |= RHSKnownZero & BKnownZero;
584 KnownOne |= RHSKnownOne & BKnownZero;
585 KnownZero |= RHSKnownOne & BKnownOne;
586 KnownOne |= RHSKnownZero & BKnownOne;
587 // assume(~(v ^ b) = a)
588 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
589 m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
591 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
592 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
593 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
594 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
595 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
597 // For those bits in B that are known to be zero, we can propagate
598 // inverted known bits from the RHS to V. For those bits in B that are
599 // known to be one, we can propagate known bits from the RHS to V.
600 KnownZero |= RHSKnownOne & BKnownZero;
601 KnownOne |= RHSKnownZero & BKnownZero;
602 KnownZero |= RHSKnownZero & BKnownOne;
603 KnownOne |= RHSKnownOne & BKnownOne;
604 // assume(v << c = a)
605 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
606 m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
608 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
609 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
610 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
611 // For those bits in RHS that are known, we can propagate them to known
612 // bits in V shifted to the right by C.
613 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
614 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
615 // assume(~(v << c) = a)
616 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
617 m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
619 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
620 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
621 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
622 // For those bits in RHS that are known, we can propagate them inverted
623 // to known bits in V shifted to the right by C.
624 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
625 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
626 // assume(v >> c = a)
627 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
628 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
632 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
633 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
634 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
635 // For those bits in RHS that are known, we can propagate them to known
636 // bits in V shifted to the right by C.
637 KnownZero |= RHSKnownZero << C->getZExtValue();
638 KnownOne |= RHSKnownOne << C->getZExtValue();
639 // assume(~(v >> c) = a)
640 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
641 m_c_ICmp(Pred, m_Not(m_CombineOr(
642 m_LShr(m_V, m_ConstantInt(C)),
643 m_AShr(m_V, m_ConstantInt(C)))),
645 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
646 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
647 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
648 // For those bits in RHS that are known, we can propagate them inverted
649 // to known bits in V shifted to the right by C.
650 KnownZero |= RHSKnownOne << C->getZExtValue();
651 KnownOne |= RHSKnownZero << C->getZExtValue();
652 // assume(v >=_s c) where c is non-negative
653 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
654 m_ICmp(Pred, m_V, m_Value(A)))) &&
655 Pred == ICmpInst::ICMP_SGE &&
656 isValidAssumeForContext(I, Q, DL)) {
657 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
658 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
660 if (RHSKnownZero.isNegative()) {
661 // We know that the sign bit is zero.
662 KnownZero |= APInt::getSignBit(BitWidth);
664 // assume(v >_s c) where c is at least -1.
665 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
666 m_ICmp(Pred, m_V, m_Value(A)))) &&
667 Pred == ICmpInst::ICMP_SGT &&
668 isValidAssumeForContext(I, Q, DL)) {
669 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
670 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
672 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
673 // We know that the sign bit is zero.
674 KnownZero |= APInt::getSignBit(BitWidth);
676 // assume(v <=_s c) where c is negative
677 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
678 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(I, m_Intrinsic<Intrinsic::assume>(
690 m_ICmp(Pred, m_V, m_Value(A)))) &&
691 Pred == ICmpInst::ICMP_SLT &&
692 isValidAssumeForContext(I, Q, DL)) {
693 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
694 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
696 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
697 // We know that the sign bit is one.
698 KnownOne |= APInt::getSignBit(BitWidth);
701 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
702 m_ICmp(Pred, m_V, m_Value(A)))) &&
703 Pred == ICmpInst::ICMP_ULE &&
704 isValidAssumeForContext(I, Q, DL)) {
705 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
706 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
708 // Whatever high bits in c are zero are known to be zero.
710 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
712 } else if (match(I, m_Intrinsic<Intrinsic::assume>(
713 m_ICmp(Pred, m_V, m_Value(A)))) &&
714 Pred == ICmpInst::ICMP_ULT &&
715 isValidAssumeForContext(I, Q, DL)) {
716 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
717 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
719 // Whatever high bits in c are zero are known to be zero (if c is a power
720 // of 2, then one more).
721 if (isKnownToBeAPowerOfTwo(A, false, Depth+1, Query(Q, I)))
723 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
726 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
731 /// Determine which bits of V are known to be either zero or one and return
732 /// them in the KnownZero/KnownOne bit sets.
734 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
735 /// we cannot optimize based on the assumption that it is zero without changing
736 /// it to be an explicit zero. If we don't change it to zero, other code could
737 /// optimized based on the contradictory assumption that it is non-zero.
738 /// Because instcombine aggressively folds operations with undef args anyway,
739 /// this won't lose us code quality.
741 /// This function is defined on values with integer type, values with pointer
742 /// type (but only if TD is non-null), and vectors of integers. In the case
743 /// where V is a vector, known zero, and known one values are the
744 /// same width as the vector element, and the bit is set only if it is true
745 /// for all of the elements in the vector.
746 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
747 const DataLayout *TD, unsigned Depth,
749 assert(V && "No Value?");
750 assert(Depth <= MaxDepth && "Limit Search Depth");
751 unsigned BitWidth = KnownZero.getBitWidth();
753 assert((V->getType()->isIntOrIntVectorTy() ||
754 V->getType()->getScalarType()->isPointerTy()) &&
755 "Not integer or pointer type!");
757 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
758 (!V->getType()->isIntOrIntVectorTy() ||
759 V->getType()->getScalarSizeInBits() == BitWidth) &&
760 KnownZero.getBitWidth() == BitWidth &&
761 KnownOne.getBitWidth() == BitWidth &&
762 "V, KnownOne and KnownZero should have same BitWidth");
764 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
765 // We know all of the bits for a constant!
766 KnownOne = CI->getValue();
767 KnownZero = ~KnownOne;
770 // Null and aggregate-zero are all-zeros.
771 if (isa<ConstantPointerNull>(V) ||
772 isa<ConstantAggregateZero>(V)) {
773 KnownOne.clearAllBits();
774 KnownZero = APInt::getAllOnesValue(BitWidth);
777 // Handle a constant vector by taking the intersection of the known bits of
778 // each element. There is no real need to handle ConstantVector here, because
779 // we don't handle undef in any particularly useful way.
780 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
781 // We know that CDS must be a vector of integers. Take the intersection of
783 KnownZero.setAllBits(); KnownOne.setAllBits();
784 APInt Elt(KnownZero.getBitWidth(), 0);
785 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
786 Elt = CDS->getElementAsInteger(i);
793 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
794 // the bits of its aliasee.
795 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
796 if (GA->mayBeOverridden()) {
797 KnownZero.clearAllBits(); KnownOne.clearAllBits();
799 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1, Q);
804 // The address of an aligned GlobalValue has trailing zeros.
805 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
806 unsigned Align = GV->getAlignment();
807 if (Align == 0 && TD) {
808 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
809 Type *ObjectType = GVar->getType()->getElementType();
810 if (ObjectType->isSized()) {
811 // If the object is defined in the current Module, we'll be giving
812 // it the preferred alignment. Otherwise, we have to assume that it
813 // may only have the minimum ABI alignment.
814 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
815 Align = TD->getPreferredAlignment(GVar);
817 Align = TD->getABITypeAlignment(ObjectType);
822 KnownZero = APInt::getLowBitsSet(BitWidth,
823 countTrailingZeros(Align));
825 KnownZero.clearAllBits();
826 KnownOne.clearAllBits();
830 if (Argument *A = dyn_cast<Argument>(V)) {
831 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
833 if (!Align && TD && A->hasStructRetAttr()) {
834 // An sret parameter has at least the ABI alignment of the return type.
835 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
836 if (EltTy->isSized())
837 Align = TD->getABITypeAlignment(EltTy);
841 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
843 // Don't give up yet... there might be an assumption that provides more
845 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
849 // Start out not knowing anything.
850 KnownZero.clearAllBits(); KnownOne.clearAllBits();
852 if (Depth == MaxDepth)
853 return; // Limit search depth.
855 // Check whether a nearby assume intrinsic can determine some known bits.
856 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
858 Operator *I = dyn_cast<Operator>(V);
861 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
862 switch (I->getOpcode()) {
864 case Instruction::Load:
865 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
866 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
868 case Instruction::And: {
869 // If either the LHS or the RHS are Zero, the result is zero.
870 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
871 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
873 // Output known-1 bits are only known if set in both the LHS & RHS.
874 KnownOne &= KnownOne2;
875 // Output known-0 are known to be clear if zero in either the LHS | RHS.
876 KnownZero |= KnownZero2;
879 case Instruction::Or: {
880 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
881 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
883 // Output known-0 bits are only known if clear in both the LHS & RHS.
884 KnownZero &= KnownZero2;
885 // Output known-1 are known to be set if set in either the LHS | RHS.
886 KnownOne |= KnownOne2;
889 case Instruction::Xor: {
890 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
891 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
893 // Output known-0 bits are known if clear or set in both the LHS & RHS.
894 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
895 // Output known-1 are known to be set if set in only one of the LHS, RHS.
896 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
897 KnownZero = KnownZeroOut;
900 case Instruction::Mul: {
901 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
902 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
903 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
907 case Instruction::UDiv: {
908 // For the purposes of computing leading zeros we can conservatively
909 // treat a udiv as a logical right shift by the power of 2 known to
910 // be less than the denominator.
911 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
912 unsigned LeadZ = KnownZero2.countLeadingOnes();
914 KnownOne2.clearAllBits();
915 KnownZero2.clearAllBits();
916 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
917 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
918 if (RHSUnknownLeadingOnes != BitWidth)
919 LeadZ = std::min(BitWidth,
920 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
922 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
925 case Instruction::Select:
926 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1, Q);
927 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
929 // Only known if known in both the LHS and RHS.
930 KnownOne &= KnownOne2;
931 KnownZero &= KnownZero2;
933 case Instruction::FPTrunc:
934 case Instruction::FPExt:
935 case Instruction::FPToUI:
936 case Instruction::FPToSI:
937 case Instruction::SIToFP:
938 case Instruction::UIToFP:
939 break; // Can't work with floating point.
940 case Instruction::PtrToInt:
941 case Instruction::IntToPtr:
942 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
943 // We can't handle these if we don't know the pointer size.
945 // FALL THROUGH and handle them the same as zext/trunc.
946 case Instruction::ZExt:
947 case Instruction::Trunc: {
948 Type *SrcTy = I->getOperand(0)->getType();
950 unsigned SrcBitWidth;
951 // Note that we handle pointer operands here because of inttoptr/ptrtoint
952 // which fall through here.
954 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
956 SrcBitWidth = SrcTy->getScalarSizeInBits();
957 if (!SrcBitWidth) break;
960 assert(SrcBitWidth && "SrcBitWidth can't be zero");
961 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
962 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
963 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
964 KnownZero = KnownZero.zextOrTrunc(BitWidth);
965 KnownOne = KnownOne.zextOrTrunc(BitWidth);
966 // Any top bits are known to be zero.
967 if (BitWidth > SrcBitWidth)
968 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
971 case Instruction::BitCast: {
972 Type *SrcTy = I->getOperand(0)->getType();
973 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
974 // TODO: For now, not handling conversions like:
975 // (bitcast i64 %x to <2 x i32>)
976 !I->getType()->isVectorTy()) {
977 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
982 case Instruction::SExt: {
983 // Compute the bits in the result that are not present in the input.
984 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
986 KnownZero = KnownZero.trunc(SrcBitWidth);
987 KnownOne = KnownOne.trunc(SrcBitWidth);
988 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
989 KnownZero = KnownZero.zext(BitWidth);
990 KnownOne = KnownOne.zext(BitWidth);
992 // If the sign bit of the input is known set or clear, then we know the
993 // top bits of the result.
994 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
995 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
996 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
997 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1000 case Instruction::Shl:
1001 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1002 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1003 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1004 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1005 KnownZero <<= ShiftAmt;
1006 KnownOne <<= ShiftAmt;
1007 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1011 case Instruction::LShr:
1012 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1013 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1014 // Compute the new bits that are at the top now.
1015 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1017 // Unsigned shift right.
1018 computeKnownBits(I->getOperand(0), KnownZero,KnownOne, TD, Depth+1, Q);
1019 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1020 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1021 // high bits known zero.
1022 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1026 case Instruction::AShr:
1027 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1028 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1029 // Compute the new bits that are at the top now.
1030 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1032 // Signed shift right.
1033 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1034 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1035 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1037 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1038 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1039 KnownZero |= HighBits;
1040 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1041 KnownOne |= HighBits;
1045 case Instruction::Sub: {
1046 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1047 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1048 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1052 case Instruction::Add: {
1053 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1054 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1055 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1059 case Instruction::SRem:
1060 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1061 APInt RA = Rem->getValue().abs();
1062 if (RA.isPowerOf2()) {
1063 APInt LowBits = RA - 1;
1064 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD,
1067 // The low bits of the first operand are unchanged by the srem.
1068 KnownZero = KnownZero2 & LowBits;
1069 KnownOne = KnownOne2 & LowBits;
1071 // If the first operand is non-negative or has all low bits zero, then
1072 // the upper bits are all zero.
1073 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1074 KnownZero |= ~LowBits;
1076 // If the first operand is negative and not all low bits are zero, then
1077 // the upper bits are all one.
1078 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1079 KnownOne |= ~LowBits;
1081 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1085 // The sign bit is the LHS's sign bit, except when the result of the
1086 // remainder is zero.
1087 if (KnownZero.isNonNegative()) {
1088 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1089 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
1091 // If it's known zero, our sign bit is also zero.
1092 if (LHSKnownZero.isNegative())
1093 KnownZero.setBit(BitWidth - 1);
1097 case Instruction::URem: {
1098 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1099 APInt RA = Rem->getValue();
1100 if (RA.isPowerOf2()) {
1101 APInt LowBits = (RA - 1);
1102 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
1104 KnownZero |= ~LowBits;
1105 KnownOne &= LowBits;
1110 // Since the result is less than or equal to either operand, any leading
1111 // zero bits in either operand must also exist in the result.
1112 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1113 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
1115 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1116 KnownZero2.countLeadingOnes());
1117 KnownOne.clearAllBits();
1118 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1122 case Instruction::Alloca: {
1123 AllocaInst *AI = cast<AllocaInst>(V);
1124 unsigned Align = AI->getAlignment();
1125 if (Align == 0 && TD)
1126 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1129 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1132 case Instruction::GetElementPtr: {
1133 // Analyze all of the subscripts of this getelementptr instruction
1134 // to determine if we can prove known low zero bits.
1135 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1136 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
1138 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1140 gep_type_iterator GTI = gep_type_begin(I);
1141 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1142 Value *Index = I->getOperand(i);
1143 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1144 // Handle struct member offset arithmetic.
1150 // Handle case when index is vector zeroinitializer
1151 Constant *CIndex = cast<Constant>(Index);
1152 if (CIndex->isZeroValue())
1155 if (CIndex->getType()->isVectorTy())
1156 Index = CIndex->getSplatValue();
1158 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1159 const StructLayout *SL = TD->getStructLayout(STy);
1160 uint64_t Offset = SL->getElementOffset(Idx);
1161 TrailZ = std::min<unsigned>(TrailZ,
1162 countTrailingZeros(Offset));
1164 // Handle array index arithmetic.
1165 Type *IndexedTy = GTI.getIndexedType();
1166 if (!IndexedTy->isSized()) {
1170 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1171 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
1172 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1173 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1, Q);
1174 TrailZ = std::min(TrailZ,
1175 unsigned(countTrailingZeros(TypeSize) +
1176 LocalKnownZero.countTrailingOnes()));
1180 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1183 case Instruction::PHI: {
1184 PHINode *P = cast<PHINode>(I);
1185 // Handle the case of a simple two-predecessor recurrence PHI.
1186 // There's a lot more that could theoretically be done here, but
1187 // this is sufficient to catch some interesting cases.
1188 if (P->getNumIncomingValues() == 2) {
1189 for (unsigned i = 0; i != 2; ++i) {
1190 Value *L = P->getIncomingValue(i);
1191 Value *R = P->getIncomingValue(!i);
1192 Operator *LU = dyn_cast<Operator>(L);
1195 unsigned Opcode = LU->getOpcode();
1196 // Check for operations that have the property that if
1197 // both their operands have low zero bits, the result
1198 // will have low zero bits.
1199 if (Opcode == Instruction::Add ||
1200 Opcode == Instruction::Sub ||
1201 Opcode == Instruction::And ||
1202 Opcode == Instruction::Or ||
1203 Opcode == Instruction::Mul) {
1204 Value *LL = LU->getOperand(0);
1205 Value *LR = LU->getOperand(1);
1206 // Find a recurrence.
1213 // Ok, we have a PHI of the form L op= R. Check for low
1215 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1, Q);
1217 // We need to take the minimum number of known bits
1218 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1219 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1, Q);
1221 KnownZero = APInt::getLowBitsSet(BitWidth,
1222 std::min(KnownZero2.countTrailingOnes(),
1223 KnownZero3.countTrailingOnes()));
1229 // Unreachable blocks may have zero-operand PHI nodes.
1230 if (P->getNumIncomingValues() == 0)
1233 // Otherwise take the unions of the known bit sets of the operands,
1234 // taking conservative care to avoid excessive recursion.
1235 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1236 // Skip if every incoming value references to ourself.
1237 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1240 KnownZero = APInt::getAllOnesValue(BitWidth);
1241 KnownOne = APInt::getAllOnesValue(BitWidth);
1242 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1243 // Skip direct self references.
1244 if (P->getIncomingValue(i) == P) continue;
1246 KnownZero2 = APInt(BitWidth, 0);
1247 KnownOne2 = APInt(BitWidth, 0);
1248 // Recurse, but cap the recursion to one level, because we don't
1249 // want to waste time spinning around in loops.
1250 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
1252 KnownZero &= KnownZero2;
1253 KnownOne &= KnownOne2;
1254 // If all bits have been ruled out, there's no need to check
1256 if (!KnownZero && !KnownOne)
1262 case Instruction::Call:
1263 case Instruction::Invoke:
1264 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1265 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1266 // If a range metadata is attached to this IntrinsicInst, intersect the
1267 // explicit range specified by the metadata and the implicit range of
1269 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1270 switch (II->getIntrinsicID()) {
1272 case Intrinsic::ctlz:
1273 case Intrinsic::cttz: {
1274 unsigned LowBits = Log2_32(BitWidth)+1;
1275 // If this call is undefined for 0, the result will be less than 2^n.
1276 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1278 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1281 case Intrinsic::ctpop: {
1282 unsigned LowBits = Log2_32(BitWidth)+1;
1283 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1286 case Intrinsic::x86_sse42_crc32_64_64:
1287 KnownZero |= APInt::getHighBitsSet(64, 32);
1292 case Instruction::ExtractValue:
1293 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1294 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1295 if (EVI->getNumIndices() != 1) break;
1296 if (EVI->getIndices()[0] == 0) {
1297 switch (II->getIntrinsicID()) {
1299 case Intrinsic::uadd_with_overflow:
1300 case Intrinsic::sadd_with_overflow:
1301 computeKnownBitsAddSub(true, II->getArgOperand(0),
1302 II->getArgOperand(1), false, KnownZero,
1303 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1305 case Intrinsic::usub_with_overflow:
1306 case Intrinsic::ssub_with_overflow:
1307 computeKnownBitsAddSub(false, II->getArgOperand(0),
1308 II->getArgOperand(1), false, KnownZero,
1309 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1311 case Intrinsic::umul_with_overflow:
1312 case Intrinsic::smul_with_overflow:
1313 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
1314 false, KnownZero, KnownOne,
1315 KnownZero2, KnownOne2, TD, Depth, Q);
1322 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1325 /// ComputeSignBit - Determine whether the sign bit is known to be zero or
1326 /// one. Convenience wrapper around computeKnownBits.
1327 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1328 const DataLayout *TD, unsigned Depth,
1330 unsigned BitWidth = getBitWidth(V->getType(), TD);
1336 APInt ZeroBits(BitWidth, 0);
1337 APInt OneBits(BitWidth, 0);
1338 computeKnownBits(V, ZeroBits, OneBits, TD, Depth, Q);
1339 KnownOne = OneBits[BitWidth - 1];
1340 KnownZero = ZeroBits[BitWidth - 1];
1343 /// isKnownToBeAPowerOfTwo - Return true if the given value is known to have exactly one
1344 /// bit set when defined. For vectors return true if every element is known to
1345 /// be a power of two when defined. Supports values with integer or pointer
1346 /// types and vectors of integers.
1347 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1349 if (Constant *C = dyn_cast<Constant>(V)) {
1350 if (C->isNullValue())
1352 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1353 return CI->getValue().isPowerOf2();
1354 // TODO: Handle vector constants.
1357 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1358 // it is shifted off the end then the result is undefined.
1359 if (match(V, m_Shl(m_One(), m_Value())))
1362 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1363 // bottom. If it is shifted off the bottom then the result is undefined.
1364 if (match(V, m_LShr(m_SignBit(), m_Value())))
1367 // The remaining tests are all recursive, so bail out if we hit the limit.
1368 if (Depth++ == MaxDepth)
1371 Value *X = nullptr, *Y = nullptr;
1372 // A shift of a power of two is a power of two or zero.
1373 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1374 match(V, m_Shr(m_Value(X), m_Value()))))
1375 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q);
1377 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1378 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1380 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1382 isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1383 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1385 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1386 // A power of two and'd with anything is a power of two or zero.
1387 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q) ||
1388 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth, Q))
1390 // X & (-X) is always a power of two or zero.
1391 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1396 // Adding a power-of-two or zero to the same power-of-two or zero yields
1397 // either the original power-of-two, a larger power-of-two or zero.
1398 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1399 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1400 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1401 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1402 match(X, m_And(m_Value(), m_Specific(Y))))
1403 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1405 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1406 match(Y, m_And(m_Value(), m_Specific(X))))
1407 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1410 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1411 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1412 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth, Q);
1414 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1415 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth, Q);
1416 // If i8 V is a power of two or zero:
1417 // ZeroBits: 1 1 1 0 1 1 1 1
1418 // ~ZeroBits: 0 0 0 1 0 0 0 0
1419 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1420 // If OrZero isn't set, we cannot give back a zero result.
1421 // Make sure either the LHS or RHS has a bit set.
1422 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1427 // An exact divide or right shift can only shift off zero bits, so the result
1428 // is a power of two only if the first operand is a power of two and not
1429 // copying a sign bit (sdiv int_min, 2).
1430 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1431 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1432 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1439 /// \brief Test whether a GEP's result is known to be non-null.
1441 /// Uses properties inherent in a GEP to try to determine whether it is known
1444 /// Currently this routine does not support vector GEPs.
1445 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
1446 unsigned Depth, const Query &Q) {
1447 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1450 // FIXME: Support vector-GEPs.
1451 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1453 // If the base pointer is non-null, we cannot walk to a null address with an
1454 // inbounds GEP in address space zero.
1455 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1458 // Past this, if we don't have DataLayout, we can't do much.
1462 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1463 // If so, then the GEP cannot produce a null pointer, as doing so would
1464 // inherently violate the inbounds contract within address space zero.
1465 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1466 GTI != GTE; ++GTI) {
1467 // Struct types are easy -- they must always be indexed by a constant.
1468 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1469 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1470 unsigned ElementIdx = OpC->getZExtValue();
1471 const StructLayout *SL = DL->getStructLayout(STy);
1472 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1473 if (ElementOffset > 0)
1478 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1479 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
1482 // Fast path the constant operand case both for efficiency and so we don't
1483 // increment Depth when just zipping down an all-constant GEP.
1484 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1490 // We post-increment Depth here because while isKnownNonZero increments it
1491 // as well, when we pop back up that increment won't persist. We don't want
1492 // to recurse 10k times just because we have 10k GEP operands. We don't
1493 // bail completely out because we want to handle constant GEPs regardless
1495 if (Depth++ >= MaxDepth)
1498 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1505 /// isKnownNonZero - Return true if the given value is known to be non-zero
1506 /// when defined. For vectors return true if every element is known to be
1507 /// non-zero when defined. Supports values with integer or pointer type and
1508 /// vectors of integers.
1509 bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
1511 if (Constant *C = dyn_cast<Constant>(V)) {
1512 if (C->isNullValue())
1514 if (isa<ConstantInt>(C))
1515 // Must be non-zero due to null test above.
1517 // TODO: Handle vectors
1521 // The remaining tests are all recursive, so bail out if we hit the limit.
1522 if (Depth++ >= MaxDepth)
1525 // Check for pointer simplifications.
1526 if (V->getType()->isPointerTy()) {
1527 if (isKnownNonNull(V))
1529 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1530 if (isGEPKnownNonNull(GEP, TD, Depth, Q))
1534 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1536 // X | Y != 0 if X != 0 or Y != 0.
1537 Value *X = nullptr, *Y = nullptr;
1538 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1539 return isKnownNonZero(X, TD, Depth, Q) ||
1540 isKnownNonZero(Y, TD, Depth, Q);
1542 // ext X != 0 if X != 0.
1543 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1544 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth, Q);
1546 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1547 // if the lowest bit is shifted off the end.
1548 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1549 // shl nuw can't remove any non-zero bits.
1550 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1551 if (BO->hasNoUnsignedWrap())
1552 return isKnownNonZero(X, TD, Depth, Q);
1554 APInt KnownZero(BitWidth, 0);
1555 APInt KnownOne(BitWidth, 0);
1556 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1560 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1561 // defined if the sign bit is shifted off the end.
1562 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1563 // shr exact can only shift out zero bits.
1564 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1566 return isKnownNonZero(X, TD, Depth, Q);
1568 bool XKnownNonNegative, XKnownNegative;
1569 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1573 // div exact can only produce a zero if the dividend is zero.
1574 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1575 return isKnownNonZero(X, TD, Depth, Q);
1578 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1579 bool XKnownNonNegative, XKnownNegative;
1580 bool YKnownNonNegative, YKnownNegative;
1581 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1582 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth, Q);
1584 // If X and Y are both non-negative (as signed values) then their sum is not
1585 // zero unless both X and Y are zero.
1586 if (XKnownNonNegative && YKnownNonNegative)
1587 if (isKnownNonZero(X, TD, Depth, Q) ||
1588 isKnownNonZero(Y, TD, Depth, Q))
1591 // If X and Y are both negative (as signed values) then their sum is not
1592 // zero unless both X and Y equal INT_MIN.
1593 if (BitWidth && XKnownNegative && YKnownNegative) {
1594 APInt KnownZero(BitWidth, 0);
1595 APInt KnownOne(BitWidth, 0);
1596 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1597 // The sign bit of X is set. If some other bit is set then X is not equal
1599 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1600 if ((KnownOne & Mask) != 0)
1602 // The sign bit of Y is set. If some other bit is set then Y is not equal
1604 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth, Q);
1605 if ((KnownOne & Mask) != 0)
1609 // The sum of a non-negative number and a power of two is not zero.
1610 if (XKnownNonNegative &&
1611 isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth, Q))
1613 if (YKnownNonNegative &&
1614 isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth, Q))
1618 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1619 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1620 // If X and Y are non-zero then so is X * Y as long as the multiplication
1621 // does not overflow.
1622 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1623 isKnownNonZero(X, TD, Depth, Q) &&
1624 isKnownNonZero(Y, TD, Depth, Q))
1627 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1628 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1629 if (isKnownNonZero(SI->getTrueValue(), TD, Depth, Q) &&
1630 isKnownNonZero(SI->getFalseValue(), TD, Depth, Q))
1634 if (!BitWidth) return false;
1635 APInt KnownZero(BitWidth, 0);
1636 APInt KnownOne(BitWidth, 0);
1637 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1638 return KnownOne != 0;
1641 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1642 /// this predicate to simplify operations downstream. Mask is known to be zero
1643 /// for bits that V cannot have.
1645 /// This function is defined on values with integer type, values with pointer
1646 /// type (but only if TD is non-null), and vectors of integers. In the case
1647 /// where V is a vector, the mask, known zero, and known one values are the
1648 /// same width as the vector element, and the bit is set only if it is true
1649 /// for all of the elements in the vector.
1650 bool MaskedValueIsZero(Value *V, const APInt &Mask,
1651 const DataLayout *TD, unsigned Depth,
1653 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1654 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1655 return (KnownZero & Mask) == Mask;
1660 /// ComputeNumSignBits - Return the number of times the sign bit of the
1661 /// register is replicated into the other bits. We know that at least 1 bit
1662 /// is always equal to the sign bit (itself), but other cases can give us
1663 /// information. For example, immediately after an "ashr X, 2", we know that
1664 /// the top 3 bits are all equal to each other, so we return 3.
1666 /// 'Op' must have a scalar integer type.
1668 unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
1669 unsigned Depth, const Query &Q) {
1670 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1671 "ComputeNumSignBits requires a DataLayout object to operate "
1672 "on non-integer values!");
1673 Type *Ty = V->getType();
1674 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1675 Ty->getScalarSizeInBits();
1677 unsigned FirstAnswer = 1;
1679 // Note that ConstantInt is handled by the general computeKnownBits case
1683 return 1; // Limit search depth.
1685 Operator *U = dyn_cast<Operator>(V);
1686 switch (Operator::getOpcode(V)) {
1688 case Instruction::SExt:
1689 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1690 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q) + Tmp;
1692 case Instruction::AShr: {
1693 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1694 // ashr X, C -> adds C sign bits. Vectors too.
1696 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1697 Tmp += ShAmt->getZExtValue();
1698 if (Tmp > TyBits) Tmp = TyBits;
1702 case Instruction::Shl: {
1704 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1705 // shl destroys sign bits.
1706 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1707 Tmp2 = ShAmt->getZExtValue();
1708 if (Tmp2 >= TyBits || // Bad shift.
1709 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1714 case Instruction::And:
1715 case Instruction::Or:
1716 case Instruction::Xor: // NOT is handled here.
1717 // Logical binary ops preserve the number of sign bits at the worst.
1718 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1720 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1721 FirstAnswer = std::min(Tmp, Tmp2);
1722 // We computed what we know about the sign bits as our first
1723 // answer. Now proceed to the generic code that uses
1724 // computeKnownBits, and pick whichever answer is better.
1728 case Instruction::Select:
1729 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1730 if (Tmp == 1) return 1; // Early out.
1731 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1, Q);
1732 return std::min(Tmp, Tmp2);
1734 case Instruction::Add:
1735 // Add can have at most one carry bit. Thus we know that the output
1736 // is, at worst, one more bit than the inputs.
1737 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1738 if (Tmp == 1) return 1; // Early out.
1740 // Special case decrementing a value (ADD X, -1):
1741 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1742 if (CRHS->isAllOnesValue()) {
1743 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1744 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1746 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1748 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1751 // If we are subtracting one from a positive number, there is no carry
1752 // out of the result.
1753 if (KnownZero.isNegative())
1757 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1758 if (Tmp2 == 1) return 1;
1759 return std::min(Tmp, Tmp2)-1;
1761 case Instruction::Sub:
1762 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1763 if (Tmp2 == 1) return 1;
1766 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1767 if (CLHS->isNullValue()) {
1768 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1769 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
1770 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1772 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1775 // If the input is known to be positive (the sign bit is known clear),
1776 // the output of the NEG has the same number of sign bits as the input.
1777 if (KnownZero.isNegative())
1780 // Otherwise, we treat this like a SUB.
1783 // Sub can have at most one carry bit. Thus we know that the output
1784 // is, at worst, one more bit than the inputs.
1785 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1786 if (Tmp == 1) return 1; // Early out.
1787 return std::min(Tmp, Tmp2)-1;
1789 case Instruction::PHI: {
1790 PHINode *PN = cast<PHINode>(U);
1791 // Don't analyze large in-degree PHIs.
1792 if (PN->getNumIncomingValues() > 4) break;
1794 // Take the minimum of all incoming values. This can't infinitely loop
1795 // because of our depth threshold.
1796 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1, Q);
1797 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1798 if (Tmp == 1) return Tmp;
1800 ComputeNumSignBits(PN->getIncomingValue(i), TD,
1806 case Instruction::Trunc:
1807 // FIXME: it's tricky to do anything useful for this, but it is an important
1808 // case for targets like X86.
1812 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1813 // use this information.
1814 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1816 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1818 if (KnownZero.isNegative()) { // sign bit is 0
1820 } else if (KnownOne.isNegative()) { // sign bit is 1;
1827 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1828 // the number of identical bits in the top of the input value.
1830 Mask <<= Mask.getBitWidth()-TyBits;
1831 // Return # leading zeros. We use 'min' here in case Val was zero before
1832 // shifting. We don't want to return '64' as for an i32 "0".
1833 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1836 /// ComputeMultiple - This function computes the integer multiple of Base that
1837 /// equals V. If successful, it returns true and returns the multiple in
1838 /// Multiple. If unsuccessful, it returns false. It looks
1839 /// through SExt instructions only if LookThroughSExt is true.
1840 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1841 bool LookThroughSExt, unsigned Depth) {
1842 const unsigned MaxDepth = 6;
1844 assert(V && "No Value?");
1845 assert(Depth <= MaxDepth && "Limit Search Depth");
1846 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1848 Type *T = V->getType();
1850 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1860 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1861 Constant *BaseVal = ConstantInt::get(T, Base);
1862 if (CO && CO == BaseVal) {
1864 Multiple = ConstantInt::get(T, 1);
1868 if (CI && CI->getZExtValue() % Base == 0) {
1869 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1873 if (Depth == MaxDepth) return false; // Limit search depth.
1875 Operator *I = dyn_cast<Operator>(V);
1876 if (!I) return false;
1878 switch (I->getOpcode()) {
1880 case Instruction::SExt:
1881 if (!LookThroughSExt) return false;
1882 // otherwise fall through to ZExt
1883 case Instruction::ZExt:
1884 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1885 LookThroughSExt, Depth+1);
1886 case Instruction::Shl:
1887 case Instruction::Mul: {
1888 Value *Op0 = I->getOperand(0);
1889 Value *Op1 = I->getOperand(1);
1891 if (I->getOpcode() == Instruction::Shl) {
1892 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1893 if (!Op1CI) return false;
1894 // Turn Op0 << Op1 into Op0 * 2^Op1
1895 APInt Op1Int = Op1CI->getValue();
1896 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1897 APInt API(Op1Int.getBitWidth(), 0);
1898 API.setBit(BitToSet);
1899 Op1 = ConstantInt::get(V->getContext(), API);
1902 Value *Mul0 = nullptr;
1903 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1904 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1905 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1906 if (Op1C->getType()->getPrimitiveSizeInBits() <
1907 MulC->getType()->getPrimitiveSizeInBits())
1908 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1909 if (Op1C->getType()->getPrimitiveSizeInBits() >
1910 MulC->getType()->getPrimitiveSizeInBits())
1911 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1913 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1914 Multiple = ConstantExpr::getMul(MulC, Op1C);
1918 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1919 if (Mul0CI->getValue() == 1) {
1920 // V == Base * Op1, so return Op1
1926 Value *Mul1 = nullptr;
1927 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1928 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1929 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1930 if (Op0C->getType()->getPrimitiveSizeInBits() <
1931 MulC->getType()->getPrimitiveSizeInBits())
1932 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1933 if (Op0C->getType()->getPrimitiveSizeInBits() >
1934 MulC->getType()->getPrimitiveSizeInBits())
1935 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1937 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1938 Multiple = ConstantExpr::getMul(MulC, Op0C);
1942 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1943 if (Mul1CI->getValue() == 1) {
1944 // V == Base * Op0, so return Op0
1952 // We could not determine if V is a multiple of Base.
1956 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1957 /// value is never equal to -0.0.
1959 /// NOTE: this function will need to be revisited when we support non-default
1962 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1963 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1964 return !CFP->getValueAPF().isNegZero();
1967 return 1; // Limit search depth.
1969 const Operator *I = dyn_cast<Operator>(V);
1970 if (!I) return false;
1972 // Check if the nsz fast-math flag is set
1973 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
1974 if (FPO->hasNoSignedZeros())
1977 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1978 if (I->getOpcode() == Instruction::FAdd)
1979 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
1980 if (CFP->isNullValue())
1983 // sitofp and uitofp turn into +0.0 for zero.
1984 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
1987 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1988 // sqrt(-0.0) = -0.0, no other negative results are possible.
1989 if (II->getIntrinsicID() == Intrinsic::sqrt)
1990 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
1992 if (const CallInst *CI = dyn_cast<CallInst>(I))
1993 if (const Function *F = CI->getCalledFunction()) {
1994 if (F->isDeclaration()) {
1996 if (F->getName() == "abs") return true;
1997 // fabs[lf](x) != -0.0
1998 if (F->getName() == "fabs") return true;
1999 if (F->getName() == "fabsf") return true;
2000 if (F->getName() == "fabsl") return true;
2001 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2002 F->getName() == "sqrtl")
2003 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2010 /// isBytewiseValue - If the specified value can be set by repeating the same
2011 /// byte in memory, return the i8 value that it is represented with. This is
2012 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2013 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2014 /// byte store (e.g. i16 0x1234), return null.
2015 Value *llvm::isBytewiseValue(Value *V) {
2016 // All byte-wide stores are splatable, even of arbitrary variables.
2017 if (V->getType()->isIntegerTy(8)) return V;
2019 // Handle 'null' ConstantArrayZero etc.
2020 if (Constant *C = dyn_cast<Constant>(V))
2021 if (C->isNullValue())
2022 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2024 // Constant float and double values can be handled as integer values if the
2025 // corresponding integer value is "byteable". An important case is 0.0.
2026 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2027 if (CFP->getType()->isFloatTy())
2028 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2029 if (CFP->getType()->isDoubleTy())
2030 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2031 // Don't handle long double formats, which have strange constraints.
2034 // We can handle constant integers that are power of two in size and a
2035 // multiple of 8 bits.
2036 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2037 unsigned Width = CI->getBitWidth();
2038 if (isPowerOf2_32(Width) && Width > 8) {
2039 // We can handle this value if the recursive binary decomposition is the
2040 // same at all levels.
2041 APInt Val = CI->getValue();
2043 while (Val.getBitWidth() != 8) {
2044 unsigned NextWidth = Val.getBitWidth()/2;
2045 Val2 = Val.lshr(NextWidth);
2046 Val2 = Val2.trunc(Val.getBitWidth()/2);
2047 Val = Val.trunc(Val.getBitWidth()/2);
2049 // If the top/bottom halves aren't the same, reject it.
2053 return ConstantInt::get(V->getContext(), Val);
2057 // A ConstantDataArray/Vector is splatable if all its members are equal and
2059 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2060 Value *Elt = CA->getElementAsConstant(0);
2061 Value *Val = isBytewiseValue(Elt);
2065 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2066 if (CA->getElementAsConstant(I) != Elt)
2072 // Conceptually, we could handle things like:
2073 // %a = zext i8 %X to i16
2074 // %b = shl i16 %a, 8
2075 // %c = or i16 %a, %b
2076 // but until there is an example that actually needs this, it doesn't seem
2077 // worth worrying about.
2082 // This is the recursive version of BuildSubAggregate. It takes a few different
2083 // arguments. Idxs is the index within the nested struct From that we are
2084 // looking at now (which is of type IndexedType). IdxSkip is the number of
2085 // indices from Idxs that should be left out when inserting into the resulting
2086 // struct. To is the result struct built so far, new insertvalue instructions
2088 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2089 SmallVectorImpl<unsigned> &Idxs,
2091 Instruction *InsertBefore) {
2092 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2094 // Save the original To argument so we can modify it
2096 // General case, the type indexed by Idxs is a struct
2097 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2098 // Process each struct element recursively
2101 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2105 // Couldn't find any inserted value for this index? Cleanup
2106 while (PrevTo != OrigTo) {
2107 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2108 PrevTo = Del->getAggregateOperand();
2109 Del->eraseFromParent();
2111 // Stop processing elements
2115 // If we successfully found a value for each of our subaggregates
2119 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2120 // the struct's elements had a value that was inserted directly. In the latter
2121 // case, perhaps we can't determine each of the subelements individually, but
2122 // we might be able to find the complete struct somewhere.
2124 // Find the value that is at that particular spot
2125 Value *V = FindInsertedValue(From, Idxs);
2130 // Insert the value in the new (sub) aggregrate
2131 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2132 "tmp", InsertBefore);
2135 // This helper takes a nested struct and extracts a part of it (which is again a
2136 // struct) into a new value. For example, given the struct:
2137 // { a, { b, { c, d }, e } }
2138 // and the indices "1, 1" this returns
2141 // It does this by inserting an insertvalue for each element in the resulting
2142 // struct, as opposed to just inserting a single struct. This will only work if
2143 // each of the elements of the substruct are known (ie, inserted into From by an
2144 // insertvalue instruction somewhere).
2146 // All inserted insertvalue instructions are inserted before InsertBefore
2147 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2148 Instruction *InsertBefore) {
2149 assert(InsertBefore && "Must have someplace to insert!");
2150 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2152 Value *To = UndefValue::get(IndexedType);
2153 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2154 unsigned IdxSkip = Idxs.size();
2156 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2159 /// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
2160 /// the scalar value indexed is already around as a register, for example if it
2161 /// were inserted directly into the aggregrate.
2163 /// If InsertBefore is not null, this function will duplicate (modified)
2164 /// insertvalues when a part of a nested struct is extracted.
2165 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2166 Instruction *InsertBefore) {
2167 // Nothing to index? Just return V then (this is useful at the end of our
2169 if (idx_range.empty())
2171 // We have indices, so V should have an indexable type.
2172 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2173 "Not looking at a struct or array?");
2174 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2175 "Invalid indices for type?");
2177 if (Constant *C = dyn_cast<Constant>(V)) {
2178 C = C->getAggregateElement(idx_range[0]);
2179 if (!C) return nullptr;
2180 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2183 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2184 // Loop the indices for the insertvalue instruction in parallel with the
2185 // requested indices
2186 const unsigned *req_idx = idx_range.begin();
2187 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2188 i != e; ++i, ++req_idx) {
2189 if (req_idx == idx_range.end()) {
2190 // We can't handle this without inserting insertvalues
2194 // The requested index identifies a part of a nested aggregate. Handle
2195 // this specially. For example,
2196 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2197 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2198 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2199 // This can be changed into
2200 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2201 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2202 // which allows the unused 0,0 element from the nested struct to be
2204 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2208 // This insert value inserts something else than what we are looking for.
2209 // See if the (aggregrate) value inserted into has the value we are
2210 // looking for, then.
2212 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2215 // If we end up here, the indices of the insertvalue match with those
2216 // requested (though possibly only partially). Now we recursively look at
2217 // the inserted value, passing any remaining indices.
2218 return FindInsertedValue(I->getInsertedValueOperand(),
2219 makeArrayRef(req_idx, idx_range.end()),
2223 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2224 // If we're extracting a value from an aggregrate that was extracted from
2225 // something else, we can extract from that something else directly instead.
2226 // However, we will need to chain I's indices with the requested indices.
2228 // Calculate the number of indices required
2229 unsigned size = I->getNumIndices() + idx_range.size();
2230 // Allocate some space to put the new indices in
2231 SmallVector<unsigned, 5> Idxs;
2233 // Add indices from the extract value instruction
2234 Idxs.append(I->idx_begin(), I->idx_end());
2236 // Add requested indices
2237 Idxs.append(idx_range.begin(), idx_range.end());
2239 assert(Idxs.size() == size
2240 && "Number of indices added not correct?");
2242 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2244 // Otherwise, we don't know (such as, extracting from a function return value
2245 // or load instruction)
2249 /// GetPointerBaseWithConstantOffset - Analyze the specified pointer to see if
2250 /// it can be expressed as a base pointer plus a constant offset. Return the
2251 /// base and offset to the caller.
2252 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2253 const DataLayout *DL) {
2254 // Without DataLayout, conservatively assume 64-bit offsets, which is
2255 // the widest we support.
2256 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
2257 APInt ByteOffset(BitWidth, 0);
2259 if (Ptr->getType()->isVectorTy())
2262 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2264 APInt GEPOffset(BitWidth, 0);
2265 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
2268 ByteOffset += GEPOffset;
2271 Ptr = GEP->getPointerOperand();
2272 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2273 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2274 Ptr = cast<Operator>(Ptr)->getOperand(0);
2275 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2276 if (GA->mayBeOverridden())
2278 Ptr = GA->getAliasee();
2283 Offset = ByteOffset.getSExtValue();
2288 /// getConstantStringInfo - This function computes the length of a
2289 /// null-terminated C string pointed to by V. If successful, it returns true
2290 /// and returns the string in Str. If unsuccessful, it returns false.
2291 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2292 uint64_t Offset, bool TrimAtNul) {
2295 // Look through bitcast instructions and geps.
2296 V = V->stripPointerCasts();
2298 // If the value is a GEP instructionor constant expression, treat it as an
2300 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2301 // Make sure the GEP has exactly three arguments.
2302 if (GEP->getNumOperands() != 3)
2305 // Make sure the index-ee is a pointer to array of i8.
2306 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2307 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2308 if (!AT || !AT->getElementType()->isIntegerTy(8))
2311 // Check to make sure that the first operand of the GEP is an integer and
2312 // has value 0 so that we are sure we're indexing into the initializer.
2313 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2314 if (!FirstIdx || !FirstIdx->isZero())
2317 // If the second index isn't a ConstantInt, then this is a variable index
2318 // into the array. If this occurs, we can't say anything meaningful about
2320 uint64_t StartIdx = 0;
2321 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2322 StartIdx = CI->getZExtValue();
2325 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
2328 // The GEP instruction, constant or instruction, must reference a global
2329 // variable that is a constant and is initialized. The referenced constant
2330 // initializer is the array that we'll use for optimization.
2331 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2332 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2335 // Handle the all-zeros case
2336 if (GV->getInitializer()->isNullValue()) {
2337 // This is a degenerate case. The initializer is constant zero so the
2338 // length of the string must be zero.
2343 // Must be a Constant Array
2344 const ConstantDataArray *Array =
2345 dyn_cast<ConstantDataArray>(GV->getInitializer());
2346 if (!Array || !Array->isString())
2349 // Get the number of elements in the array
2350 uint64_t NumElts = Array->getType()->getArrayNumElements();
2352 // Start out with the entire array in the StringRef.
2353 Str = Array->getAsString();
2355 if (Offset > NumElts)
2358 // Skip over 'offset' bytes.
2359 Str = Str.substr(Offset);
2362 // Trim off the \0 and anything after it. If the array is not nul
2363 // terminated, we just return the whole end of string. The client may know
2364 // some other way that the string is length-bound.
2365 Str = Str.substr(0, Str.find('\0'));
2370 // These next two are very similar to the above, but also look through PHI
2372 // TODO: See if we can integrate these two together.
2374 /// GetStringLengthH - If we can compute the length of the string pointed to by
2375 /// the specified pointer, return 'len+1'. If we can't, return 0.
2376 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2377 // Look through noop bitcast instructions.
2378 V = V->stripPointerCasts();
2380 // If this is a PHI node, there are two cases: either we have already seen it
2382 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2383 if (!PHIs.insert(PN))
2384 return ~0ULL; // already in the set.
2386 // If it was new, see if all the input strings are the same length.
2387 uint64_t LenSoFar = ~0ULL;
2388 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2389 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2390 if (Len == 0) return 0; // Unknown length -> unknown.
2392 if (Len == ~0ULL) continue;
2394 if (Len != LenSoFar && LenSoFar != ~0ULL)
2395 return 0; // Disagree -> unknown.
2399 // Success, all agree.
2403 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2404 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2405 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2406 if (Len1 == 0) return 0;
2407 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2408 if (Len2 == 0) return 0;
2409 if (Len1 == ~0ULL) return Len2;
2410 if (Len2 == ~0ULL) return Len1;
2411 if (Len1 != Len2) return 0;
2415 // Otherwise, see if we can read the string.
2417 if (!getConstantStringInfo(V, StrData))
2420 return StrData.size()+1;
2423 /// GetStringLength - If we can compute the length of the string pointed to by
2424 /// the specified pointer, return 'len+1'. If we can't, return 0.
2425 uint64_t llvm::GetStringLength(Value *V) {
2426 if (!V->getType()->isPointerTy()) return 0;
2428 SmallPtrSet<PHINode*, 32> PHIs;
2429 uint64_t Len = GetStringLengthH(V, PHIs);
2430 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2431 // an empty string as a length.
2432 return Len == ~0ULL ? 1 : Len;
2436 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
2437 if (!V->getType()->isPointerTy())
2439 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2440 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2441 V = GEP->getPointerOperand();
2442 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2443 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2444 V = cast<Operator>(V)->getOperand(0);
2445 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2446 if (GA->mayBeOverridden())
2448 V = GA->getAliasee();
2450 // See if InstructionSimplify knows any relevant tricks.
2451 if (Instruction *I = dyn_cast<Instruction>(V))
2452 // TODO: Acquire a DominatorTree and AssumptionTracker and use them.
2453 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
2460 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2466 llvm::GetUnderlyingObjects(Value *V,
2467 SmallVectorImpl<Value *> &Objects,
2468 const DataLayout *TD,
2469 unsigned MaxLookup) {
2470 SmallPtrSet<Value *, 4> Visited;
2471 SmallVector<Value *, 4> Worklist;
2472 Worklist.push_back(V);
2474 Value *P = Worklist.pop_back_val();
2475 P = GetUnderlyingObject(P, TD, MaxLookup);
2477 if (!Visited.insert(P))
2480 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2481 Worklist.push_back(SI->getTrueValue());
2482 Worklist.push_back(SI->getFalseValue());
2486 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2487 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2488 Worklist.push_back(PN->getIncomingValue(i));
2492 Objects.push_back(P);
2493 } while (!Worklist.empty());
2496 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
2497 /// are lifetime markers.
2499 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2500 for (const User *U : V->users()) {
2501 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2502 if (!II) return false;
2504 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2505 II->getIntrinsicID() != Intrinsic::lifetime_end)
2511 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
2512 const DataLayout *TD) {
2513 const Operator *Inst = dyn_cast<Operator>(V);
2517 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2518 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2522 switch (Inst->getOpcode()) {
2525 case Instruction::UDiv:
2526 case Instruction::URem:
2527 // x / y is undefined if y == 0, but calculations like x / 3 are safe.
2528 return isKnownNonZero(Inst->getOperand(1), TD);
2529 case Instruction::SDiv:
2530 case Instruction::SRem: {
2531 Value *Op = Inst->getOperand(1);
2532 // x / y is undefined if y == 0
2533 if (!isKnownNonZero(Op, TD))
2535 // x / y might be undefined if y == -1
2536 unsigned BitWidth = getBitWidth(Op->getType(), TD);
2539 APInt KnownZero(BitWidth, 0);
2540 APInt KnownOne(BitWidth, 0);
2541 computeKnownBits(Op, KnownZero, KnownOne, TD);
2544 case Instruction::Load: {
2545 const LoadInst *LI = cast<LoadInst>(Inst);
2546 if (!LI->isUnordered() ||
2547 // Speculative load may create a race that did not exist in the source.
2548 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2550 return LI->getPointerOperand()->isDereferenceablePointer(TD);
2552 case Instruction::Call: {
2553 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2554 switch (II->getIntrinsicID()) {
2555 // These synthetic intrinsics have no side-effects and just mark
2556 // information about their operands.
2557 // FIXME: There are other no-op synthetic instructions that potentially
2558 // should be considered at least *safe* to speculate...
2559 case Intrinsic::dbg_declare:
2560 case Intrinsic::dbg_value:
2563 case Intrinsic::bswap:
2564 case Intrinsic::ctlz:
2565 case Intrinsic::ctpop:
2566 case Intrinsic::cttz:
2567 case Intrinsic::objectsize:
2568 case Intrinsic::sadd_with_overflow:
2569 case Intrinsic::smul_with_overflow:
2570 case Intrinsic::ssub_with_overflow:
2571 case Intrinsic::uadd_with_overflow:
2572 case Intrinsic::umul_with_overflow:
2573 case Intrinsic::usub_with_overflow:
2575 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2576 // errno like libm sqrt would.
2577 case Intrinsic::sqrt:
2578 case Intrinsic::fma:
2579 case Intrinsic::fmuladd:
2580 case Intrinsic::fabs:
2582 // TODO: some fp intrinsics are marked as having the same error handling
2583 // as libm. They're safe to speculate when they won't error.
2584 // TODO: are convert_{from,to}_fp16 safe?
2585 // TODO: can we list target-specific intrinsics here?
2589 return false; // The called function could have undefined behavior or
2590 // side-effects, even if marked readnone nounwind.
2592 case Instruction::VAArg:
2593 case Instruction::Alloca:
2594 case Instruction::Invoke:
2595 case Instruction::PHI:
2596 case Instruction::Store:
2597 case Instruction::Ret:
2598 case Instruction::Br:
2599 case Instruction::IndirectBr:
2600 case Instruction::Switch:
2601 case Instruction::Unreachable:
2602 case Instruction::Fence:
2603 case Instruction::LandingPad:
2604 case Instruction::AtomicRMW:
2605 case Instruction::AtomicCmpXchg:
2606 case Instruction::Resume:
2607 return false; // Misc instructions which have effects
2611 /// isKnownNonNull - Return true if we know that the specified value is never
2613 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2614 // Alloca never returns null, malloc might.
2615 if (isa<AllocaInst>(V)) return true;
2617 // A byval, inalloca, or nonnull argument is never null.
2618 if (const Argument *A = dyn_cast<Argument>(V))
2619 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2621 // Global values are not null unless extern weak.
2622 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2623 return !GV->hasExternalWeakLinkage();
2625 // A Load tagged w/nonnull metadata is never null.
2626 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2627 return LI->getMetadata(LLVMContext::MD_nonnull);
2629 if (ImmutableCallSite CS = V)
2630 if (CS.isReturnNonNull())
2633 // operator new never returns null.
2634 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))