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();
515 auto m_V = m_CombineOr(m_Specific(V),
516 m_CombineOr(m_PtrToInt(m_Specific(V)),
517 m_BitCast(m_Specific(V))));
519 CmpInst::Predicate Pred;
522 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
523 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
524 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
525 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
526 KnownZero |= RHSKnownZero;
527 KnownOne |= RHSKnownOne;
529 } else if (match(Arg, m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)),
531 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
532 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
533 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
534 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
535 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
537 // For those bits in the mask that are known to be one, we can propagate
538 // known bits from the RHS to V.
539 KnownZero |= RHSKnownZero & MaskKnownOne;
540 KnownOne |= RHSKnownOne & MaskKnownOne;
541 // assume(~(v & b) = a)
542 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
544 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
545 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
546 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
547 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
548 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
550 // For those bits in the mask that are known to be one, we can propagate
551 // inverted known bits from the RHS to V.
552 KnownZero |= RHSKnownOne & MaskKnownOne;
553 KnownOne |= RHSKnownZero & MaskKnownOne;
555 } else if (match(Arg, m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)),
557 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
558 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
559 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
560 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
561 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
563 // For those bits in B that are known to be zero, we can propagate known
564 // bits from the RHS to V.
565 KnownZero |= RHSKnownZero & BKnownZero;
566 KnownOne |= RHSKnownOne & BKnownZero;
567 // assume(~(v | b) = a)
568 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
570 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
571 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
572 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
573 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
574 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
576 // For those bits in B that are known to be zero, we can propagate
577 // inverted known bits from the RHS to V.
578 KnownZero |= RHSKnownOne & BKnownZero;
579 KnownOne |= RHSKnownZero & BKnownZero;
581 } else if (match(Arg, m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)),
583 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
584 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
585 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
586 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
587 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
589 // For those bits in B that are known to be zero, we can propagate known
590 // bits from the RHS to V. For those bits in B that are known to be one,
591 // we can propagate inverted known bits from the RHS to V.
592 KnownZero |= RHSKnownZero & BKnownZero;
593 KnownOne |= RHSKnownOne & BKnownZero;
594 KnownZero |= RHSKnownOne & BKnownOne;
595 KnownOne |= RHSKnownZero & BKnownOne;
596 // assume(~(v ^ b) = a)
597 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
599 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
600 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
601 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
602 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
603 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
605 // For those bits in B that are known to be zero, we can propagate
606 // inverted known bits from the RHS to V. For those bits in B that are
607 // known to be one, we can propagate known bits from the RHS to V.
608 KnownZero |= RHSKnownOne & BKnownZero;
609 KnownOne |= RHSKnownZero & BKnownZero;
610 KnownZero |= RHSKnownZero & BKnownOne;
611 KnownOne |= RHSKnownOne & BKnownOne;
612 // assume(v << c = a)
613 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
615 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
616 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
617 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
618 // For those bits in RHS that are known, we can propagate them to known
619 // bits in V shifted to the right by C.
620 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
621 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
622 // assume(~(v << c) = a)
623 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
625 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
626 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
627 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
628 // For those bits in RHS that are known, we can propagate them inverted
629 // to known bits in V shifted to the right by C.
630 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
631 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
632 // assume(v >> c = a)
633 } else if (match(Arg,
634 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
638 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
639 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
640 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
641 // For those bits in RHS that are known, we can propagate them to known
642 // bits in V shifted to the right by C.
643 KnownZero |= RHSKnownZero << C->getZExtValue();
644 KnownOne |= RHSKnownOne << C->getZExtValue();
645 // assume(~(v >> c) = a)
646 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
647 m_LShr(m_V, m_ConstantInt(C)),
648 m_AShr(m_V, m_ConstantInt(C)))),
650 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
651 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
652 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
653 // For those bits in RHS that are known, we can propagate them inverted
654 // to known bits in V shifted to the right by C.
655 KnownZero |= RHSKnownOne << C->getZExtValue();
656 KnownOne |= RHSKnownZero << C->getZExtValue();
657 // assume(v >=_s c) where c is non-negative
658 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
659 Pred == ICmpInst::ICMP_SGE &&
660 isValidAssumeForContext(I, Q, DL)) {
661 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
662 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
664 if (RHSKnownZero.isNegative()) {
665 // We know that the sign bit is zero.
666 KnownZero |= APInt::getSignBit(BitWidth);
668 // assume(v >_s c) where c is at least -1.
669 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
670 Pred == ICmpInst::ICMP_SGT &&
671 isValidAssumeForContext(I, Q, DL)) {
672 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
673 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
675 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
676 // We know that the sign bit is zero.
677 KnownZero |= APInt::getSignBit(BitWidth);
679 // assume(v <=_s c) where c is negative
680 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
681 Pred == ICmpInst::ICMP_SLE &&
682 isValidAssumeForContext(I, Q, DL)) {
683 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
684 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
686 if (RHSKnownOne.isNegative()) {
687 // We know that the sign bit is one.
688 KnownOne |= APInt::getSignBit(BitWidth);
690 // assume(v <_s c) where c is non-positive
691 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
692 Pred == ICmpInst::ICMP_SLT &&
693 isValidAssumeForContext(I, Q, DL)) {
694 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
695 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
697 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
698 // We know that the sign bit is one.
699 KnownOne |= APInt::getSignBit(BitWidth);
702 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
703 Pred == ICmpInst::ICMP_ULE &&
704 isValidAssumeForContext(I, Q, DL)) {
705 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
706 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
708 // Whatever high bits in c are zero are known to be zero.
710 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
712 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
713 Pred == ICmpInst::ICMP_ULT &&
714 isValidAssumeForContext(I, Q, DL)) {
715 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
716 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
718 // Whatever high bits in c are zero are known to be zero (if c is a power
719 // of 2, then one more).
720 if (isKnownToBeAPowerOfTwo(A, false, Depth+1, Query(Q, I)))
722 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
725 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
730 /// Determine which bits of V are known to be either zero or one and return
731 /// them in the KnownZero/KnownOne bit sets.
733 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
734 /// we cannot optimize based on the assumption that it is zero without changing
735 /// it to be an explicit zero. If we don't change it to zero, other code could
736 /// optimized based on the contradictory assumption that it is non-zero.
737 /// Because instcombine aggressively folds operations with undef args anyway,
738 /// this won't lose us code quality.
740 /// This function is defined on values with integer type, values with pointer
741 /// type (but only if TD is non-null), and vectors of integers. In the case
742 /// where V is a vector, known zero, and known one values are the
743 /// same width as the vector element, and the bit is set only if it is true
744 /// for all of the elements in the vector.
745 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
746 const DataLayout *TD, unsigned Depth,
748 assert(V && "No Value?");
749 assert(Depth <= MaxDepth && "Limit Search Depth");
750 unsigned BitWidth = KnownZero.getBitWidth();
752 assert((V->getType()->isIntOrIntVectorTy() ||
753 V->getType()->getScalarType()->isPointerTy()) &&
754 "Not integer or pointer type!");
756 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
757 (!V->getType()->isIntOrIntVectorTy() ||
758 V->getType()->getScalarSizeInBits() == BitWidth) &&
759 KnownZero.getBitWidth() == BitWidth &&
760 KnownOne.getBitWidth() == BitWidth &&
761 "V, KnownOne and KnownZero should have same BitWidth");
763 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
764 // We know all of the bits for a constant!
765 KnownOne = CI->getValue();
766 KnownZero = ~KnownOne;
769 // Null and aggregate-zero are all-zeros.
770 if (isa<ConstantPointerNull>(V) ||
771 isa<ConstantAggregateZero>(V)) {
772 KnownOne.clearAllBits();
773 KnownZero = APInt::getAllOnesValue(BitWidth);
776 // Handle a constant vector by taking the intersection of the known bits of
777 // each element. There is no real need to handle ConstantVector here, because
778 // we don't handle undef in any particularly useful way.
779 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
780 // We know that CDS must be a vector of integers. Take the intersection of
782 KnownZero.setAllBits(); KnownOne.setAllBits();
783 APInt Elt(KnownZero.getBitWidth(), 0);
784 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
785 Elt = CDS->getElementAsInteger(i);
792 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
793 // the bits of its aliasee.
794 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
795 if (GA->mayBeOverridden()) {
796 KnownZero.clearAllBits(); KnownOne.clearAllBits();
798 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth+1, Q);
803 // The address of an aligned GlobalValue has trailing zeros.
804 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
805 unsigned Align = GV->getAlignment();
806 if (Align == 0 && TD) {
807 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
808 Type *ObjectType = GVar->getType()->getElementType();
809 if (ObjectType->isSized()) {
810 // If the object is defined in the current Module, we'll be giving
811 // it the preferred alignment. Otherwise, we have to assume that it
812 // may only have the minimum ABI alignment.
813 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
814 Align = TD->getPreferredAlignment(GVar);
816 Align = TD->getABITypeAlignment(ObjectType);
821 KnownZero = APInt::getLowBitsSet(BitWidth,
822 countTrailingZeros(Align));
824 KnownZero.clearAllBits();
825 KnownOne.clearAllBits();
829 if (Argument *A = dyn_cast<Argument>(V)) {
830 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
832 if (!Align && TD && A->hasStructRetAttr()) {
833 // An sret parameter has at least the ABI alignment of the return type.
834 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
835 if (EltTy->isSized())
836 Align = TD->getABITypeAlignment(EltTy);
840 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
842 // Don't give up yet... there might be an assumption that provides more
844 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
848 // Start out not knowing anything.
849 KnownZero.clearAllBits(); KnownOne.clearAllBits();
851 if (Depth == MaxDepth)
852 return; // Limit search depth.
854 // Check whether a nearby assume intrinsic can determine some known bits.
855 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
857 Operator *I = dyn_cast<Operator>(V);
860 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
861 switch (I->getOpcode()) {
863 case Instruction::Load:
864 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
865 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
867 case Instruction::And: {
868 // If either the LHS or the RHS are Zero, the result is zero.
869 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
870 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
872 // Output known-1 bits are only known if set in both the LHS & RHS.
873 KnownOne &= KnownOne2;
874 // Output known-0 are known to be clear if zero in either the LHS | RHS.
875 KnownZero |= KnownZero2;
878 case Instruction::Or: {
879 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
880 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
882 // Output known-0 bits are only known if clear in both the LHS & RHS.
883 KnownZero &= KnownZero2;
884 // Output known-1 are known to be set if set in either the LHS | RHS.
885 KnownOne |= KnownOne2;
888 case Instruction::Xor: {
889 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
890 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
892 // Output known-0 bits are known if clear or set in both the LHS & RHS.
893 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
894 // Output known-1 are known to be set if set in only one of the LHS, RHS.
895 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
896 KnownZero = KnownZeroOut;
899 case Instruction::Mul: {
900 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
901 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
902 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
906 case Instruction::UDiv: {
907 // For the purposes of computing leading zeros we can conservatively
908 // treat a udiv as a logical right shift by the power of 2 known to
909 // be less than the denominator.
910 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
911 unsigned LeadZ = KnownZero2.countLeadingOnes();
913 KnownOne2.clearAllBits();
914 KnownZero2.clearAllBits();
915 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
916 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
917 if (RHSUnknownLeadingOnes != BitWidth)
918 LeadZ = std::min(BitWidth,
919 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
921 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
924 case Instruction::Select:
925 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1, Q);
926 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
928 // Only known if known in both the LHS and RHS.
929 KnownOne &= KnownOne2;
930 KnownZero &= KnownZero2;
932 case Instruction::FPTrunc:
933 case Instruction::FPExt:
934 case Instruction::FPToUI:
935 case Instruction::FPToSI:
936 case Instruction::SIToFP:
937 case Instruction::UIToFP:
938 break; // Can't work with floating point.
939 case Instruction::PtrToInt:
940 case Instruction::IntToPtr:
941 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
942 // We can't handle these if we don't know the pointer size.
944 // FALL THROUGH and handle them the same as zext/trunc.
945 case Instruction::ZExt:
946 case Instruction::Trunc: {
947 Type *SrcTy = I->getOperand(0)->getType();
949 unsigned SrcBitWidth;
950 // Note that we handle pointer operands here because of inttoptr/ptrtoint
951 // which fall through here.
953 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
955 SrcBitWidth = SrcTy->getScalarSizeInBits();
956 if (!SrcBitWidth) break;
959 assert(SrcBitWidth && "SrcBitWidth can't be zero");
960 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
961 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
962 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
963 KnownZero = KnownZero.zextOrTrunc(BitWidth);
964 KnownOne = KnownOne.zextOrTrunc(BitWidth);
965 // Any top bits are known to be zero.
966 if (BitWidth > SrcBitWidth)
967 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
970 case Instruction::BitCast: {
971 Type *SrcTy = I->getOperand(0)->getType();
972 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
973 // TODO: For now, not handling conversions like:
974 // (bitcast i64 %x to <2 x i32>)
975 !I->getType()->isVectorTy()) {
976 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
981 case Instruction::SExt: {
982 // Compute the bits in the result that are not present in the input.
983 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
985 KnownZero = KnownZero.trunc(SrcBitWidth);
986 KnownOne = KnownOne.trunc(SrcBitWidth);
987 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
988 KnownZero = KnownZero.zext(BitWidth);
989 KnownOne = KnownOne.zext(BitWidth);
991 // If the sign bit of the input is known set or clear, then we know the
992 // top bits of the result.
993 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
994 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
995 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
996 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
999 case Instruction::Shl:
1000 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1001 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1002 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1003 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1004 KnownZero <<= ShiftAmt;
1005 KnownOne <<= ShiftAmt;
1006 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1009 case Instruction::LShr:
1010 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1011 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1012 // Compute the new bits that are at the top now.
1013 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1015 // Unsigned shift right.
1016 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1017 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1018 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1019 // high bits known zero.
1020 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1023 case Instruction::AShr:
1024 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1025 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1026 // Compute the new bits that are at the top now.
1027 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1029 // Signed shift right.
1030 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1031 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1032 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1034 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1035 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1036 KnownZero |= HighBits;
1037 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1038 KnownOne |= HighBits;
1041 case Instruction::Sub: {
1042 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1043 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1044 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1048 case Instruction::Add: {
1049 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1050 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1051 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1055 case Instruction::SRem:
1056 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1057 APInt RA = Rem->getValue().abs();
1058 if (RA.isPowerOf2()) {
1059 APInt LowBits = RA - 1;
1060 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD,
1063 // The low bits of the first operand are unchanged by the srem.
1064 KnownZero = KnownZero2 & LowBits;
1065 KnownOne = KnownOne2 & LowBits;
1067 // If the first operand is non-negative or has all low bits zero, then
1068 // the upper bits are all zero.
1069 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1070 KnownZero |= ~LowBits;
1072 // If the first operand is negative and not all low bits are zero, then
1073 // the upper bits are all one.
1074 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1075 KnownOne |= ~LowBits;
1077 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1081 // The sign bit is the LHS's sign bit, except when the result of the
1082 // remainder is zero.
1083 if (KnownZero.isNonNegative()) {
1084 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1085 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
1087 // If it's known zero, our sign bit is also zero.
1088 if (LHSKnownZero.isNegative())
1089 KnownZero.setBit(BitWidth - 1);
1093 case Instruction::URem: {
1094 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1095 APInt RA = Rem->getValue();
1096 if (RA.isPowerOf2()) {
1097 APInt LowBits = (RA - 1);
1098 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
1100 KnownZero |= ~LowBits;
1101 KnownOne &= LowBits;
1106 // Since the result is less than or equal to either operand, any leading
1107 // zero bits in either operand must also exist in the result.
1108 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1109 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
1111 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1112 KnownZero2.countLeadingOnes());
1113 KnownOne.clearAllBits();
1114 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1118 case Instruction::Alloca: {
1119 AllocaInst *AI = cast<AllocaInst>(V);
1120 unsigned Align = AI->getAlignment();
1121 if (Align == 0 && TD)
1122 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1125 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1128 case Instruction::GetElementPtr: {
1129 // Analyze all of the subscripts of this getelementptr instruction
1130 // to determine if we can prove known low zero bits.
1131 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1132 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
1134 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1136 gep_type_iterator GTI = gep_type_begin(I);
1137 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1138 Value *Index = I->getOperand(i);
1139 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1140 // Handle struct member offset arithmetic.
1146 // Handle case when index is vector zeroinitializer
1147 Constant *CIndex = cast<Constant>(Index);
1148 if (CIndex->isZeroValue())
1151 if (CIndex->getType()->isVectorTy())
1152 Index = CIndex->getSplatValue();
1154 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1155 const StructLayout *SL = TD->getStructLayout(STy);
1156 uint64_t Offset = SL->getElementOffset(Idx);
1157 TrailZ = std::min<unsigned>(TrailZ,
1158 countTrailingZeros(Offset));
1160 // Handle array index arithmetic.
1161 Type *IndexedTy = GTI.getIndexedType();
1162 if (!IndexedTy->isSized()) {
1166 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1167 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
1168 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1169 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1, Q);
1170 TrailZ = std::min(TrailZ,
1171 unsigned(countTrailingZeros(TypeSize) +
1172 LocalKnownZero.countTrailingOnes()));
1176 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1179 case Instruction::PHI: {
1180 PHINode *P = cast<PHINode>(I);
1181 // Handle the case of a simple two-predecessor recurrence PHI.
1182 // There's a lot more that could theoretically be done here, but
1183 // this is sufficient to catch some interesting cases.
1184 if (P->getNumIncomingValues() == 2) {
1185 for (unsigned i = 0; i != 2; ++i) {
1186 Value *L = P->getIncomingValue(i);
1187 Value *R = P->getIncomingValue(!i);
1188 Operator *LU = dyn_cast<Operator>(L);
1191 unsigned Opcode = LU->getOpcode();
1192 // Check for operations that have the property that if
1193 // both their operands have low zero bits, the result
1194 // will have low zero bits.
1195 if (Opcode == Instruction::Add ||
1196 Opcode == Instruction::Sub ||
1197 Opcode == Instruction::And ||
1198 Opcode == Instruction::Or ||
1199 Opcode == Instruction::Mul) {
1200 Value *LL = LU->getOperand(0);
1201 Value *LR = LU->getOperand(1);
1202 // Find a recurrence.
1209 // Ok, we have a PHI of the form L op= R. Check for low
1211 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1, Q);
1213 // We need to take the minimum number of known bits
1214 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1215 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1, Q);
1217 KnownZero = APInt::getLowBitsSet(BitWidth,
1218 std::min(KnownZero2.countTrailingOnes(),
1219 KnownZero3.countTrailingOnes()));
1225 // Unreachable blocks may have zero-operand PHI nodes.
1226 if (P->getNumIncomingValues() == 0)
1229 // Otherwise take the unions of the known bit sets of the operands,
1230 // taking conservative care to avoid excessive recursion.
1231 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1232 // Skip if every incoming value references to ourself.
1233 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1236 KnownZero = APInt::getAllOnesValue(BitWidth);
1237 KnownOne = APInt::getAllOnesValue(BitWidth);
1238 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1239 // Skip direct self references.
1240 if (P->getIncomingValue(i) == P) continue;
1242 KnownZero2 = APInt(BitWidth, 0);
1243 KnownOne2 = APInt(BitWidth, 0);
1244 // Recurse, but cap the recursion to one level, because we don't
1245 // want to waste time spinning around in loops.
1246 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
1248 KnownZero &= KnownZero2;
1249 KnownOne &= KnownOne2;
1250 // If all bits have been ruled out, there's no need to check
1252 if (!KnownZero && !KnownOne)
1258 case Instruction::Call:
1259 case Instruction::Invoke:
1260 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1261 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1262 // If a range metadata is attached to this IntrinsicInst, intersect the
1263 // explicit range specified by the metadata and the implicit range of
1265 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1266 switch (II->getIntrinsicID()) {
1268 case Intrinsic::ctlz:
1269 case Intrinsic::cttz: {
1270 unsigned LowBits = Log2_32(BitWidth)+1;
1271 // If this call is undefined for 0, the result will be less than 2^n.
1272 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1274 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1277 case Intrinsic::ctpop: {
1278 unsigned LowBits = Log2_32(BitWidth)+1;
1279 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1282 case Intrinsic::x86_sse42_crc32_64_64:
1283 KnownZero |= APInt::getHighBitsSet(64, 32);
1288 case Instruction::ExtractValue:
1289 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1290 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1291 if (EVI->getNumIndices() != 1) break;
1292 if (EVI->getIndices()[0] == 0) {
1293 switch (II->getIntrinsicID()) {
1295 case Intrinsic::uadd_with_overflow:
1296 case Intrinsic::sadd_with_overflow:
1297 computeKnownBitsAddSub(true, II->getArgOperand(0),
1298 II->getArgOperand(1), false, KnownZero,
1299 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1301 case Intrinsic::usub_with_overflow:
1302 case Intrinsic::ssub_with_overflow:
1303 computeKnownBitsAddSub(false, II->getArgOperand(0),
1304 II->getArgOperand(1), false, KnownZero,
1305 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1307 case Intrinsic::umul_with_overflow:
1308 case Intrinsic::smul_with_overflow:
1309 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
1310 false, KnownZero, KnownOne,
1311 KnownZero2, KnownOne2, TD, Depth, Q);
1318 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1321 /// Determine whether the sign bit is known to be zero or one.
1322 /// Convenience wrapper around computeKnownBits.
1323 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1324 const DataLayout *TD, unsigned Depth,
1326 unsigned BitWidth = getBitWidth(V->getType(), TD);
1332 APInt ZeroBits(BitWidth, 0);
1333 APInt OneBits(BitWidth, 0);
1334 computeKnownBits(V, ZeroBits, OneBits, TD, Depth, Q);
1335 KnownOne = OneBits[BitWidth - 1];
1336 KnownZero = ZeroBits[BitWidth - 1];
1339 /// Return true if the given value is known to have exactly one
1340 /// bit set when defined. For vectors return true if every element is known to
1341 /// be a power of two when defined. Supports values with integer or pointer
1342 /// types and vectors of integers.
1343 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1345 if (Constant *C = dyn_cast<Constant>(V)) {
1346 if (C->isNullValue())
1348 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1349 return CI->getValue().isPowerOf2();
1350 // TODO: Handle vector constants.
1353 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1354 // it is shifted off the end then the result is undefined.
1355 if (match(V, m_Shl(m_One(), m_Value())))
1358 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1359 // bottom. If it is shifted off the bottom then the result is undefined.
1360 if (match(V, m_LShr(m_SignBit(), m_Value())))
1363 // The remaining tests are all recursive, so bail out if we hit the limit.
1364 if (Depth++ == MaxDepth)
1367 Value *X = nullptr, *Y = nullptr;
1368 // A shift of a power of two is a power of two or zero.
1369 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1370 match(V, m_Shr(m_Value(X), m_Value()))))
1371 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q);
1373 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1374 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1376 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1378 isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1379 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1381 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1382 // A power of two and'd with anything is a power of two or zero.
1383 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q) ||
1384 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth, Q))
1386 // X & (-X) is always a power of two or zero.
1387 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1392 // Adding a power-of-two or zero to the same power-of-two or zero yields
1393 // either the original power-of-two, a larger power-of-two or zero.
1394 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1395 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1396 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1397 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1398 match(X, m_And(m_Value(), m_Specific(Y))))
1399 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1401 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1402 match(Y, m_And(m_Value(), m_Specific(X))))
1403 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1406 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1407 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1408 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth, Q);
1410 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1411 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth, Q);
1412 // If i8 V is a power of two or zero:
1413 // ZeroBits: 1 1 1 0 1 1 1 1
1414 // ~ZeroBits: 0 0 0 1 0 0 0 0
1415 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1416 // If OrZero isn't set, we cannot give back a zero result.
1417 // Make sure either the LHS or RHS has a bit set.
1418 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1423 // An exact divide or right shift can only shift off zero bits, so the result
1424 // is a power of two only if the first operand is a power of two and not
1425 // copying a sign bit (sdiv int_min, 2).
1426 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1427 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1428 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1435 /// \brief Test whether a GEP's result is known to be non-null.
1437 /// Uses properties inherent in a GEP to try to determine whether it is known
1440 /// Currently this routine does not support vector GEPs.
1441 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
1442 unsigned Depth, const Query &Q) {
1443 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1446 // FIXME: Support vector-GEPs.
1447 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1449 // If the base pointer is non-null, we cannot walk to a null address with an
1450 // inbounds GEP in address space zero.
1451 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1454 // Past this, if we don't have DataLayout, we can't do much.
1458 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1459 // If so, then the GEP cannot produce a null pointer, as doing so would
1460 // inherently violate the inbounds contract within address space zero.
1461 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1462 GTI != GTE; ++GTI) {
1463 // Struct types are easy -- they must always be indexed by a constant.
1464 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1465 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1466 unsigned ElementIdx = OpC->getZExtValue();
1467 const StructLayout *SL = DL->getStructLayout(STy);
1468 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1469 if (ElementOffset > 0)
1474 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1475 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
1478 // Fast path the constant operand case both for efficiency and so we don't
1479 // increment Depth when just zipping down an all-constant GEP.
1480 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1486 // We post-increment Depth here because while isKnownNonZero increments it
1487 // as well, when we pop back up that increment won't persist. We don't want
1488 // to recurse 10k times just because we have 10k GEP operands. We don't
1489 // bail completely out because we want to handle constant GEPs regardless
1491 if (Depth++ >= MaxDepth)
1494 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1501 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1502 /// ensure that the value it's attached to is never Value? 'RangeType' is
1503 /// is the type of the value described by the range.
1504 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1505 const APInt& Value) {
1506 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1507 assert(NumRanges >= 1);
1508 for (unsigned i = 0; i < NumRanges; ++i) {
1509 ConstantInt *Lower =
1510 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1511 ConstantInt *Upper =
1512 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1513 ConstantRange Range(Lower->getValue(), Upper->getValue());
1514 if (Range.contains(Value))
1520 /// Return true if the given value is known to be non-zero when defined.
1521 /// For vectors return true if every element is known to be non-zero when
1522 /// defined. Supports values with integer or pointer type and vectors of
1524 bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
1526 if (Constant *C = dyn_cast<Constant>(V)) {
1527 if (C->isNullValue())
1529 if (isa<ConstantInt>(C))
1530 // Must be non-zero due to null test above.
1532 // TODO: Handle vectors
1536 if (Instruction* I = dyn_cast<Instruction>(V)) {
1537 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1538 // If the possible ranges don't contain zero, then the value is
1539 // definitely non-zero.
1540 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1541 const APInt ZeroValue(Ty->getBitWidth(), 0);
1542 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1548 // The remaining tests are all recursive, so bail out if we hit the limit.
1549 if (Depth++ >= MaxDepth)
1552 // Check for pointer simplifications.
1553 if (V->getType()->isPointerTy()) {
1554 if (isKnownNonNull(V))
1556 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1557 if (isGEPKnownNonNull(GEP, TD, Depth, Q))
1561 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1563 // X | Y != 0 if X != 0 or Y != 0.
1564 Value *X = nullptr, *Y = nullptr;
1565 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1566 return isKnownNonZero(X, TD, Depth, Q) ||
1567 isKnownNonZero(Y, TD, Depth, Q);
1569 // ext X != 0 if X != 0.
1570 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1571 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth, Q);
1573 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1574 // if the lowest bit is shifted off the end.
1575 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1576 // shl nuw can't remove any non-zero bits.
1577 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1578 if (BO->hasNoUnsignedWrap())
1579 return isKnownNonZero(X, TD, Depth, Q);
1581 APInt KnownZero(BitWidth, 0);
1582 APInt KnownOne(BitWidth, 0);
1583 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1587 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1588 // defined if the sign bit is shifted off the end.
1589 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1590 // shr exact can only shift out zero bits.
1591 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1593 return isKnownNonZero(X, TD, Depth, Q);
1595 bool XKnownNonNegative, XKnownNegative;
1596 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1600 // div exact can only produce a zero if the dividend is zero.
1601 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1602 return isKnownNonZero(X, TD, Depth, Q);
1605 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1606 bool XKnownNonNegative, XKnownNegative;
1607 bool YKnownNonNegative, YKnownNegative;
1608 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1609 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth, Q);
1611 // If X and Y are both non-negative (as signed values) then their sum is not
1612 // zero unless both X and Y are zero.
1613 if (XKnownNonNegative && YKnownNonNegative)
1614 if (isKnownNonZero(X, TD, Depth, Q) ||
1615 isKnownNonZero(Y, TD, Depth, Q))
1618 // If X and Y are both negative (as signed values) then their sum is not
1619 // zero unless both X and Y equal INT_MIN.
1620 if (BitWidth && XKnownNegative && YKnownNegative) {
1621 APInt KnownZero(BitWidth, 0);
1622 APInt KnownOne(BitWidth, 0);
1623 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1624 // The sign bit of X is set. If some other bit is set then X is not equal
1626 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1627 if ((KnownOne & Mask) != 0)
1629 // The sign bit of Y is set. If some other bit is set then Y is not equal
1631 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth, Q);
1632 if ((KnownOne & Mask) != 0)
1636 // The sum of a non-negative number and a power of two is not zero.
1637 if (XKnownNonNegative &&
1638 isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth, Q))
1640 if (YKnownNonNegative &&
1641 isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth, Q))
1645 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1646 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1647 // If X and Y are non-zero then so is X * Y as long as the multiplication
1648 // does not overflow.
1649 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1650 isKnownNonZero(X, TD, Depth, Q) &&
1651 isKnownNonZero(Y, TD, Depth, Q))
1654 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1655 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1656 if (isKnownNonZero(SI->getTrueValue(), TD, Depth, Q) &&
1657 isKnownNonZero(SI->getFalseValue(), TD, Depth, Q))
1661 if (!BitWidth) return false;
1662 APInt KnownZero(BitWidth, 0);
1663 APInt KnownOne(BitWidth, 0);
1664 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1665 return KnownOne != 0;
1668 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1669 /// simplify operations downstream. Mask is known to be zero for bits that V
1672 /// This function is defined on values with integer type, values with pointer
1673 /// type (but only if TD is non-null), and vectors of integers. In the case
1674 /// where V is a vector, the mask, known zero, and known one values are the
1675 /// same width as the vector element, and the bit is set only if it is true
1676 /// for all of the elements in the vector.
1677 bool MaskedValueIsZero(Value *V, const APInt &Mask,
1678 const DataLayout *TD, unsigned Depth,
1680 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1681 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1682 return (KnownZero & Mask) == Mask;
1687 /// Return the number of times the sign bit of the register is replicated into
1688 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1689 /// (itself), but other cases can give us information. For example, immediately
1690 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1691 /// other, so we return 3.
1693 /// 'Op' must have a scalar integer type.
1695 unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
1696 unsigned Depth, const Query &Q) {
1697 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1698 "ComputeNumSignBits requires a DataLayout object to operate "
1699 "on non-integer values!");
1700 Type *Ty = V->getType();
1701 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1702 Ty->getScalarSizeInBits();
1704 unsigned FirstAnswer = 1;
1706 // Note that ConstantInt is handled by the general computeKnownBits case
1710 return 1; // Limit search depth.
1712 Operator *U = dyn_cast<Operator>(V);
1713 switch (Operator::getOpcode(V)) {
1715 case Instruction::SExt:
1716 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1717 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q) + Tmp;
1719 case Instruction::AShr: {
1720 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1721 // ashr X, C -> adds C sign bits. Vectors too.
1723 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1724 Tmp += ShAmt->getZExtValue();
1725 if (Tmp > TyBits) Tmp = TyBits;
1729 case Instruction::Shl: {
1731 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1732 // shl destroys sign bits.
1733 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1734 Tmp2 = ShAmt->getZExtValue();
1735 if (Tmp2 >= TyBits || // Bad shift.
1736 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1741 case Instruction::And:
1742 case Instruction::Or:
1743 case Instruction::Xor: // NOT is handled here.
1744 // Logical binary ops preserve the number of sign bits at the worst.
1745 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1747 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1748 FirstAnswer = std::min(Tmp, Tmp2);
1749 // We computed what we know about the sign bits as our first
1750 // answer. Now proceed to the generic code that uses
1751 // computeKnownBits, and pick whichever answer is better.
1755 case Instruction::Select:
1756 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1757 if (Tmp == 1) return 1; // Early out.
1758 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1, Q);
1759 return std::min(Tmp, Tmp2);
1761 case Instruction::Add:
1762 // Add can have at most one carry bit. Thus we know that the output
1763 // is, at worst, one more bit than the inputs.
1764 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1765 if (Tmp == 1) return 1; // Early out.
1767 // Special case decrementing a value (ADD X, -1):
1768 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
1769 if (CRHS->isAllOnesValue()) {
1770 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1771 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1773 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1775 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1778 // If we are subtracting one from a positive number, there is no carry
1779 // out of the result.
1780 if (KnownZero.isNegative())
1784 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1785 if (Tmp2 == 1) return 1;
1786 return std::min(Tmp, Tmp2)-1;
1788 case Instruction::Sub:
1789 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1790 if (Tmp2 == 1) return 1;
1793 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
1794 if (CLHS->isNullValue()) {
1795 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1796 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
1797 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1799 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1802 // If the input is known to be positive (the sign bit is known clear),
1803 // the output of the NEG has the same number of sign bits as the input.
1804 if (KnownZero.isNegative())
1807 // Otherwise, we treat this like a SUB.
1810 // Sub can have at most one carry bit. Thus we know that the output
1811 // is, at worst, one more bit than the inputs.
1812 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1813 if (Tmp == 1) return 1; // Early out.
1814 return std::min(Tmp, Tmp2)-1;
1816 case Instruction::PHI: {
1817 PHINode *PN = cast<PHINode>(U);
1818 // Don't analyze large in-degree PHIs.
1819 if (PN->getNumIncomingValues() > 4) break;
1821 // Take the minimum of all incoming values. This can't infinitely loop
1822 // because of our depth threshold.
1823 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1, Q);
1824 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
1825 if (Tmp == 1) return Tmp;
1827 ComputeNumSignBits(PN->getIncomingValue(i), TD,
1833 case Instruction::Trunc:
1834 // FIXME: it's tricky to do anything useful for this, but it is an important
1835 // case for targets like X86.
1839 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1840 // use this information.
1841 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1843 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1845 if (KnownZero.isNegative()) { // sign bit is 0
1847 } else if (KnownOne.isNegative()) { // sign bit is 1;
1854 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1855 // the number of identical bits in the top of the input value.
1857 Mask <<= Mask.getBitWidth()-TyBits;
1858 // Return # leading zeros. We use 'min' here in case Val was zero before
1859 // shifting. We don't want to return '64' as for an i32 "0".
1860 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1863 /// This function computes the integer multiple of Base that equals V.
1864 /// If successful, it returns true and returns the multiple in
1865 /// Multiple. If unsuccessful, it returns false. It looks
1866 /// through SExt instructions only if LookThroughSExt is true.
1867 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1868 bool LookThroughSExt, unsigned Depth) {
1869 const unsigned MaxDepth = 6;
1871 assert(V && "No Value?");
1872 assert(Depth <= MaxDepth && "Limit Search Depth");
1873 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1875 Type *T = V->getType();
1877 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1887 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1888 Constant *BaseVal = ConstantInt::get(T, Base);
1889 if (CO && CO == BaseVal) {
1891 Multiple = ConstantInt::get(T, 1);
1895 if (CI && CI->getZExtValue() % Base == 0) {
1896 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1900 if (Depth == MaxDepth) return false; // Limit search depth.
1902 Operator *I = dyn_cast<Operator>(V);
1903 if (!I) return false;
1905 switch (I->getOpcode()) {
1907 case Instruction::SExt:
1908 if (!LookThroughSExt) return false;
1909 // otherwise fall through to ZExt
1910 case Instruction::ZExt:
1911 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1912 LookThroughSExt, Depth+1);
1913 case Instruction::Shl:
1914 case Instruction::Mul: {
1915 Value *Op0 = I->getOperand(0);
1916 Value *Op1 = I->getOperand(1);
1918 if (I->getOpcode() == Instruction::Shl) {
1919 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1920 if (!Op1CI) return false;
1921 // Turn Op0 << Op1 into Op0 * 2^Op1
1922 APInt Op1Int = Op1CI->getValue();
1923 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1924 APInt API(Op1Int.getBitWidth(), 0);
1925 API.setBit(BitToSet);
1926 Op1 = ConstantInt::get(V->getContext(), API);
1929 Value *Mul0 = nullptr;
1930 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1931 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1932 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1933 if (Op1C->getType()->getPrimitiveSizeInBits() <
1934 MulC->getType()->getPrimitiveSizeInBits())
1935 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1936 if (Op1C->getType()->getPrimitiveSizeInBits() >
1937 MulC->getType()->getPrimitiveSizeInBits())
1938 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1940 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1941 Multiple = ConstantExpr::getMul(MulC, Op1C);
1945 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1946 if (Mul0CI->getValue() == 1) {
1947 // V == Base * Op1, so return Op1
1953 Value *Mul1 = nullptr;
1954 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1955 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1956 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1957 if (Op0C->getType()->getPrimitiveSizeInBits() <
1958 MulC->getType()->getPrimitiveSizeInBits())
1959 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1960 if (Op0C->getType()->getPrimitiveSizeInBits() >
1961 MulC->getType()->getPrimitiveSizeInBits())
1962 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1964 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1965 Multiple = ConstantExpr::getMul(MulC, Op0C);
1969 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1970 if (Mul1CI->getValue() == 1) {
1971 // V == Base * Op0, so return Op0
1979 // We could not determine if V is a multiple of Base.
1983 /// Return true if we can prove that the specified FP value is never equal to
1986 /// NOTE: this function will need to be revisited when we support non-default
1989 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
1990 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1991 return !CFP->getValueAPF().isNegZero();
1994 return 1; // Limit search depth.
1996 const Operator *I = dyn_cast<Operator>(V);
1997 if (!I) return false;
1999 // Check if the nsz fast-math flag is set
2000 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2001 if (FPO->hasNoSignedZeros())
2004 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2005 if (I->getOpcode() == Instruction::FAdd)
2006 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2007 if (CFP->isNullValue())
2010 // sitofp and uitofp turn into +0.0 for zero.
2011 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2014 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2015 // sqrt(-0.0) = -0.0, no other negative results are possible.
2016 if (II->getIntrinsicID() == Intrinsic::sqrt)
2017 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2019 if (const CallInst *CI = dyn_cast<CallInst>(I))
2020 if (const Function *F = CI->getCalledFunction()) {
2021 if (F->isDeclaration()) {
2023 if (F->getName() == "abs") return true;
2024 // fabs[lf](x) != -0.0
2025 if (F->getName() == "fabs") return true;
2026 if (F->getName() == "fabsf") return true;
2027 if (F->getName() == "fabsl") return true;
2028 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2029 F->getName() == "sqrtl")
2030 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2037 /// If the specified value can be set by repeating the same byte in memory,
2038 /// return the i8 value that it is represented with. This is
2039 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2040 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2041 /// byte store (e.g. i16 0x1234), return null.
2042 Value *llvm::isBytewiseValue(Value *V) {
2043 // All byte-wide stores are splatable, even of arbitrary variables.
2044 if (V->getType()->isIntegerTy(8)) return V;
2046 // Handle 'null' ConstantArrayZero etc.
2047 if (Constant *C = dyn_cast<Constant>(V))
2048 if (C->isNullValue())
2049 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2051 // Constant float and double values can be handled as integer values if the
2052 // corresponding integer value is "byteable". An important case is 0.0.
2053 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2054 if (CFP->getType()->isFloatTy())
2055 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2056 if (CFP->getType()->isDoubleTy())
2057 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2058 // Don't handle long double formats, which have strange constraints.
2061 // We can handle constant integers that are power of two in size and a
2062 // multiple of 8 bits.
2063 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2064 unsigned Width = CI->getBitWidth();
2065 if (isPowerOf2_32(Width) && Width > 8) {
2066 // We can handle this value if the recursive binary decomposition is the
2067 // same at all levels.
2068 APInt Val = CI->getValue();
2070 while (Val.getBitWidth() != 8) {
2071 unsigned NextWidth = Val.getBitWidth()/2;
2072 Val2 = Val.lshr(NextWidth);
2073 Val2 = Val2.trunc(Val.getBitWidth()/2);
2074 Val = Val.trunc(Val.getBitWidth()/2);
2076 // If the top/bottom halves aren't the same, reject it.
2080 return ConstantInt::get(V->getContext(), Val);
2084 // A ConstantDataArray/Vector is splatable if all its members are equal and
2086 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2087 Value *Elt = CA->getElementAsConstant(0);
2088 Value *Val = isBytewiseValue(Elt);
2092 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2093 if (CA->getElementAsConstant(I) != Elt)
2099 // Conceptually, we could handle things like:
2100 // %a = zext i8 %X to i16
2101 // %b = shl i16 %a, 8
2102 // %c = or i16 %a, %b
2103 // but until there is an example that actually needs this, it doesn't seem
2104 // worth worrying about.
2109 // This is the recursive version of BuildSubAggregate. It takes a few different
2110 // arguments. Idxs is the index within the nested struct From that we are
2111 // looking at now (which is of type IndexedType). IdxSkip is the number of
2112 // indices from Idxs that should be left out when inserting into the resulting
2113 // struct. To is the result struct built so far, new insertvalue instructions
2115 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2116 SmallVectorImpl<unsigned> &Idxs,
2118 Instruction *InsertBefore) {
2119 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2121 // Save the original To argument so we can modify it
2123 // General case, the type indexed by Idxs is a struct
2124 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2125 // Process each struct element recursively
2128 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2132 // Couldn't find any inserted value for this index? Cleanup
2133 while (PrevTo != OrigTo) {
2134 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2135 PrevTo = Del->getAggregateOperand();
2136 Del->eraseFromParent();
2138 // Stop processing elements
2142 // If we successfully found a value for each of our subaggregates
2146 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2147 // the struct's elements had a value that was inserted directly. In the latter
2148 // case, perhaps we can't determine each of the subelements individually, but
2149 // we might be able to find the complete struct somewhere.
2151 // Find the value that is at that particular spot
2152 Value *V = FindInsertedValue(From, Idxs);
2157 // Insert the value in the new (sub) aggregrate
2158 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2159 "tmp", InsertBefore);
2162 // This helper takes a nested struct and extracts a part of it (which is again a
2163 // struct) into a new value. For example, given the struct:
2164 // { a, { b, { c, d }, e } }
2165 // and the indices "1, 1" this returns
2168 // It does this by inserting an insertvalue for each element in the resulting
2169 // struct, as opposed to just inserting a single struct. This will only work if
2170 // each of the elements of the substruct are known (ie, inserted into From by an
2171 // insertvalue instruction somewhere).
2173 // All inserted insertvalue instructions are inserted before InsertBefore
2174 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2175 Instruction *InsertBefore) {
2176 assert(InsertBefore && "Must have someplace to insert!");
2177 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2179 Value *To = UndefValue::get(IndexedType);
2180 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2181 unsigned IdxSkip = Idxs.size();
2183 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2186 /// Given an aggregrate and an sequence of indices, see if
2187 /// the scalar value indexed is already around as a register, for example if it
2188 /// were inserted directly into the aggregrate.
2190 /// If InsertBefore is not null, this function will duplicate (modified)
2191 /// insertvalues when a part of a nested struct is extracted.
2192 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2193 Instruction *InsertBefore) {
2194 // Nothing to index? Just return V then (this is useful at the end of our
2196 if (idx_range.empty())
2198 // We have indices, so V should have an indexable type.
2199 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2200 "Not looking at a struct or array?");
2201 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2202 "Invalid indices for type?");
2204 if (Constant *C = dyn_cast<Constant>(V)) {
2205 C = C->getAggregateElement(idx_range[0]);
2206 if (!C) return nullptr;
2207 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2210 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2211 // Loop the indices for the insertvalue instruction in parallel with the
2212 // requested indices
2213 const unsigned *req_idx = idx_range.begin();
2214 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2215 i != e; ++i, ++req_idx) {
2216 if (req_idx == idx_range.end()) {
2217 // We can't handle this without inserting insertvalues
2221 // The requested index identifies a part of a nested aggregate. Handle
2222 // this specially. For example,
2223 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2224 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2225 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2226 // This can be changed into
2227 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2228 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2229 // which allows the unused 0,0 element from the nested struct to be
2231 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2235 // This insert value inserts something else than what we are looking for.
2236 // See if the (aggregrate) value inserted into has the value we are
2237 // looking for, then.
2239 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2242 // If we end up here, the indices of the insertvalue match with those
2243 // requested (though possibly only partially). Now we recursively look at
2244 // the inserted value, passing any remaining indices.
2245 return FindInsertedValue(I->getInsertedValueOperand(),
2246 makeArrayRef(req_idx, idx_range.end()),
2250 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2251 // If we're extracting a value from an aggregrate that was extracted from
2252 // something else, we can extract from that something else directly instead.
2253 // However, we will need to chain I's indices with the requested indices.
2255 // Calculate the number of indices required
2256 unsigned size = I->getNumIndices() + idx_range.size();
2257 // Allocate some space to put the new indices in
2258 SmallVector<unsigned, 5> Idxs;
2260 // Add indices from the extract value instruction
2261 Idxs.append(I->idx_begin(), I->idx_end());
2263 // Add requested indices
2264 Idxs.append(idx_range.begin(), idx_range.end());
2266 assert(Idxs.size() == size
2267 && "Number of indices added not correct?");
2269 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2271 // Otherwise, we don't know (such as, extracting from a function return value
2272 // or load instruction)
2276 /// Analyze the specified pointer to see if it can be expressed as a base
2277 /// pointer plus a constant offset. Return the base and offset to the caller.
2278 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2279 const DataLayout *DL) {
2280 // Without DataLayout, conservatively assume 64-bit offsets, which is
2281 // the widest we support.
2282 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
2283 APInt ByteOffset(BitWidth, 0);
2285 if (Ptr->getType()->isVectorTy())
2288 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2290 APInt GEPOffset(BitWidth, 0);
2291 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
2294 ByteOffset += GEPOffset;
2297 Ptr = GEP->getPointerOperand();
2298 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2299 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2300 Ptr = cast<Operator>(Ptr)->getOperand(0);
2301 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2302 if (GA->mayBeOverridden())
2304 Ptr = GA->getAliasee();
2309 Offset = ByteOffset.getSExtValue();
2314 /// This function computes the length of a null-terminated C string pointed to
2315 /// by V. If successful, it returns true and returns the string in Str.
2316 /// If unsuccessful, it returns false.
2317 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2318 uint64_t Offset, bool TrimAtNul) {
2321 // Look through bitcast instructions and geps.
2322 V = V->stripPointerCasts();
2324 // If the value is a GEP instructionor constant expression, treat it as an
2326 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2327 // Make sure the GEP has exactly three arguments.
2328 if (GEP->getNumOperands() != 3)
2331 // Make sure the index-ee is a pointer to array of i8.
2332 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2333 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2334 if (!AT || !AT->getElementType()->isIntegerTy(8))
2337 // Check to make sure that the first operand of the GEP is an integer and
2338 // has value 0 so that we are sure we're indexing into the initializer.
2339 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2340 if (!FirstIdx || !FirstIdx->isZero())
2343 // If the second index isn't a ConstantInt, then this is a variable index
2344 // into the array. If this occurs, we can't say anything meaningful about
2346 uint64_t StartIdx = 0;
2347 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2348 StartIdx = CI->getZExtValue();
2351 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
2354 // The GEP instruction, constant or instruction, must reference a global
2355 // variable that is a constant and is initialized. The referenced constant
2356 // initializer is the array that we'll use for optimization.
2357 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2358 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2361 // Handle the all-zeros case
2362 if (GV->getInitializer()->isNullValue()) {
2363 // This is a degenerate case. The initializer is constant zero so the
2364 // length of the string must be zero.
2369 // Must be a Constant Array
2370 const ConstantDataArray *Array =
2371 dyn_cast<ConstantDataArray>(GV->getInitializer());
2372 if (!Array || !Array->isString())
2375 // Get the number of elements in the array
2376 uint64_t NumElts = Array->getType()->getArrayNumElements();
2378 // Start out with the entire array in the StringRef.
2379 Str = Array->getAsString();
2381 if (Offset > NumElts)
2384 // Skip over 'offset' bytes.
2385 Str = Str.substr(Offset);
2388 // Trim off the \0 and anything after it. If the array is not nul
2389 // terminated, we just return the whole end of string. The client may know
2390 // some other way that the string is length-bound.
2391 Str = Str.substr(0, Str.find('\0'));
2396 // These next two are very similar to the above, but also look through PHI
2398 // TODO: See if we can integrate these two together.
2400 /// If we can compute the length of the string pointed to by
2401 /// the specified pointer, return 'len+1'. If we can't, return 0.
2402 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2403 // Look through noop bitcast instructions.
2404 V = V->stripPointerCasts();
2406 // If this is a PHI node, there are two cases: either we have already seen it
2408 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2409 if (!PHIs.insert(PN).second)
2410 return ~0ULL; // already in the set.
2412 // If it was new, see if all the input strings are the same length.
2413 uint64_t LenSoFar = ~0ULL;
2414 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2415 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2416 if (Len == 0) return 0; // Unknown length -> unknown.
2418 if (Len == ~0ULL) continue;
2420 if (Len != LenSoFar && LenSoFar != ~0ULL)
2421 return 0; // Disagree -> unknown.
2425 // Success, all agree.
2429 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2430 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2431 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2432 if (Len1 == 0) return 0;
2433 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2434 if (Len2 == 0) return 0;
2435 if (Len1 == ~0ULL) return Len2;
2436 if (Len2 == ~0ULL) return Len1;
2437 if (Len1 != Len2) return 0;
2441 // Otherwise, see if we can read the string.
2443 if (!getConstantStringInfo(V, StrData))
2446 return StrData.size()+1;
2449 /// If we can compute the length of the string pointed to by
2450 /// the specified pointer, return 'len+1'. If we can't, return 0.
2451 uint64_t llvm::GetStringLength(Value *V) {
2452 if (!V->getType()->isPointerTy()) return 0;
2454 SmallPtrSet<PHINode*, 32> PHIs;
2455 uint64_t Len = GetStringLengthH(V, PHIs);
2456 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2457 // an empty string as a length.
2458 return Len == ~0ULL ? 1 : Len;
2462 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
2463 if (!V->getType()->isPointerTy())
2465 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2466 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2467 V = GEP->getPointerOperand();
2468 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2469 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2470 V = cast<Operator>(V)->getOperand(0);
2471 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2472 if (GA->mayBeOverridden())
2474 V = GA->getAliasee();
2476 // See if InstructionSimplify knows any relevant tricks.
2477 if (Instruction *I = dyn_cast<Instruction>(V))
2478 // TODO: Acquire a DominatorTree and AssumptionTracker and use them.
2479 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
2486 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2492 llvm::GetUnderlyingObjects(Value *V,
2493 SmallVectorImpl<Value *> &Objects,
2494 const DataLayout *TD,
2495 unsigned MaxLookup) {
2496 SmallPtrSet<Value *, 4> Visited;
2497 SmallVector<Value *, 4> Worklist;
2498 Worklist.push_back(V);
2500 Value *P = Worklist.pop_back_val();
2501 P = GetUnderlyingObject(P, TD, MaxLookup);
2503 if (!Visited.insert(P).second)
2506 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2507 Worklist.push_back(SI->getTrueValue());
2508 Worklist.push_back(SI->getFalseValue());
2512 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2513 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2514 Worklist.push_back(PN->getIncomingValue(i));
2518 Objects.push_back(P);
2519 } while (!Worklist.empty());
2522 /// Return true if the only users of this pointer are lifetime markers.
2523 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2524 for (const User *U : V->users()) {
2525 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2526 if (!II) return false;
2528 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2529 II->getIntrinsicID() != Intrinsic::lifetime_end)
2535 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
2536 const DataLayout *TD) {
2537 const Operator *Inst = dyn_cast<Operator>(V);
2541 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2542 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2546 switch (Inst->getOpcode()) {
2549 case Instruction::UDiv:
2550 case Instruction::URem: {
2551 // x / y is undefined if y == 0.
2553 if (match(Inst->getOperand(1), m_APInt(V)))
2557 case Instruction::SDiv:
2558 case Instruction::SRem: {
2559 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
2561 if (match(Inst->getOperand(1), m_APInt(Y))) {
2564 // The numerator can't be MinSignedValue if the denominator is -1.
2565 if (match(Inst->getOperand(0), m_APInt(X)))
2566 return !Y->isMinSignedValue();
2567 // The numerator *might* be MinSignedValue.
2570 // The denominator is not 0 or -1, it's safe to proceed.
2576 case Instruction::Load: {
2577 const LoadInst *LI = cast<LoadInst>(Inst);
2578 if (!LI->isUnordered() ||
2579 // Speculative load may create a race that did not exist in the source.
2580 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2582 return LI->getPointerOperand()->isDereferenceablePointer(TD);
2584 case Instruction::Call: {
2585 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2586 switch (II->getIntrinsicID()) {
2587 // These synthetic intrinsics have no side-effects and just mark
2588 // information about their operands.
2589 // FIXME: There are other no-op synthetic instructions that potentially
2590 // should be considered at least *safe* to speculate...
2591 case Intrinsic::dbg_declare:
2592 case Intrinsic::dbg_value:
2595 case Intrinsic::bswap:
2596 case Intrinsic::ctlz:
2597 case Intrinsic::ctpop:
2598 case Intrinsic::cttz:
2599 case Intrinsic::objectsize:
2600 case Intrinsic::sadd_with_overflow:
2601 case Intrinsic::smul_with_overflow:
2602 case Intrinsic::ssub_with_overflow:
2603 case Intrinsic::uadd_with_overflow:
2604 case Intrinsic::umul_with_overflow:
2605 case Intrinsic::usub_with_overflow:
2607 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2608 // errno like libm sqrt would.
2609 case Intrinsic::sqrt:
2610 case Intrinsic::fma:
2611 case Intrinsic::fmuladd:
2612 case Intrinsic::fabs:
2613 case Intrinsic::minnum:
2614 case Intrinsic::maxnum:
2616 // TODO: some fp intrinsics are marked as having the same error handling
2617 // as libm. They're safe to speculate when they won't error.
2618 // TODO: are convert_{from,to}_fp16 safe?
2619 // TODO: can we list target-specific intrinsics here?
2623 return false; // The called function could have undefined behavior or
2624 // side-effects, even if marked readnone nounwind.
2626 case Instruction::VAArg:
2627 case Instruction::Alloca:
2628 case Instruction::Invoke:
2629 case Instruction::PHI:
2630 case Instruction::Store:
2631 case Instruction::Ret:
2632 case Instruction::Br:
2633 case Instruction::IndirectBr:
2634 case Instruction::Switch:
2635 case Instruction::Unreachable:
2636 case Instruction::Fence:
2637 case Instruction::LandingPad:
2638 case Instruction::AtomicRMW:
2639 case Instruction::AtomicCmpXchg:
2640 case Instruction::Resume:
2641 return false; // Misc instructions which have effects
2645 /// Return true if we know that the specified value is never null.
2646 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2647 // Alloca never returns null, malloc might.
2648 if (isa<AllocaInst>(V)) return true;
2650 // A byval, inalloca, or nonnull argument is never null.
2651 if (const Argument *A = dyn_cast<Argument>(V))
2652 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2654 // Global values are not null unless extern weak.
2655 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2656 return !GV->hasExternalWeakLinkage();
2658 // A Load tagged w/nonnull metadata is never null.
2659 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2660 return LI->getMetadata(LLVMContext::MD_nonnull);
2662 if (ImmutableCallSite CS = V)
2663 if (CS.isReturnNonNull())
2666 // operator new never returns null.
2667 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))