1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains routines that help analyze properties that chains of
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/Analysis/AssumptionTracker.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/MemoryBuiltins.h"
20 #include "llvm/IR/CallSite.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/MathExtras.h"
38 using namespace llvm::PatternMatch;
40 const unsigned MaxDepth = 6;
42 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
43 /// 0). For vector types, returns the element type's bitwidth.
44 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
45 if (unsigned BitWidth = Ty->getScalarSizeInBits())
48 return TD ? TD->getPointerTypeSizeInBits(Ty) : 0;
51 // Many of these functions have internal versions that take an assumption
52 // exclusion set. This is because of the potential for mutual recursion to
53 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
54 // classic case of this is assume(x = y), which will attempt to determine
55 // bits in x from bits in y, which will attempt to determine bits in y from
56 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
57 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
58 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
59 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
62 // Simplifying using an assume can only be done in a particular control-flow
63 // context (the context instruction provides that context). If an assume and
64 // the context instruction are not in the same block then the DT helps in
65 // figuring out if we can use it.
68 AssumptionTracker *AT;
69 const Instruction *CxtI;
70 const DominatorTree *DT;
72 Query(AssumptionTracker *AT = nullptr, const Instruction *CxtI = nullptr,
73 const DominatorTree *DT = nullptr)
74 : AT(AT), CxtI(CxtI), DT(DT) {}
76 Query(const Query &Q, const Value *NewExcl)
77 : ExclInvs(Q.ExclInvs), AT(Q.AT), CxtI(Q.CxtI), DT(Q.DT) {
78 ExclInvs.insert(NewExcl);
81 } // end anonymous namespace
83 // Given the provided Value and, potentially, a context instruction, return
84 // the preferred context instruction (if any).
85 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
86 // If we've been provided with a context instruction, then use that (provided
87 // it has been inserted).
88 if (CxtI && CxtI->getParent())
91 // If the value is really an already-inserted instruction, then use that.
92 CxtI = dyn_cast<Instruction>(V);
93 if (CxtI && CxtI->getParent())
99 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
100 const DataLayout *TD, unsigned Depth,
103 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
104 const DataLayout *TD, unsigned Depth,
105 AssumptionTracker *AT, const Instruction *CxtI,
106 const DominatorTree *DT) {
107 ::computeKnownBits(V, KnownZero, KnownOne, TD, Depth,
108 Query(AT, safeCxtI(V, CxtI), DT));
111 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
112 const DataLayout *TD, unsigned Depth,
115 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
116 const DataLayout *TD, unsigned Depth,
117 AssumptionTracker *AT, const Instruction *CxtI,
118 const DominatorTree *DT) {
119 ::ComputeSignBit(V, KnownZero, KnownOne, TD, Depth,
120 Query(AT, safeCxtI(V, CxtI), DT));
123 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
126 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
127 AssumptionTracker *AT,
128 const Instruction *CxtI,
129 const DominatorTree *DT) {
130 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
131 Query(AT, safeCxtI(V, CxtI), DT));
134 static bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
137 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
138 AssumptionTracker *AT, const Instruction *CxtI,
139 const DominatorTree *DT) {
140 return ::isKnownNonZero(V, TD, Depth, Query(AT, safeCxtI(V, CxtI), DT));
143 static bool MaskedValueIsZero(Value *V, const APInt &Mask,
144 const DataLayout *TD, unsigned Depth,
147 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
148 const DataLayout *TD, unsigned Depth,
149 AssumptionTracker *AT, const Instruction *CxtI,
150 const DominatorTree *DT) {
151 return ::MaskedValueIsZero(V, Mask, TD, Depth,
152 Query(AT, safeCxtI(V, CxtI), DT));
155 static unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
156 unsigned Depth, const Query &Q);
158 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
159 unsigned Depth, AssumptionTracker *AT,
160 const Instruction *CxtI,
161 const DominatorTree *DT) {
162 return ::ComputeNumSignBits(V, TD, Depth, Query(AT, safeCxtI(V, CxtI), DT));
165 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
166 APInt &KnownZero, APInt &KnownOne,
167 APInt &KnownZero2, APInt &KnownOne2,
168 const DataLayout *TD, unsigned Depth,
171 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
172 // We know that the top bits of C-X are clear if X contains less bits
173 // than C (i.e. no wrap-around can happen). For example, 20-X is
174 // positive if we can prove that X is >= 0 and < 16.
175 if (!CLHS->getValue().isNegative()) {
176 unsigned BitWidth = KnownZero.getBitWidth();
177 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
178 // NLZ can't be BitWidth with no sign bit
179 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
180 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
182 // If all of the MaskV bits are known to be zero, then we know the
183 // output top bits are zero, because we now know that the output is
185 if ((KnownZero2 & MaskV) == MaskV) {
186 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
187 // Top bits known zero.
188 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
194 unsigned BitWidth = KnownZero.getBitWidth();
196 // If an initial sequence of bits in the result is not needed, the
197 // corresponding bits in the operands are not needed.
198 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
199 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1, Q);
200 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
202 // Carry in a 1 for a subtract, rather than a 0.
203 APInt CarryIn(BitWidth, 0);
205 // Sum = LHS + ~RHS + 1
206 std::swap(KnownZero2, KnownOne2);
210 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
211 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
213 // Compute known bits of the carry.
214 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
215 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
217 // Compute set of known bits (where all three relevant bits are known).
218 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
219 APInt RHSKnown = KnownZero2 | KnownOne2;
220 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
221 APInt Known = LHSKnown & RHSKnown & CarryKnown;
223 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
224 "known bits of sum differ");
226 // Compute known bits of the result.
227 KnownZero = ~PossibleSumOne & Known;
228 KnownOne = PossibleSumOne & Known;
230 // Are we still trying to solve for the sign bit?
231 if (!Known.isNegative()) {
233 // Adding two non-negative numbers, or subtracting a negative number from
234 // a non-negative one, can't wrap into negative.
235 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
236 KnownZero |= APInt::getSignBit(BitWidth);
237 // Adding two negative numbers, or subtracting a non-negative number from
238 // a negative one, can't wrap into non-negative.
239 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
240 KnownOne |= APInt::getSignBit(BitWidth);
245 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
246 APInt &KnownZero, APInt &KnownOne,
247 APInt &KnownZero2, APInt &KnownOne2,
248 const DataLayout *TD, unsigned Depth,
250 unsigned BitWidth = KnownZero.getBitWidth();
251 computeKnownBits(Op1, KnownZero, KnownOne, TD, Depth+1, Q);
252 computeKnownBits(Op0, KnownZero2, KnownOne2, TD, Depth+1, Q);
254 bool isKnownNegative = false;
255 bool isKnownNonNegative = false;
256 // If the multiplication is known not to overflow, compute the sign bit.
259 // The product of a number with itself is non-negative.
260 isKnownNonNegative = true;
262 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
263 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
264 bool isKnownNegativeOp1 = KnownOne.isNegative();
265 bool isKnownNegativeOp0 = KnownOne2.isNegative();
266 // The product of two numbers with the same sign is non-negative.
267 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
268 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
269 // The product of a negative number and a non-negative number is either
271 if (!isKnownNonNegative)
272 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
273 isKnownNonZero(Op0, TD, Depth, Q)) ||
274 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
275 isKnownNonZero(Op1, TD, Depth, Q));
279 // If low bits are zero in either operand, output low known-0 bits.
280 // Also compute a conserative estimate for high known-0 bits.
281 // More trickiness is possible, but this is sufficient for the
282 // interesting case of alignment computation.
283 KnownOne.clearAllBits();
284 unsigned TrailZ = KnownZero.countTrailingOnes() +
285 KnownZero2.countTrailingOnes();
286 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
287 KnownZero2.countLeadingOnes(),
288 BitWidth) - BitWidth;
290 TrailZ = std::min(TrailZ, BitWidth);
291 LeadZ = std::min(LeadZ, BitWidth);
292 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
293 APInt::getHighBitsSet(BitWidth, LeadZ);
295 // Only make use of no-wrap flags if we failed to compute the sign bit
296 // directly. This matters if the multiplication always overflows, in
297 // which case we prefer to follow the result of the direct computation,
298 // though as the program is invoking undefined behaviour we can choose
299 // whatever we like here.
300 if (isKnownNonNegative && !KnownOne.isNegative())
301 KnownZero.setBit(BitWidth - 1);
302 else if (isKnownNegative && !KnownZero.isNegative())
303 KnownOne.setBit(BitWidth - 1);
306 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
308 unsigned BitWidth = KnownZero.getBitWidth();
309 unsigned NumRanges = Ranges.getNumOperands() / 2;
310 assert(NumRanges >= 1);
312 // Use the high end of the ranges to find leading zeros.
313 unsigned MinLeadingZeros = BitWidth;
314 for (unsigned i = 0; i < NumRanges; ++i) {
316 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
318 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
319 ConstantRange Range(Lower->getValue(), Upper->getValue());
320 if (Range.isWrappedSet())
321 MinLeadingZeros = 0; // -1 has no zeros
322 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
323 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
326 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
329 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
330 SmallVector<const Value *, 16> WorkSet(1, I);
331 SmallPtrSet<const Value *, 32> Visited;
332 SmallPtrSet<const Value *, 16> EphValues;
334 while (!WorkSet.empty()) {
335 const Value *V = WorkSet.pop_back_val();
336 if (!Visited.insert(V).second)
339 // If all uses of this value are ephemeral, then so is this value.
340 bool FoundNEUse = false;
341 for (const User *I : V->users())
342 if (!EphValues.count(I)) {
352 if (const User *U = dyn_cast<User>(V))
353 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
355 if (isSafeToSpeculativelyExecute(*J))
356 WorkSet.push_back(*J);
364 // Is this an intrinsic that cannot be speculated but also cannot trap?
365 static bool isAssumeLikeIntrinsic(const Instruction *I) {
366 if (const CallInst *CI = dyn_cast<CallInst>(I))
367 if (Function *F = CI->getCalledFunction())
368 switch (F->getIntrinsicID()) {
370 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
371 case Intrinsic::assume:
372 case Intrinsic::dbg_declare:
373 case Intrinsic::dbg_value:
374 case Intrinsic::invariant_start:
375 case Intrinsic::invariant_end:
376 case Intrinsic::lifetime_start:
377 case Intrinsic::lifetime_end:
378 case Intrinsic::objectsize:
379 case Intrinsic::ptr_annotation:
380 case Intrinsic::var_annotation:
387 static bool isValidAssumeForContext(Value *V, const Query &Q,
388 const DataLayout *DL) {
389 Instruction *Inv = cast<Instruction>(V);
391 // There are two restrictions on the use of an assume:
392 // 1. The assume must dominate the context (or the control flow must
393 // reach the assume whenever it reaches the context).
394 // 2. The context must not be in the assume's set of ephemeral values
395 // (otherwise we will use the assume to prove that the condition
396 // feeding the assume is trivially true, thus causing the removal of
400 if (Q.DT->dominates(Inv, Q.CxtI)) {
402 } else if (Inv->getParent() == Q.CxtI->getParent()) {
403 // The context comes first, but they're both in the same block. Make sure
404 // there is nothing in between that might interrupt the control flow.
405 for (BasicBlock::const_iterator I =
406 std::next(BasicBlock::const_iterator(Q.CxtI)),
407 IE(Inv); I != IE; ++I)
408 if (!isSafeToSpeculativelyExecute(I, DL) &&
409 !isAssumeLikeIntrinsic(I))
412 return !isEphemeralValueOf(Inv, Q.CxtI);
418 // When we don't have a DT, we do a limited search...
419 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
421 } else if (Inv->getParent() == Q.CxtI->getParent()) {
422 // Search forward from the assume until we reach the context (or the end
423 // of the block); the common case is that the assume will come first.
424 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
425 IE = Inv->getParent()->end(); I != IE; ++I)
429 // The context must come first...
430 for (BasicBlock::const_iterator I =
431 std::next(BasicBlock::const_iterator(Q.CxtI)),
432 IE(Inv); I != IE; ++I)
433 if (!isSafeToSpeculativelyExecute(I, DL) &&
434 !isAssumeLikeIntrinsic(I))
437 return !isEphemeralValueOf(Inv, Q.CxtI);
443 bool llvm::isValidAssumeForContext(const Instruction *I,
444 const Instruction *CxtI,
445 const DataLayout *DL,
446 const DominatorTree *DT) {
447 return ::isValidAssumeForContext(const_cast<Instruction*>(I),
448 Query(nullptr, CxtI, DT), DL);
451 template<typename LHS, typename RHS>
452 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
453 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
454 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
455 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
458 template<typename LHS, typename RHS>
459 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
460 BinaryOp_match<RHS, LHS, Instruction::And>>
461 m_c_And(const LHS &L, const RHS &R) {
462 return m_CombineOr(m_And(L, R), m_And(R, L));
465 template<typename LHS, typename RHS>
466 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
467 BinaryOp_match<RHS, LHS, Instruction::Or>>
468 m_c_Or(const LHS &L, const RHS &R) {
469 return m_CombineOr(m_Or(L, R), m_Or(R, L));
472 template<typename LHS, typename RHS>
473 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
474 BinaryOp_match<RHS, LHS, Instruction::Xor>>
475 m_c_Xor(const LHS &L, const RHS &R) {
476 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
479 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
481 const DataLayout *DL,
482 unsigned Depth, const Query &Q) {
483 // Use of assumptions is context-sensitive. If we don't have a context, we
485 if (!Q.AT || !Q.CxtI)
488 unsigned BitWidth = KnownZero.getBitWidth();
490 Function *F = const_cast<Function*>(Q.CxtI->getParent()->getParent());
491 for (auto &CI : Q.AT->assumptions(F)) {
493 if (Q.ExclInvs.count(I))
496 // Warning: This loop can end up being somewhat performance sensetive.
497 // We're running this loop for once for each value queried resulting in a
498 // runtime of ~O(#assumes * #values).
500 assert(isa<IntrinsicInst>(I) &&
501 dyn_cast<IntrinsicInst>(I)->getIntrinsicID() == Intrinsic::assume &&
502 "must be an assume intrinsic");
504 Value *Arg = I->getArgOperand(0);
507 isValidAssumeForContext(I, Q, DL)) {
508 assert(BitWidth == 1 && "assume operand is not i1?");
509 KnownZero.clearAllBits();
510 KnownOne.setAllBits();
514 // The remaining tests are all recursive, so bail out if we hit the limit.
515 if (Depth == MaxDepth)
519 auto m_V = m_CombineOr(m_Specific(V),
520 m_CombineOr(m_PtrToInt(m_Specific(V)),
521 m_BitCast(m_Specific(V))));
523 CmpInst::Predicate Pred;
526 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
527 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
528 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
529 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
530 KnownZero |= RHSKnownZero;
531 KnownOne |= RHSKnownOne;
533 } else if (match(Arg, m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)),
535 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
536 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
537 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
538 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
539 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
541 // For those bits in the mask that are known to be one, we can propagate
542 // known bits from the RHS to V.
543 KnownZero |= RHSKnownZero & MaskKnownOne;
544 KnownOne |= RHSKnownOne & MaskKnownOne;
545 // assume(~(v & b) = a)
546 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
548 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
549 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
550 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
551 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
552 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
554 // For those bits in the mask that are known to be one, we can propagate
555 // inverted known bits from the RHS to V.
556 KnownZero |= RHSKnownOne & MaskKnownOne;
557 KnownOne |= RHSKnownZero & MaskKnownOne;
559 } else if (match(Arg, m_c_ICmp(Pred, 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 known
568 // bits from the RHS to V.
569 KnownZero |= RHSKnownZero & BKnownZero;
570 KnownOne |= RHSKnownOne & BKnownZero;
571 // assume(~(v | b) = a)
572 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
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
581 // inverted known bits from the RHS to V.
582 KnownZero |= RHSKnownOne & BKnownZero;
583 KnownOne |= RHSKnownZero & BKnownZero;
585 } else if (match(Arg, m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)),
587 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
588 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
589 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
590 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
591 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
593 // For those bits in B that are known to be zero, we can propagate known
594 // bits from the RHS to V. For those bits in B that are known to be one,
595 // we can propagate inverted known bits from the RHS to V.
596 KnownZero |= RHSKnownZero & BKnownZero;
597 KnownOne |= RHSKnownOne & BKnownZero;
598 KnownZero |= RHSKnownOne & BKnownOne;
599 KnownOne |= RHSKnownZero & BKnownOne;
600 // assume(~(v ^ b) = a)
601 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
603 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
604 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
605 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
606 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
607 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
609 // For those bits in B that are known to be zero, we can propagate
610 // inverted known bits from the RHS to V. For those bits in B that are
611 // known to be one, we can propagate known bits from the RHS to V.
612 KnownZero |= RHSKnownOne & BKnownZero;
613 KnownOne |= RHSKnownZero & BKnownZero;
614 KnownZero |= RHSKnownZero & BKnownOne;
615 KnownOne |= RHSKnownOne & BKnownOne;
616 // assume(v << c = a)
617 } else if (match(Arg, m_c_ICmp(Pred, 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 to known
623 // bits in V shifted to the right by C.
624 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
625 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
626 // assume(~(v << c) = a)
627 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
629 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
630 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
631 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
632 // For those bits in RHS that are known, we can propagate them inverted
633 // to known bits in V shifted to the right by C.
634 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
635 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
636 // assume(v >> c = a)
637 } else if (match(Arg,
638 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
642 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
643 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
644 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
645 // For those bits in RHS that are known, we can propagate them to known
646 // bits in V shifted to the right by C.
647 KnownZero |= RHSKnownZero << C->getZExtValue();
648 KnownOne |= RHSKnownOne << C->getZExtValue();
649 // assume(~(v >> c) = a)
650 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
651 m_LShr(m_V, m_ConstantInt(C)),
652 m_AShr(m_V, m_ConstantInt(C)))),
654 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
655 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
656 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
657 // For those bits in RHS that are known, we can propagate them inverted
658 // to known bits in V shifted to the right by C.
659 KnownZero |= RHSKnownOne << C->getZExtValue();
660 KnownOne |= RHSKnownZero << C->getZExtValue();
661 // assume(v >=_s c) where c is non-negative
662 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
663 Pred == ICmpInst::ICMP_SGE &&
664 isValidAssumeForContext(I, Q, DL)) {
665 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
666 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
668 if (RHSKnownZero.isNegative()) {
669 // We know that the sign bit is zero.
670 KnownZero |= APInt::getSignBit(BitWidth);
672 // assume(v >_s c) where c is at least -1.
673 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
674 Pred == ICmpInst::ICMP_SGT &&
675 isValidAssumeForContext(I, Q, DL)) {
676 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
677 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
679 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
680 // We know that the sign bit is zero.
681 KnownZero |= APInt::getSignBit(BitWidth);
683 // assume(v <=_s c) where c is negative
684 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
685 Pred == ICmpInst::ICMP_SLE &&
686 isValidAssumeForContext(I, Q, DL)) {
687 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
688 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
690 if (RHSKnownOne.isNegative()) {
691 // We know that the sign bit is one.
692 KnownOne |= APInt::getSignBit(BitWidth);
694 // assume(v <_s c) where c is non-positive
695 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
696 Pred == ICmpInst::ICMP_SLT &&
697 isValidAssumeForContext(I, Q, DL)) {
698 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
699 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
701 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
702 // We know that the sign bit is one.
703 KnownOne |= APInt::getSignBit(BitWidth);
706 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
707 Pred == ICmpInst::ICMP_ULE &&
708 isValidAssumeForContext(I, Q, DL)) {
709 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
710 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
712 // Whatever high bits in c are zero are known to be zero.
714 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
716 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
717 Pred == ICmpInst::ICMP_ULT &&
718 isValidAssumeForContext(I, Q, DL)) {
719 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
720 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
722 // Whatever high bits in c are zero are known to be zero (if c is a power
723 // of 2, then one more).
724 if (isKnownToBeAPowerOfTwo(A, false, Depth+1, Query(Q, I)))
726 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
729 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
734 /// Determine which bits of V are known to be either zero or one and return
735 /// them in the KnownZero/KnownOne bit sets.
737 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
738 /// we cannot optimize based on the assumption that it is zero without changing
739 /// it to be an explicit zero. If we don't change it to zero, other code could
740 /// optimized based on the contradictory assumption that it is non-zero.
741 /// Because instcombine aggressively folds operations with undef args anyway,
742 /// this won't lose us code quality.
744 /// This function is defined on values with integer type, values with pointer
745 /// type (but only if TD is non-null), and vectors of integers. In the case
746 /// where V is a vector, known zero, and known one values are the
747 /// same width as the vector element, and the bit is set only if it is true
748 /// for all of the elements in the vector.
749 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
750 const DataLayout *TD, unsigned Depth,
752 assert(V && "No Value?");
753 assert(Depth <= MaxDepth && "Limit Search Depth");
754 unsigned BitWidth = KnownZero.getBitWidth();
756 assert((V->getType()->isIntOrIntVectorTy() ||
757 V->getType()->getScalarType()->isPointerTy()) &&
758 "Not integer or pointer type!");
760 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
761 (!V->getType()->isIntOrIntVectorTy() ||
762 V->getType()->getScalarSizeInBits() == BitWidth) &&
763 KnownZero.getBitWidth() == BitWidth &&
764 KnownOne.getBitWidth() == BitWidth &&
765 "V, KnownOne and KnownZero should have same BitWidth");
767 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
768 // We know all of the bits for a constant!
769 KnownOne = CI->getValue();
770 KnownZero = ~KnownOne;
773 // Null and aggregate-zero are all-zeros.
774 if (isa<ConstantPointerNull>(V) ||
775 isa<ConstantAggregateZero>(V)) {
776 KnownOne.clearAllBits();
777 KnownZero = APInt::getAllOnesValue(BitWidth);
780 // Handle a constant vector by taking the intersection of the known bits of
781 // each element. There is no real need to handle ConstantVector here, because
782 // we don't handle undef in any particularly useful way.
783 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
784 // We know that CDS must be a vector of integers. Take the intersection of
786 KnownZero.setAllBits(); KnownOne.setAllBits();
787 APInt Elt(KnownZero.getBitWidth(), 0);
788 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
789 Elt = CDS->getElementAsInteger(i);
796 // The address of an aligned GlobalValue has trailing zeros.
797 if (auto *GO = dyn_cast<GlobalObject>(V)) {
798 unsigned Align = GO->getAlignment();
799 if (Align == 0 && TD) {
800 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
801 Type *ObjectType = GVar->getType()->getElementType();
802 if (ObjectType->isSized()) {
803 // If the object is defined in the current Module, we'll be giving
804 // it the preferred alignment. Otherwise, we have to assume that it
805 // may only have the minimum ABI alignment.
806 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
807 Align = TD->getPreferredAlignment(GVar);
809 Align = TD->getABITypeAlignment(ObjectType);
814 KnownZero = APInt::getLowBitsSet(BitWidth,
815 countTrailingZeros(Align));
817 KnownZero.clearAllBits();
818 KnownOne.clearAllBits();
822 if (Argument *A = dyn_cast<Argument>(V)) {
823 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
825 if (!Align && TD && A->hasStructRetAttr()) {
826 // An sret parameter has at least the ABI alignment of the return type.
827 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
828 if (EltTy->isSized())
829 Align = TD->getABITypeAlignment(EltTy);
833 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
835 KnownZero.clearAllBits();
836 KnownOne.clearAllBits();
838 // Don't give up yet... there might be an assumption that provides more
840 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
844 // Start out not knowing anything.
845 KnownZero.clearAllBits(); KnownOne.clearAllBits();
847 // Limit search depth.
848 // All recursive calls that increase depth must come after this.
849 if (Depth == MaxDepth)
852 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
853 // the bits of its aliasee.
854 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
855 if (!GA->mayBeOverridden())
856 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth + 1, Q);
860 // Check whether a nearby assume intrinsic can determine some known bits.
861 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
863 Operator *I = dyn_cast<Operator>(V);
866 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
867 switch (I->getOpcode()) {
869 case Instruction::Load:
870 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
871 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
873 case Instruction::And: {
874 // If either the LHS or the RHS are Zero, the result is zero.
875 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
876 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
878 // Output known-1 bits are only known if set in both the LHS & RHS.
879 KnownOne &= KnownOne2;
880 // Output known-0 are known to be clear if zero in either the LHS | RHS.
881 KnownZero |= KnownZero2;
884 case Instruction::Or: {
885 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
886 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
888 // Output known-0 bits are only known if clear in both the LHS & RHS.
889 KnownZero &= KnownZero2;
890 // Output known-1 are known to be set if set in either the LHS | RHS.
891 KnownOne |= KnownOne2;
894 case Instruction::Xor: {
895 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
896 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
898 // Output known-0 bits are known if clear or set in both the LHS & RHS.
899 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
900 // Output known-1 are known to be set if set in only one of the LHS, RHS.
901 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
902 KnownZero = KnownZeroOut;
905 case Instruction::Mul: {
906 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
907 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
908 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
912 case Instruction::UDiv: {
913 // For the purposes of computing leading zeros we can conservatively
914 // treat a udiv as a logical right shift by the power of 2 known to
915 // be less than the denominator.
916 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
917 unsigned LeadZ = KnownZero2.countLeadingOnes();
919 KnownOne2.clearAllBits();
920 KnownZero2.clearAllBits();
921 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
922 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
923 if (RHSUnknownLeadingOnes != BitWidth)
924 LeadZ = std::min(BitWidth,
925 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
927 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
930 case Instruction::Select:
931 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1, Q);
932 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
934 // Only known if known in both the LHS and RHS.
935 KnownOne &= KnownOne2;
936 KnownZero &= KnownZero2;
938 case Instruction::FPTrunc:
939 case Instruction::FPExt:
940 case Instruction::FPToUI:
941 case Instruction::FPToSI:
942 case Instruction::SIToFP:
943 case Instruction::UIToFP:
944 break; // Can't work with floating point.
945 case Instruction::PtrToInt:
946 case Instruction::IntToPtr:
947 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
948 // We can't handle these if we don't know the pointer size.
950 // FALL THROUGH and handle them the same as zext/trunc.
951 case Instruction::ZExt:
952 case Instruction::Trunc: {
953 Type *SrcTy = I->getOperand(0)->getType();
955 unsigned SrcBitWidth;
956 // Note that we handle pointer operands here because of inttoptr/ptrtoint
957 // which fall through here.
959 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
961 SrcBitWidth = SrcTy->getScalarSizeInBits();
962 if (!SrcBitWidth) break;
965 assert(SrcBitWidth && "SrcBitWidth can't be zero");
966 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
967 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
968 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
969 KnownZero = KnownZero.zextOrTrunc(BitWidth);
970 KnownOne = KnownOne.zextOrTrunc(BitWidth);
971 // Any top bits are known to be zero.
972 if (BitWidth > SrcBitWidth)
973 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
976 case Instruction::BitCast: {
977 Type *SrcTy = I->getOperand(0)->getType();
978 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
979 // TODO: For now, not handling conversions like:
980 // (bitcast i64 %x to <2 x i32>)
981 !I->getType()->isVectorTy()) {
982 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
987 case Instruction::SExt: {
988 // Compute the bits in the result that are not present in the input.
989 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
991 KnownZero = KnownZero.trunc(SrcBitWidth);
992 KnownOne = KnownOne.trunc(SrcBitWidth);
993 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
994 KnownZero = KnownZero.zext(BitWidth);
995 KnownOne = KnownOne.zext(BitWidth);
997 // If the sign bit of the input is known set or clear, then we know the
998 // top bits of the result.
999 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1000 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1001 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1002 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1005 case Instruction::Shl:
1006 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1007 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1008 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1009 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1010 KnownZero <<= ShiftAmt;
1011 KnownOne <<= ShiftAmt;
1012 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1015 case Instruction::LShr:
1016 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1017 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1018 // Compute the new bits that are at the top now.
1019 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1021 // Unsigned shift right.
1022 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1023 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1024 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1025 // high bits known zero.
1026 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1029 case Instruction::AShr:
1030 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1031 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1032 // Compute the new bits that are at the top now.
1033 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1035 // Signed shift right.
1036 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1037 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1038 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1040 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1041 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1042 KnownZero |= HighBits;
1043 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1044 KnownOne |= HighBits;
1047 case Instruction::Sub: {
1048 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1049 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1050 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1054 case Instruction::Add: {
1055 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1056 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1057 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1061 case Instruction::SRem:
1062 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1063 APInt RA = Rem->getValue().abs();
1064 if (RA.isPowerOf2()) {
1065 APInt LowBits = RA - 1;
1066 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD,
1069 // The low bits of the first operand are unchanged by the srem.
1070 KnownZero = KnownZero2 & LowBits;
1071 KnownOne = KnownOne2 & LowBits;
1073 // If the first operand is non-negative or has all low bits zero, then
1074 // the upper bits are all zero.
1075 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1076 KnownZero |= ~LowBits;
1078 // If the first operand is negative and not all low bits are zero, then
1079 // the upper bits are all one.
1080 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1081 KnownOne |= ~LowBits;
1083 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1087 // The sign bit is the LHS's sign bit, except when the result of the
1088 // remainder is zero.
1089 if (KnownZero.isNonNegative()) {
1090 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1091 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
1093 // If it's known zero, our sign bit is also zero.
1094 if (LHSKnownZero.isNegative())
1095 KnownZero.setBit(BitWidth - 1);
1099 case Instruction::URem: {
1100 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1101 APInt RA = Rem->getValue();
1102 if (RA.isPowerOf2()) {
1103 APInt LowBits = (RA - 1);
1104 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
1106 KnownZero |= ~LowBits;
1107 KnownOne &= LowBits;
1112 // Since the result is less than or equal to either operand, any leading
1113 // zero bits in either operand must also exist in the result.
1114 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1115 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
1117 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1118 KnownZero2.countLeadingOnes());
1119 KnownOne.clearAllBits();
1120 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1124 case Instruction::Alloca: {
1125 AllocaInst *AI = cast<AllocaInst>(V);
1126 unsigned Align = AI->getAlignment();
1127 if (Align == 0 && TD)
1128 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1131 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1134 case Instruction::GetElementPtr: {
1135 // Analyze all of the subscripts of this getelementptr instruction
1136 // to determine if we can prove known low zero bits.
1137 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1138 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
1140 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1142 gep_type_iterator GTI = gep_type_begin(I);
1143 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1144 Value *Index = I->getOperand(i);
1145 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1146 // Handle struct member offset arithmetic.
1152 // Handle case when index is vector zeroinitializer
1153 Constant *CIndex = cast<Constant>(Index);
1154 if (CIndex->isZeroValue())
1157 if (CIndex->getType()->isVectorTy())
1158 Index = CIndex->getSplatValue();
1160 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1161 const StructLayout *SL = TD->getStructLayout(STy);
1162 uint64_t Offset = SL->getElementOffset(Idx);
1163 TrailZ = std::min<unsigned>(TrailZ,
1164 countTrailingZeros(Offset));
1166 // Handle array index arithmetic.
1167 Type *IndexedTy = GTI.getIndexedType();
1168 if (!IndexedTy->isSized()) {
1172 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1173 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
1174 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1175 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1, Q);
1176 TrailZ = std::min(TrailZ,
1177 unsigned(countTrailingZeros(TypeSize) +
1178 LocalKnownZero.countTrailingOnes()));
1182 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1185 case Instruction::PHI: {
1186 PHINode *P = cast<PHINode>(I);
1187 // Handle the case of a simple two-predecessor recurrence PHI.
1188 // There's a lot more that could theoretically be done here, but
1189 // this is sufficient to catch some interesting cases.
1190 if (P->getNumIncomingValues() == 2) {
1191 for (unsigned i = 0; i != 2; ++i) {
1192 Value *L = P->getIncomingValue(i);
1193 Value *R = P->getIncomingValue(!i);
1194 Operator *LU = dyn_cast<Operator>(L);
1197 unsigned Opcode = LU->getOpcode();
1198 // Check for operations that have the property that if
1199 // both their operands have low zero bits, the result
1200 // will have low zero bits.
1201 if (Opcode == Instruction::Add ||
1202 Opcode == Instruction::Sub ||
1203 Opcode == Instruction::And ||
1204 Opcode == Instruction::Or ||
1205 Opcode == Instruction::Mul) {
1206 Value *LL = LU->getOperand(0);
1207 Value *LR = LU->getOperand(1);
1208 // Find a recurrence.
1215 // Ok, we have a PHI of the form L op= R. Check for low
1217 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1, Q);
1219 // We need to take the minimum number of known bits
1220 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1221 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1, Q);
1223 KnownZero = APInt::getLowBitsSet(BitWidth,
1224 std::min(KnownZero2.countTrailingOnes(),
1225 KnownZero3.countTrailingOnes()));
1231 // Unreachable blocks may have zero-operand PHI nodes.
1232 if (P->getNumIncomingValues() == 0)
1235 // Otherwise take the unions of the known bit sets of the operands,
1236 // taking conservative care to avoid excessive recursion.
1237 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1238 // Skip if every incoming value references to ourself.
1239 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1242 KnownZero = APInt::getAllOnesValue(BitWidth);
1243 KnownOne = APInt::getAllOnesValue(BitWidth);
1244 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1245 // Skip direct self references.
1246 if (P->getIncomingValue(i) == P) continue;
1248 KnownZero2 = APInt(BitWidth, 0);
1249 KnownOne2 = APInt(BitWidth, 0);
1250 // Recurse, but cap the recursion to one level, because we don't
1251 // want to waste time spinning around in loops.
1252 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
1254 KnownZero &= KnownZero2;
1255 KnownOne &= KnownOne2;
1256 // If all bits have been ruled out, there's no need to check
1258 if (!KnownZero && !KnownOne)
1264 case Instruction::Call:
1265 case Instruction::Invoke:
1266 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1267 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1268 // If a range metadata is attached to this IntrinsicInst, intersect the
1269 // explicit range specified by the metadata and the implicit range of
1271 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1272 switch (II->getIntrinsicID()) {
1274 case Intrinsic::ctlz:
1275 case Intrinsic::cttz: {
1276 unsigned LowBits = Log2_32(BitWidth)+1;
1277 // If this call is undefined for 0, the result will be less than 2^n.
1278 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1280 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1283 case Intrinsic::ctpop: {
1284 unsigned LowBits = Log2_32(BitWidth)+1;
1285 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1288 case Intrinsic::x86_sse42_crc32_64_64:
1289 KnownZero |= APInt::getHighBitsSet(64, 32);
1294 case Instruction::ExtractValue:
1295 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1296 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1297 if (EVI->getNumIndices() != 1) break;
1298 if (EVI->getIndices()[0] == 0) {
1299 switch (II->getIntrinsicID()) {
1301 case Intrinsic::uadd_with_overflow:
1302 case Intrinsic::sadd_with_overflow:
1303 computeKnownBitsAddSub(true, II->getArgOperand(0),
1304 II->getArgOperand(1), false, KnownZero,
1305 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1307 case Intrinsic::usub_with_overflow:
1308 case Intrinsic::ssub_with_overflow:
1309 computeKnownBitsAddSub(false, II->getArgOperand(0),
1310 II->getArgOperand(1), false, KnownZero,
1311 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1313 case Intrinsic::umul_with_overflow:
1314 case Intrinsic::smul_with_overflow:
1315 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
1316 false, KnownZero, KnownOne,
1317 KnownZero2, KnownOne2, TD, Depth, Q);
1324 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1327 /// Determine whether the sign bit is known to be zero or one.
1328 /// Convenience wrapper around computeKnownBits.
1329 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1330 const DataLayout *TD, unsigned Depth,
1332 unsigned BitWidth = getBitWidth(V->getType(), TD);
1338 APInt ZeroBits(BitWidth, 0);
1339 APInt OneBits(BitWidth, 0);
1340 computeKnownBits(V, ZeroBits, OneBits, TD, Depth, Q);
1341 KnownOne = OneBits[BitWidth - 1];
1342 KnownZero = ZeroBits[BitWidth - 1];
1345 /// Return true if the given value is known to have exactly one
1346 /// bit set when defined. For vectors return true if every element is known to
1347 /// be a power of two when defined. Supports values with integer or pointer
1348 /// types and vectors of integers.
1349 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1351 if (Constant *C = dyn_cast<Constant>(V)) {
1352 if (C->isNullValue())
1354 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1355 return CI->getValue().isPowerOf2();
1356 // TODO: Handle vector constants.
1359 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1360 // it is shifted off the end then the result is undefined.
1361 if (match(V, m_Shl(m_One(), m_Value())))
1364 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1365 // bottom. If it is shifted off the bottom then the result is undefined.
1366 if (match(V, m_LShr(m_SignBit(), m_Value())))
1369 // The remaining tests are all recursive, so bail out if we hit the limit.
1370 if (Depth++ == MaxDepth)
1373 Value *X = nullptr, *Y = nullptr;
1374 // A shift of a power of two is a power of two or zero.
1375 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1376 match(V, m_Shr(m_Value(X), m_Value()))))
1377 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q);
1379 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1380 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1382 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1384 isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1385 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1387 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1388 // A power of two and'd with anything is a power of two or zero.
1389 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q) ||
1390 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth, Q))
1392 // X & (-X) is always a power of two or zero.
1393 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1398 // Adding a power-of-two or zero to the same power-of-two or zero yields
1399 // either the original power-of-two, a larger power-of-two or zero.
1400 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1401 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1402 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1403 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1404 match(X, m_And(m_Value(), m_Specific(Y))))
1405 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1407 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1408 match(Y, m_And(m_Value(), m_Specific(X))))
1409 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1412 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1413 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1414 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth, Q);
1416 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1417 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth, Q);
1418 // If i8 V is a power of two or zero:
1419 // ZeroBits: 1 1 1 0 1 1 1 1
1420 // ~ZeroBits: 0 0 0 1 0 0 0 0
1421 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1422 // If OrZero isn't set, we cannot give back a zero result.
1423 // Make sure either the LHS or RHS has a bit set.
1424 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1429 // An exact divide or right shift can only shift off zero bits, so the result
1430 // is a power of two only if the first operand is a power of two and not
1431 // copying a sign bit (sdiv int_min, 2).
1432 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1433 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1434 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1441 /// \brief Test whether a GEP's result is known to be non-null.
1443 /// Uses properties inherent in a GEP to try to determine whether it is known
1446 /// Currently this routine does not support vector GEPs.
1447 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
1448 unsigned Depth, const Query &Q) {
1449 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1452 // FIXME: Support vector-GEPs.
1453 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1455 // If the base pointer is non-null, we cannot walk to a null address with an
1456 // inbounds GEP in address space zero.
1457 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1460 // Past this, if we don't have DataLayout, we can't do much.
1464 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1465 // If so, then the GEP cannot produce a null pointer, as doing so would
1466 // inherently violate the inbounds contract within address space zero.
1467 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1468 GTI != GTE; ++GTI) {
1469 // Struct types are easy -- they must always be indexed by a constant.
1470 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1471 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1472 unsigned ElementIdx = OpC->getZExtValue();
1473 const StructLayout *SL = DL->getStructLayout(STy);
1474 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1475 if (ElementOffset > 0)
1480 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1481 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
1484 // Fast path the constant operand case both for efficiency and so we don't
1485 // increment Depth when just zipping down an all-constant GEP.
1486 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1492 // We post-increment Depth here because while isKnownNonZero increments it
1493 // as well, when we pop back up that increment won't persist. We don't want
1494 // to recurse 10k times just because we have 10k GEP operands. We don't
1495 // bail completely out because we want to handle constant GEPs regardless
1497 if (Depth++ >= MaxDepth)
1500 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1507 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1508 /// ensure that the value it's attached to is never Value? 'RangeType' is
1509 /// is the type of the value described by the range.
1510 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1511 const APInt& Value) {
1512 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1513 assert(NumRanges >= 1);
1514 for (unsigned i = 0; i < NumRanges; ++i) {
1515 ConstantInt *Lower =
1516 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1517 ConstantInt *Upper =
1518 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1519 ConstantRange Range(Lower->getValue(), Upper->getValue());
1520 if (Range.contains(Value))
1526 /// Return true if the given value is known to be non-zero when defined.
1527 /// For vectors return true if every element is known to be non-zero when
1528 /// defined. Supports values with integer or pointer type and vectors of
1530 bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
1532 if (Constant *C = dyn_cast<Constant>(V)) {
1533 if (C->isNullValue())
1535 if (isa<ConstantInt>(C))
1536 // Must be non-zero due to null test above.
1538 // TODO: Handle vectors
1542 if (Instruction* I = dyn_cast<Instruction>(V)) {
1543 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1544 // If the possible ranges don't contain zero, then the value is
1545 // definitely non-zero.
1546 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1547 const APInt ZeroValue(Ty->getBitWidth(), 0);
1548 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1554 // The remaining tests are all recursive, so bail out if we hit the limit.
1555 if (Depth++ >= MaxDepth)
1558 // Check for pointer simplifications.
1559 if (V->getType()->isPointerTy()) {
1560 if (isKnownNonNull(V))
1562 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1563 if (isGEPKnownNonNull(GEP, TD, Depth, Q))
1567 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1569 // X | Y != 0 if X != 0 or Y != 0.
1570 Value *X = nullptr, *Y = nullptr;
1571 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1572 return isKnownNonZero(X, TD, Depth, Q) ||
1573 isKnownNonZero(Y, TD, Depth, Q);
1575 // ext X != 0 if X != 0.
1576 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1577 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth, Q);
1579 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1580 // if the lowest bit is shifted off the end.
1581 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1582 // shl nuw can't remove any non-zero bits.
1583 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1584 if (BO->hasNoUnsignedWrap())
1585 return isKnownNonZero(X, TD, Depth, Q);
1587 APInt KnownZero(BitWidth, 0);
1588 APInt KnownOne(BitWidth, 0);
1589 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1593 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1594 // defined if the sign bit is shifted off the end.
1595 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1596 // shr exact can only shift out zero bits.
1597 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1599 return isKnownNonZero(X, TD, Depth, Q);
1601 bool XKnownNonNegative, XKnownNegative;
1602 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1606 // div exact can only produce a zero if the dividend is zero.
1607 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1608 return isKnownNonZero(X, TD, Depth, Q);
1611 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1612 bool XKnownNonNegative, XKnownNegative;
1613 bool YKnownNonNegative, YKnownNegative;
1614 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1615 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth, Q);
1617 // If X and Y are both non-negative (as signed values) then their sum is not
1618 // zero unless both X and Y are zero.
1619 if (XKnownNonNegative && YKnownNonNegative)
1620 if (isKnownNonZero(X, TD, Depth, Q) ||
1621 isKnownNonZero(Y, TD, Depth, Q))
1624 // If X and Y are both negative (as signed values) then their sum is not
1625 // zero unless both X and Y equal INT_MIN.
1626 if (BitWidth && XKnownNegative && YKnownNegative) {
1627 APInt KnownZero(BitWidth, 0);
1628 APInt KnownOne(BitWidth, 0);
1629 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1630 // The sign bit of X is set. If some other bit is set then X is not equal
1632 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1633 if ((KnownOne & Mask) != 0)
1635 // The sign bit of Y is set. If some other bit is set then Y is not equal
1637 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth, Q);
1638 if ((KnownOne & Mask) != 0)
1642 // The sum of a non-negative number and a power of two is not zero.
1643 if (XKnownNonNegative &&
1644 isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth, Q))
1646 if (YKnownNonNegative &&
1647 isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth, Q))
1651 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1652 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1653 // If X and Y are non-zero then so is X * Y as long as the multiplication
1654 // does not overflow.
1655 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1656 isKnownNonZero(X, TD, Depth, Q) &&
1657 isKnownNonZero(Y, TD, Depth, Q))
1660 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1661 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1662 if (isKnownNonZero(SI->getTrueValue(), TD, Depth, Q) &&
1663 isKnownNonZero(SI->getFalseValue(), TD, Depth, Q))
1667 if (!BitWidth) return false;
1668 APInt KnownZero(BitWidth, 0);
1669 APInt KnownOne(BitWidth, 0);
1670 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1671 return KnownOne != 0;
1674 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1675 /// simplify operations downstream. Mask is known to be zero for bits that V
1678 /// This function is defined on values with integer type, values with pointer
1679 /// type (but only if TD is non-null), and vectors of integers. In the case
1680 /// where V is a vector, the mask, known zero, and known one values are the
1681 /// same width as the vector element, and the bit is set only if it is true
1682 /// for all of the elements in the vector.
1683 bool MaskedValueIsZero(Value *V, const APInt &Mask,
1684 const DataLayout *TD, unsigned Depth,
1686 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1687 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1688 return (KnownZero & Mask) == Mask;
1693 /// Return the number of times the sign bit of the register is replicated into
1694 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1695 /// (itself), but other cases can give us information. For example, immediately
1696 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1697 /// other, so we return 3.
1699 /// 'Op' must have a scalar integer type.
1701 unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
1702 unsigned Depth, const Query &Q) {
1703 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1704 "ComputeNumSignBits requires a DataLayout object to operate "
1705 "on non-integer values!");
1706 Type *Ty = V->getType();
1707 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1708 Ty->getScalarSizeInBits();
1710 unsigned FirstAnswer = 1;
1712 // Note that ConstantInt is handled by the general computeKnownBits case
1716 return 1; // Limit search depth.
1718 Operator *U = dyn_cast<Operator>(V);
1719 switch (Operator::getOpcode(V)) {
1721 case Instruction::SExt:
1722 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1723 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q) + Tmp;
1725 case Instruction::AShr: {
1726 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1727 // ashr X, C -> adds C sign bits. Vectors too.
1729 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1730 Tmp += ShAmt->getZExtValue();
1731 if (Tmp > TyBits) Tmp = TyBits;
1735 case Instruction::Shl: {
1737 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1738 // shl destroys sign bits.
1739 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1740 Tmp2 = ShAmt->getZExtValue();
1741 if (Tmp2 >= TyBits || // Bad shift.
1742 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1747 case Instruction::And:
1748 case Instruction::Or:
1749 case Instruction::Xor: // NOT is handled here.
1750 // Logical binary ops preserve the number of sign bits at the worst.
1751 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1753 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1754 FirstAnswer = std::min(Tmp, Tmp2);
1755 // We computed what we know about the sign bits as our first
1756 // answer. Now proceed to the generic code that uses
1757 // computeKnownBits, and pick whichever answer is better.
1761 case Instruction::Select:
1762 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1763 if (Tmp == 1) return 1; // Early out.
1764 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1, Q);
1765 return std::min(Tmp, Tmp2);
1767 case Instruction::Add:
1768 // Add can have at most one carry bit. Thus we know that the output
1769 // is, at worst, one more bit than the inputs.
1770 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1771 if (Tmp == 1) return 1; // Early out.
1773 // Special case decrementing a value (ADD X, -1):
1774 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
1775 if (CRHS->isAllOnesValue()) {
1776 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1777 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1779 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1781 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1784 // If we are subtracting one from a positive number, there is no carry
1785 // out of the result.
1786 if (KnownZero.isNegative())
1790 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1791 if (Tmp2 == 1) return 1;
1792 return std::min(Tmp, Tmp2)-1;
1794 case Instruction::Sub:
1795 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1796 if (Tmp2 == 1) return 1;
1799 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
1800 if (CLHS->isNullValue()) {
1801 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1802 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
1803 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1805 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1808 // If the input is known to be positive (the sign bit is known clear),
1809 // the output of the NEG has the same number of sign bits as the input.
1810 if (KnownZero.isNegative())
1813 // Otherwise, we treat this like a SUB.
1816 // Sub can have at most one carry bit. Thus we know that the output
1817 // is, at worst, one more bit than the inputs.
1818 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1819 if (Tmp == 1) return 1; // Early out.
1820 return std::min(Tmp, Tmp2)-1;
1822 case Instruction::PHI: {
1823 PHINode *PN = cast<PHINode>(U);
1824 unsigned NumIncomingValues = PN->getNumIncomingValues();
1825 // Don't analyze large in-degree PHIs.
1826 if (NumIncomingValues > 4) break;
1827 // Unreachable blocks may have zero-operand PHI nodes.
1828 if (NumIncomingValues == 0) break;
1830 // Take the minimum of all incoming values. This can't infinitely loop
1831 // because of our depth threshold.
1832 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1, Q);
1833 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
1834 if (Tmp == 1) return Tmp;
1836 ComputeNumSignBits(PN->getIncomingValue(i), TD,
1842 case Instruction::Trunc:
1843 // FIXME: it's tricky to do anything useful for this, but it is an important
1844 // case for targets like X86.
1848 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1849 // use this information.
1850 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1852 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1854 if (KnownZero.isNegative()) { // sign bit is 0
1856 } else if (KnownOne.isNegative()) { // sign bit is 1;
1863 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1864 // the number of identical bits in the top of the input value.
1866 Mask <<= Mask.getBitWidth()-TyBits;
1867 // Return # leading zeros. We use 'min' here in case Val was zero before
1868 // shifting. We don't want to return '64' as for an i32 "0".
1869 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1872 /// This function computes the integer multiple of Base that equals V.
1873 /// If successful, it returns true and returns the multiple in
1874 /// Multiple. If unsuccessful, it returns false. It looks
1875 /// through SExt instructions only if LookThroughSExt is true.
1876 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1877 bool LookThroughSExt, unsigned Depth) {
1878 const unsigned MaxDepth = 6;
1880 assert(V && "No Value?");
1881 assert(Depth <= MaxDepth && "Limit Search Depth");
1882 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1884 Type *T = V->getType();
1886 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1896 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1897 Constant *BaseVal = ConstantInt::get(T, Base);
1898 if (CO && CO == BaseVal) {
1900 Multiple = ConstantInt::get(T, 1);
1904 if (CI && CI->getZExtValue() % Base == 0) {
1905 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1909 if (Depth == MaxDepth) return false; // Limit search depth.
1911 Operator *I = dyn_cast<Operator>(V);
1912 if (!I) return false;
1914 switch (I->getOpcode()) {
1916 case Instruction::SExt:
1917 if (!LookThroughSExt) return false;
1918 // otherwise fall through to ZExt
1919 case Instruction::ZExt:
1920 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1921 LookThroughSExt, Depth+1);
1922 case Instruction::Shl:
1923 case Instruction::Mul: {
1924 Value *Op0 = I->getOperand(0);
1925 Value *Op1 = I->getOperand(1);
1927 if (I->getOpcode() == Instruction::Shl) {
1928 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1929 if (!Op1CI) return false;
1930 // Turn Op0 << Op1 into Op0 * 2^Op1
1931 APInt Op1Int = Op1CI->getValue();
1932 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1933 APInt API(Op1Int.getBitWidth(), 0);
1934 API.setBit(BitToSet);
1935 Op1 = ConstantInt::get(V->getContext(), API);
1938 Value *Mul0 = nullptr;
1939 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1940 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1941 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1942 if (Op1C->getType()->getPrimitiveSizeInBits() <
1943 MulC->getType()->getPrimitiveSizeInBits())
1944 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1945 if (Op1C->getType()->getPrimitiveSizeInBits() >
1946 MulC->getType()->getPrimitiveSizeInBits())
1947 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1949 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1950 Multiple = ConstantExpr::getMul(MulC, Op1C);
1954 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1955 if (Mul0CI->getValue() == 1) {
1956 // V == Base * Op1, so return Op1
1962 Value *Mul1 = nullptr;
1963 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1964 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1965 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1966 if (Op0C->getType()->getPrimitiveSizeInBits() <
1967 MulC->getType()->getPrimitiveSizeInBits())
1968 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1969 if (Op0C->getType()->getPrimitiveSizeInBits() >
1970 MulC->getType()->getPrimitiveSizeInBits())
1971 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1973 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1974 Multiple = ConstantExpr::getMul(MulC, Op0C);
1978 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1979 if (Mul1CI->getValue() == 1) {
1980 // V == Base * Op0, so return Op0
1988 // We could not determine if V is a multiple of Base.
1992 /// Return true if we can prove that the specified FP value is never equal to
1995 /// NOTE: this function will need to be revisited when we support non-default
1998 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1999 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2000 return !CFP->getValueAPF().isNegZero();
2003 return 1; // Limit search depth.
2005 const Operator *I = dyn_cast<Operator>(V);
2006 if (!I) return false;
2008 // Check if the nsz fast-math flag is set
2009 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2010 if (FPO->hasNoSignedZeros())
2013 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2014 if (I->getOpcode() == Instruction::FAdd)
2015 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2016 if (CFP->isNullValue())
2019 // sitofp and uitofp turn into +0.0 for zero.
2020 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2023 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2024 // sqrt(-0.0) = -0.0, no other negative results are possible.
2025 if (II->getIntrinsicID() == Intrinsic::sqrt)
2026 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2028 if (const CallInst *CI = dyn_cast<CallInst>(I))
2029 if (const Function *F = CI->getCalledFunction()) {
2030 if (F->isDeclaration()) {
2032 if (F->getName() == "abs") return true;
2033 // fabs[lf](x) != -0.0
2034 if (F->getName() == "fabs") return true;
2035 if (F->getName() == "fabsf") return true;
2036 if (F->getName() == "fabsl") return true;
2037 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2038 F->getName() == "sqrtl")
2039 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2046 /// If the specified value can be set by repeating the same byte in memory,
2047 /// return the i8 value that it is represented with. This is
2048 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2049 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2050 /// byte store (e.g. i16 0x1234), return null.
2051 Value *llvm::isBytewiseValue(Value *V) {
2052 // All byte-wide stores are splatable, even of arbitrary variables.
2053 if (V->getType()->isIntegerTy(8)) return V;
2055 // Handle 'null' ConstantArrayZero etc.
2056 if (Constant *C = dyn_cast<Constant>(V))
2057 if (C->isNullValue())
2058 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2060 // Constant float and double values can be handled as integer values if the
2061 // corresponding integer value is "byteable". An important case is 0.0.
2062 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2063 if (CFP->getType()->isFloatTy())
2064 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2065 if (CFP->getType()->isDoubleTy())
2066 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2067 // Don't handle long double formats, which have strange constraints.
2070 // We can handle constant integers that are power of two in size and a
2071 // multiple of 8 bits.
2072 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2073 unsigned Width = CI->getBitWidth();
2074 if (isPowerOf2_32(Width) && Width > 8) {
2075 // We can handle this value if the recursive binary decomposition is the
2076 // same at all levels.
2077 APInt Val = CI->getValue();
2079 while (Val.getBitWidth() != 8) {
2080 unsigned NextWidth = Val.getBitWidth()/2;
2081 Val2 = Val.lshr(NextWidth);
2082 Val2 = Val2.trunc(Val.getBitWidth()/2);
2083 Val = Val.trunc(Val.getBitWidth()/2);
2085 // If the top/bottom halves aren't the same, reject it.
2089 return ConstantInt::get(V->getContext(), Val);
2093 // A ConstantDataArray/Vector is splatable if all its members are equal and
2095 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2096 Value *Elt = CA->getElementAsConstant(0);
2097 Value *Val = isBytewiseValue(Elt);
2101 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2102 if (CA->getElementAsConstant(I) != Elt)
2108 // Conceptually, we could handle things like:
2109 // %a = zext i8 %X to i16
2110 // %b = shl i16 %a, 8
2111 // %c = or i16 %a, %b
2112 // but until there is an example that actually needs this, it doesn't seem
2113 // worth worrying about.
2118 // This is the recursive version of BuildSubAggregate. It takes a few different
2119 // arguments. Idxs is the index within the nested struct From that we are
2120 // looking at now (which is of type IndexedType). IdxSkip is the number of
2121 // indices from Idxs that should be left out when inserting into the resulting
2122 // struct. To is the result struct built so far, new insertvalue instructions
2124 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2125 SmallVectorImpl<unsigned> &Idxs,
2127 Instruction *InsertBefore) {
2128 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2130 // Save the original To argument so we can modify it
2132 // General case, the type indexed by Idxs is a struct
2133 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2134 // Process each struct element recursively
2137 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2141 // Couldn't find any inserted value for this index? Cleanup
2142 while (PrevTo != OrigTo) {
2143 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2144 PrevTo = Del->getAggregateOperand();
2145 Del->eraseFromParent();
2147 // Stop processing elements
2151 // If we successfully found a value for each of our subaggregates
2155 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2156 // the struct's elements had a value that was inserted directly. In the latter
2157 // case, perhaps we can't determine each of the subelements individually, but
2158 // we might be able to find the complete struct somewhere.
2160 // Find the value that is at that particular spot
2161 Value *V = FindInsertedValue(From, Idxs);
2166 // Insert the value in the new (sub) aggregrate
2167 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2168 "tmp", InsertBefore);
2171 // This helper takes a nested struct and extracts a part of it (which is again a
2172 // struct) into a new value. For example, given the struct:
2173 // { a, { b, { c, d }, e } }
2174 // and the indices "1, 1" this returns
2177 // It does this by inserting an insertvalue for each element in the resulting
2178 // struct, as opposed to just inserting a single struct. This will only work if
2179 // each of the elements of the substruct are known (ie, inserted into From by an
2180 // insertvalue instruction somewhere).
2182 // All inserted insertvalue instructions are inserted before InsertBefore
2183 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2184 Instruction *InsertBefore) {
2185 assert(InsertBefore && "Must have someplace to insert!");
2186 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2188 Value *To = UndefValue::get(IndexedType);
2189 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2190 unsigned IdxSkip = Idxs.size();
2192 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2195 /// Given an aggregrate and an sequence of indices, see if
2196 /// the scalar value indexed is already around as a register, for example if it
2197 /// were inserted directly into the aggregrate.
2199 /// If InsertBefore is not null, this function will duplicate (modified)
2200 /// insertvalues when a part of a nested struct is extracted.
2201 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2202 Instruction *InsertBefore) {
2203 // Nothing to index? Just return V then (this is useful at the end of our
2205 if (idx_range.empty())
2207 // We have indices, so V should have an indexable type.
2208 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2209 "Not looking at a struct or array?");
2210 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2211 "Invalid indices for type?");
2213 if (Constant *C = dyn_cast<Constant>(V)) {
2214 C = C->getAggregateElement(idx_range[0]);
2215 if (!C) return nullptr;
2216 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2219 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2220 // Loop the indices for the insertvalue instruction in parallel with the
2221 // requested indices
2222 const unsigned *req_idx = idx_range.begin();
2223 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2224 i != e; ++i, ++req_idx) {
2225 if (req_idx == idx_range.end()) {
2226 // We can't handle this without inserting insertvalues
2230 // The requested index identifies a part of a nested aggregate. Handle
2231 // this specially. For example,
2232 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2233 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2234 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2235 // This can be changed into
2236 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2237 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2238 // which allows the unused 0,0 element from the nested struct to be
2240 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2244 // This insert value inserts something else than what we are looking for.
2245 // See if the (aggregrate) value inserted into has the value we are
2246 // looking for, then.
2248 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2251 // If we end up here, the indices of the insertvalue match with those
2252 // requested (though possibly only partially). Now we recursively look at
2253 // the inserted value, passing any remaining indices.
2254 return FindInsertedValue(I->getInsertedValueOperand(),
2255 makeArrayRef(req_idx, idx_range.end()),
2259 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2260 // If we're extracting a value from an aggregrate that was extracted from
2261 // something else, we can extract from that something else directly instead.
2262 // However, we will need to chain I's indices with the requested indices.
2264 // Calculate the number of indices required
2265 unsigned size = I->getNumIndices() + idx_range.size();
2266 // Allocate some space to put the new indices in
2267 SmallVector<unsigned, 5> Idxs;
2269 // Add indices from the extract value instruction
2270 Idxs.append(I->idx_begin(), I->idx_end());
2272 // Add requested indices
2273 Idxs.append(idx_range.begin(), idx_range.end());
2275 assert(Idxs.size() == size
2276 && "Number of indices added not correct?");
2278 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2280 // Otherwise, we don't know (such as, extracting from a function return value
2281 // or load instruction)
2285 /// Analyze the specified pointer to see if it can be expressed as a base
2286 /// pointer plus a constant offset. Return the base and offset to the caller.
2287 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2288 const DataLayout *DL) {
2289 // Without DataLayout, conservatively assume 64-bit offsets, which is
2290 // the widest we support.
2291 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
2292 APInt ByteOffset(BitWidth, 0);
2294 if (Ptr->getType()->isVectorTy())
2297 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2299 APInt GEPOffset(BitWidth, 0);
2300 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
2303 ByteOffset += GEPOffset;
2306 Ptr = GEP->getPointerOperand();
2307 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2308 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2309 Ptr = cast<Operator>(Ptr)->getOperand(0);
2310 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2311 if (GA->mayBeOverridden())
2313 Ptr = GA->getAliasee();
2318 Offset = ByteOffset.getSExtValue();
2323 /// This function computes the length of a null-terminated C string pointed to
2324 /// by V. If successful, it returns true and returns the string in Str.
2325 /// If unsuccessful, it returns false.
2326 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2327 uint64_t Offset, bool TrimAtNul) {
2330 // Look through bitcast instructions and geps.
2331 V = V->stripPointerCasts();
2333 // If the value is a GEP instructionor constant expression, treat it as an
2335 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2336 // Make sure the GEP has exactly three arguments.
2337 if (GEP->getNumOperands() != 3)
2340 // Make sure the index-ee is a pointer to array of i8.
2341 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2342 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2343 if (!AT || !AT->getElementType()->isIntegerTy(8))
2346 // Check to make sure that the first operand of the GEP is an integer and
2347 // has value 0 so that we are sure we're indexing into the initializer.
2348 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2349 if (!FirstIdx || !FirstIdx->isZero())
2352 // If the second index isn't a ConstantInt, then this is a variable index
2353 // into the array. If this occurs, we can't say anything meaningful about
2355 uint64_t StartIdx = 0;
2356 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2357 StartIdx = CI->getZExtValue();
2360 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
2363 // The GEP instruction, constant or instruction, must reference a global
2364 // variable that is a constant and is initialized. The referenced constant
2365 // initializer is the array that we'll use for optimization.
2366 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2367 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2370 // Handle the all-zeros case
2371 if (GV->getInitializer()->isNullValue()) {
2372 // This is a degenerate case. The initializer is constant zero so the
2373 // length of the string must be zero.
2378 // Must be a Constant Array
2379 const ConstantDataArray *Array =
2380 dyn_cast<ConstantDataArray>(GV->getInitializer());
2381 if (!Array || !Array->isString())
2384 // Get the number of elements in the array
2385 uint64_t NumElts = Array->getType()->getArrayNumElements();
2387 // Start out with the entire array in the StringRef.
2388 Str = Array->getAsString();
2390 if (Offset > NumElts)
2393 // Skip over 'offset' bytes.
2394 Str = Str.substr(Offset);
2397 // Trim off the \0 and anything after it. If the array is not nul
2398 // terminated, we just return the whole end of string. The client may know
2399 // some other way that the string is length-bound.
2400 Str = Str.substr(0, Str.find('\0'));
2405 // These next two are very similar to the above, but also look through PHI
2407 // TODO: See if we can integrate these two together.
2409 /// If we can compute the length of the string pointed to by
2410 /// the specified pointer, return 'len+1'. If we can't, return 0.
2411 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2412 // Look through noop bitcast instructions.
2413 V = V->stripPointerCasts();
2415 // If this is a PHI node, there are two cases: either we have already seen it
2417 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2418 if (!PHIs.insert(PN).second)
2419 return ~0ULL; // already in the set.
2421 // If it was new, see if all the input strings are the same length.
2422 uint64_t LenSoFar = ~0ULL;
2423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2424 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2425 if (Len == 0) return 0; // Unknown length -> unknown.
2427 if (Len == ~0ULL) continue;
2429 if (Len != LenSoFar && LenSoFar != ~0ULL)
2430 return 0; // Disagree -> unknown.
2434 // Success, all agree.
2438 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2439 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2440 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2441 if (Len1 == 0) return 0;
2442 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2443 if (Len2 == 0) return 0;
2444 if (Len1 == ~0ULL) return Len2;
2445 if (Len2 == ~0ULL) return Len1;
2446 if (Len1 != Len2) return 0;
2450 // Otherwise, see if we can read the string.
2452 if (!getConstantStringInfo(V, StrData))
2455 return StrData.size()+1;
2458 /// If we can compute the length of the string pointed to by
2459 /// the specified pointer, return 'len+1'. If we can't, return 0.
2460 uint64_t llvm::GetStringLength(Value *V) {
2461 if (!V->getType()->isPointerTy()) return 0;
2463 SmallPtrSet<PHINode*, 32> PHIs;
2464 uint64_t Len = GetStringLengthH(V, PHIs);
2465 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2466 // an empty string as a length.
2467 return Len == ~0ULL ? 1 : Len;
2471 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
2472 if (!V->getType()->isPointerTy())
2474 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2475 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2476 V = GEP->getPointerOperand();
2477 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2478 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2479 V = cast<Operator>(V)->getOperand(0);
2480 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2481 if (GA->mayBeOverridden())
2483 V = GA->getAliasee();
2485 // See if InstructionSimplify knows any relevant tricks.
2486 if (Instruction *I = dyn_cast<Instruction>(V))
2487 // TODO: Acquire a DominatorTree and AssumptionTracker and use them.
2488 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
2495 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2501 llvm::GetUnderlyingObjects(Value *V,
2502 SmallVectorImpl<Value *> &Objects,
2503 const DataLayout *TD,
2504 unsigned MaxLookup) {
2505 SmallPtrSet<Value *, 4> Visited;
2506 SmallVector<Value *, 4> Worklist;
2507 Worklist.push_back(V);
2509 Value *P = Worklist.pop_back_val();
2510 P = GetUnderlyingObject(P, TD, MaxLookup);
2512 if (!Visited.insert(P).second)
2515 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2516 Worklist.push_back(SI->getTrueValue());
2517 Worklist.push_back(SI->getFalseValue());
2521 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2522 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2523 Worklist.push_back(PN->getIncomingValue(i));
2527 Objects.push_back(P);
2528 } while (!Worklist.empty());
2531 /// Return true if the only users of this pointer are lifetime markers.
2532 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2533 for (const User *U : V->users()) {
2534 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2535 if (!II) return false;
2537 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2538 II->getIntrinsicID() != Intrinsic::lifetime_end)
2544 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
2545 const DataLayout *TD) {
2546 const Operator *Inst = dyn_cast<Operator>(V);
2550 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2551 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2555 switch (Inst->getOpcode()) {
2558 case Instruction::UDiv:
2559 case Instruction::URem: {
2560 // x / y is undefined if y == 0.
2562 if (match(Inst->getOperand(1), m_APInt(V)))
2566 case Instruction::SDiv:
2567 case Instruction::SRem: {
2568 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
2570 if (match(Inst->getOperand(1), m_APInt(Y))) {
2573 // The numerator can't be MinSignedValue if the denominator is -1.
2574 if (match(Inst->getOperand(0), m_APInt(X)))
2575 return !Y->isMinSignedValue();
2576 // The numerator *might* be MinSignedValue.
2579 // The denominator is not 0 or -1, it's safe to proceed.
2585 case Instruction::Load: {
2586 const LoadInst *LI = cast<LoadInst>(Inst);
2587 if (!LI->isUnordered() ||
2588 // Speculative load may create a race that did not exist in the source.
2589 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2591 return LI->getPointerOperand()->isDereferenceablePointer(TD);
2593 case Instruction::Call: {
2594 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2595 switch (II->getIntrinsicID()) {
2596 // These synthetic intrinsics have no side-effects and just mark
2597 // information about their operands.
2598 // FIXME: There are other no-op synthetic instructions that potentially
2599 // should be considered at least *safe* to speculate...
2600 case Intrinsic::dbg_declare:
2601 case Intrinsic::dbg_value:
2604 case Intrinsic::bswap:
2605 case Intrinsic::ctlz:
2606 case Intrinsic::ctpop:
2607 case Intrinsic::cttz:
2608 case Intrinsic::objectsize:
2609 case Intrinsic::sadd_with_overflow:
2610 case Intrinsic::smul_with_overflow:
2611 case Intrinsic::ssub_with_overflow:
2612 case Intrinsic::uadd_with_overflow:
2613 case Intrinsic::umul_with_overflow:
2614 case Intrinsic::usub_with_overflow:
2616 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2617 // errno like libm sqrt would.
2618 case Intrinsic::sqrt:
2619 case Intrinsic::fma:
2620 case Intrinsic::fmuladd:
2621 case Intrinsic::fabs:
2622 case Intrinsic::minnum:
2623 case Intrinsic::maxnum:
2625 // TODO: some fp intrinsics are marked as having the same error handling
2626 // as libm. They're safe to speculate when they won't error.
2627 // TODO: are convert_{from,to}_fp16 safe?
2628 // TODO: can we list target-specific intrinsics here?
2632 return false; // The called function could have undefined behavior or
2633 // side-effects, even if marked readnone nounwind.
2635 case Instruction::VAArg:
2636 case Instruction::Alloca:
2637 case Instruction::Invoke:
2638 case Instruction::PHI:
2639 case Instruction::Store:
2640 case Instruction::Ret:
2641 case Instruction::Br:
2642 case Instruction::IndirectBr:
2643 case Instruction::Switch:
2644 case Instruction::Unreachable:
2645 case Instruction::Fence:
2646 case Instruction::LandingPad:
2647 case Instruction::AtomicRMW:
2648 case Instruction::AtomicCmpXchg:
2649 case Instruction::Resume:
2650 return false; // Misc instructions which have effects
2654 /// Return true if we know that the specified value is never null.
2655 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2656 // Alloca never returns null, malloc might.
2657 if (isa<AllocaInst>(V)) return true;
2659 // A byval, inalloca, or nonnull argument is never null.
2660 if (const Argument *A = dyn_cast<Argument>(V))
2661 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2663 // Global values are not null unless extern weak.
2664 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2665 return !GV->hasExternalWeakLinkage();
2667 // A Load tagged w/nonnull metadata is never null.
2668 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2669 return LI->getMetadata(LLVMContext::MD_nonnull);
2671 if (ImmutableCallSite CS = V)
2672 if (CS.isReturnNonNull())
2675 // operator new never returns null.
2676 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
2682 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
2683 const DataLayout *DL,
2684 AssumptionTracker *AT,
2685 const Instruction *CxtI,
2686 const DominatorTree *DT) {
2687 // Multiplying n * m significant bits yields a result of n + m significant
2688 // bits. If the total number of significant bits does not exceed the
2689 // result bit width (minus 1), there is no overflow.
2690 // This means if we have enough leading zero bits in the operands
2691 // we can guarantee that the result does not overflow.
2692 // Ref: "Hacker's Delight" by Henry Warren
2693 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
2694 APInt LHSKnownZero(BitWidth, 0);
2695 APInt LHSKnownOne(BitWidth, 0);
2696 APInt RHSKnownZero(BitWidth, 0);
2697 APInt RHSKnownOne(BitWidth, 0);
2698 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AT, CxtI, DT);
2699 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AT, CxtI, DT);
2700 // Note that underestimating the number of zero bits gives a more
2701 // conservative answer.
2702 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
2703 RHSKnownZero.countLeadingOnes();
2704 // First handle the easy case: if we have enough zero bits there's
2705 // definitely no overflow.
2706 if (ZeroBits >= BitWidth)
2707 return OverflowResult::NeverOverflows;
2709 // Get the largest possible values for each operand.
2710 APInt LHSMax = ~LHSKnownZero;
2711 APInt RHSMax = ~RHSKnownZero;
2713 // We know the multiply operation doesn't overflow if the maximum values for
2714 // each operand will not overflow after we multiply them together.
2716 LHSMax.umul_ov(RHSMax, MaxOverflow);
2718 return OverflowResult::NeverOverflows;
2720 // We know it always overflows if multiplying the smallest possible values for
2721 // the operands also results in overflow.
2723 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
2725 return OverflowResult::AlwaysOverflows;
2727 return OverflowResult::MayOverflow;